Daily Papers

Diffusion Language Models (DLMs) have recently achieved significant success due to their any-order generation capabilities. However, existing inference methods typically rely on local, immediate-step metrics such as confidence or entropy which inherently lack a more reliable perspective. This limitation frequently leads to inconsistent sampling trajectories and suboptimal generation quality. To address this, we propose Coherent Contextual Decoding (CCD), a novel inference framework built upon two core innovations. First, CCD employs a trajectory rectification mechanism that leverages historical context to enhance sequence coherence, enabling the early rejection of suboptimal paths. We demonstrate that this mechanism is theoretically equivalent to modeling the consistency of historical steps via the conditional mutual information between context and token predictions. Building on this theoretical insight, we further address the inefficiency of conventional uniform decoding budgets. Instead of rigid allocations based on diffusion steps, we introduce an adaptive sampling strategy that dynamically adjusts the unmasking budget for each step according to our consistency metric. Consequently, our method significantly improves the quality of generation trajectories while accelerating the sampling process. Empirically, our method achieves a simultaneous enhancement in both inference speed and performance across diverse benchmarks on Dream and LLaDA, delivering up to 3.48x speedup alongside 3.91% performance improvement.

Diffusion large language models (dLLMs) accelerate language generation by predicting multiple masks in a single forward pass. However, existing dLLMs can suffer from unreliable predictions in early denoising stages under aggressive parallelism strategies, leading to errors that can propagate to later stages. To tackle this issue, we present Consistency Forcing (CForce) for dLLMs, a distillation method to force the mask predictions of early stages to align with those of later stages. CForce trains the model on pre-collected self-rollout trajectories, thereby improving training-inference alignment. We introduce Confidence Adaptive KL Divergence as a distillation objective to conjoin the merits of forward and reverse KL. We further provide a theoretical analysis for the consistency objective to explain why CForce can approximately minimize the prediction error of early stages. Critically, the same formulation applies to both mask-to-token decoding and edit-capable decoding; in the edit-capable case, later token-to-token refinements provide additional supervision for earlier masked-state predictions. Experiments on non-edit and edit-capable LLaDA models show improved speed-quality trade-offs, especially under high-parallelism decoding budgets. Code is available at: https://github.com/inclusionAI/dFactory.

Masked diffusion language models (MDLMs) generate text by iteratively unmasking tokens, but their standard decoder reduces each step to a binary action: a position is either committed to a single token or left fully masked, discarding rich predictive information rather than carrying it forward, and forcing premature, irrevocable commitments that lead to poor performance under a limited decoding budget. In this paper, we reinterpret mask prediction as a clean-state prediction (x-prediction) and show that it can be used to induce a continuous flow in the input embedding space. Building on this view, we propose a continuous decoding framework for MDLMs where tokens can accumulate partial progress at each diffusion step and remain revisable. To match the uneven contextual constraints across positions in language, we replace the globally synchronous schedule in image diffusion with a confidence-based asynchronous update in which the diffusion progress is token-wise accumulated. Additionally, we introduce a lightweight policy network and formulate its training as a reinforcement learning problem. Applied to pretrained LLaDA, our decoder retains 83--97% of full-budget accuracy using under 15% of the diffusion steps, largely outperforming discrete mask-prediction decoding at matched budgets.

Test-time scaling has been shown to substantially improve large language models' (LLMs) mathematical reasoning. However, for a large portion of mathematical corpora, especially theorem proving, RLVR's scalability is limited: intermediate reasoning is crucial, while final answers are difficult to directly and reliably verify. Meanwhile, token-level SFT often degenerates into rote memorization rather than inducing longer chains of thought. Inspired by BERT's self-supervised tasks, we propose MR-RLVR (Masked-and-Reordered RLVR), which constructs process-level self-supervised rewards via "masked-then-fill" and "step reordering" to extract learnable signals from intermediate reasoning. Our training pipeline comprises two stages: we first perform self-supervised training on sampled mathematical calculation and proof data; we then conduct RLVR fine-tuning on mathematical calculation datasets where only outcomes are verifiable. We implement MR-RLVR on Qwen2.5-3B and DeepSeek-R1-Distill-Qwen-1.5B, and evaluate on AIME24, AIME25, AMC23, and MATH500. Under a fixed sampling and decoding budget, MR-RLVR achieves average relative gains over the original RLVR of +9.86% Pass@1, +5.27% Pass@5, and +4.00% Pass@8. These results indicate that incorporating process-aware self-supervised signals can effectively enhance RLVR's scalability and performance in only outcome-verifiable settings.

Studies of bias in LLM-as-judge systems typically build synthetic corpora by prompting an LLM to generate a hallucinated answer to pair with a factual one, then presenting both to a judge. We report a case in which this generation step silently failed, and use it to argue that the failure mode is structural rather than incidental. In a multilingual (Turkish/English) faithfulness-judgment corpus, a decoding-budget parameter shared between judging and generation calls truncated one producer's hallucinated answers to a few words. The resulting items produced a large, statistically robust effect: a 32-point cross-lingual collapse in one judge's selection accuracy, replicated from N=50 to N=500, explained by a three-layer mechanistic account, and confirmed by a controlled producer-swap experiment, none of which was real. The effect vanished to ceiling once the shared parameter was corrected, and only manual reading of the raw generations, not any aggregate statistical check, exposed the fault. A second measured bias (Markdown-formatting preference) was not fabricated but distorted by the same fault, its magnitude and in one case its sign shifting with stimulus length, a mode aggregate metrics cannot distinguish from the first. We frame the underlying vulnerability using the test oracle problem: corpora whose negative examples are LLM-generated carry no mechanical way to verify item integrity, while corpora built by deterministic perturbation of a gold answer carry an item-level oracle for free. A positive control supports this claim directly: an analogous fault injected into a minimal perturbation-based corpus is caught with 100% accuracy by a zero-cost, zero-human gold-to-negative string comparison. We close with a validation protocol, derived from our own case, for analysts working in the oracle-less regime that we argue describes most contemporary multilingual LLM-as-judge corpora.

Quantization-aware distillation (QAD) restores much of the short-form question-answering performance lost to sub-3-bit quantization, yet leaves mathematical and code reasoning substantially impaired. Long generations often degenerate into repetitive loops, exhausting the decoding budget without completing a solution. We trace this gap to quantization-amplified exposure bias: QAD trains on fixed corpus prefixes, while quantization-induced deviations compound along the model's own autoregressive trajectories. To address this mismatch, we introduce an on-policy distillation (OPD) stage that places teacher supervision where the quantized model actually goes. Starting from a QAD checkpoint, the student generates through the quantized forward path used at deployment and receives feedback from a frozen full-precision teacher on its own prefixes, combining dense token-level guidance with task-verifier rewards. Across four models at 2.79 and 1.88 effective bits, OPD raises average BF16 performance retention from 35% to 70% on MATH-500 and from 66% to 91% on HumanEval while preserving short-form performance, with reasoning gains substantially exceeding those of continued teacher-forced QAD in matched-budget comparisons. By coupling QAD's stable low-bit initialization with OPD's on-policy reasoning recovery, our framework provides a comprehensive sub-3-bit solution that preserves broad capabilities while restoring long-form reasoning.

MLLM-based detectors formulate vision-language grounding as autoregressive coordinate generation, yet in dense scenes, they often collapse into a box repetition loop, repeatedly emitting identical boxes until the decoding budget is exhausted. We trace this failure to historical-answer attention takeover: Under full-history autoregression, repeated coordinate tokens in the growing answer history attract disproportionate attention, crowding out the attention sinks needed to attend to new visual evidence. % To break this loop, in this paper, we propose Compressed Window Auto-Regression (WinAR). This new mechanism partitions a long answer sequence into consecutive generation windows and restricts full token-level autoregression to the current window only while compressing each previous window into a fixed set of latent tokens that condition subsequent decoding. This design reduces the repeated exposure to raw historical coordinates that triggers the attention takeover, yet preserves cross-window dependencies essential for coherent multi-object generation. On RSDense, which contains 200-2,000 objects per image, our model WinARDet improves RexOmni F1-score (IoU=0.5) from 32.16 to 39.36 over the baseline, while reducing the redundancy rate from 35.48% to 13.00%. WinARDet also generalizes across decoding strategies and output formats, including parallel box decoding and polygon generation, providing a practical paradigm for long-horizon vision-language grounding across scene densities.

Language models can learn from experience, but raw solution trajectories are often too long and noisy to provide effective guidance. In this work, we propose Learning to Coach (L2C), a framework that trains a dedicated LLM-as-a-Coach to extract actionable experiential knowledge from an actor model's previous trajectory. The actor remains frozen, while the LLM-as-a-Coach is trained to maximize a reward given by the correctness of the actor's guided response. We study two such rewards: a same-instance reward, which improves subsequent responses on the original problem, and a cross-instance reward, which elicits knowledge that transfers to other instances. Across mathematical reasoning and interactive text-games, L2C consistently outperforms self-refinement and an untrained LLM-as-a-Coach. Running experiential learning for more iterations further improves accuracy and uses additional inference compute more effectively than enlarging the actor's decoding budget. The trained LLM-as-a-Coach also transfers to out-of-distribution tasks and adapts its guidance to the specific actor it coaches.

While large language models (LLMs) excel at handling long-context sequences, they require substantial prefill computation and key-value (KV) cache, which can heavily burden computational efficiency and memory usage in both prefill and decoding stages. Recent works that compress KV caches with prefill acceleration reduce this cost but inadvertently tie the prefill compute reduction to the decoding KV budget. This coupling arises from overlooking the layer-dependent variation of critical context, often leading to accuracy degradation. To address this issue, we introduce FastKV, a KV cache compression framework designed to reduce latency in both prefill and decoding by leveraging the stabilization of token importance in later layers. FastKV performs full-context computation until a Token-Selective Propagation (TSP) layer, which forwards only the most informative tokens to subsequent layers. From these propagated tokens, FastKV independently selects salient KV entries for caching, thereby decoupling KV budget from the prefill compute reduction based on the TSP decision. This independent control of the TSP rate and KV retention rate enables flexible optimization of efficiency and accuracy. Experimental results show that FastKV achieves speedups of up to 1.82times in prefill and 2.87times in decoding compared to the full-context baseline, while matching the accuracy of the baselines that only accelerate the decoding stage. Our code is available at https://github.com/dongwonjo/FastKV.

The optimal training configurations of large language models (LLMs) with respect to model sizes and compute budgets have been extensively studied. But how to optimally configure LLMs during inference has not been explored in sufficient depth. We study compute-optimal inference: designing models and inference strategies that optimally trade off additional inference-time compute for improved performance. As a first step towards understanding and designing compute-optimal inference methods, we assessed the effectiveness and computational efficiency of multiple inference strategies such as Greedy Search, Majority Voting, Best-of-N, Weighted Voting, and their variants on two different Tree Search algorithms, involving different model sizes and computational budgets. We found that a smaller language model with a novel tree search algorithm typically achieves a Pareto-optimal trade-off. These results highlight the potential benefits of deploying smaller models equipped with more sophisticated decoding algorithms in budget-constrained scenarios, e.g., on end-devices, to enhance problem-solving accuracy. For instance, we show that the Llemma-7B model can achieve competitive accuracy to a Llemma-34B model on MATH500 while using 2times less FLOPs. Our findings could potentially apply to any generation task with a well-defined measure of success.

Long-context inference remains challenging for large language models due to attention dilution and out-of-distribution degradation. Context selection mitigates this limitation by attending to a subset of key-value cache entries, yet most methods allocate a fixed context budget throughout decoding despite highly non-uniform token-level contextual demands. To address this issue, we propose Uncertainty-Triggered Adaptive Context Allocation (UT-ACA), an inference-time framework that dynamically adjusts the context window based on token-wise uncertainty. UT-ACA learns an uncertainty detector that combines semantic embeddings with logit-based confidence while accounting for uncertainty accumulation across decoding steps. When insufficient evidence is indicated, UT-ACA selectively rolls back, expands the context window, and regenerates the token with additional support. Experiments show that UT-ACA substantially reduces average context usage while preserving generation quality in long-context settings.

Recent work improves reasoning in large reasoning models (LRMs) by test-time scaling through self-consistency and parallel thinking, adding generic "thinking tokens", or presenting the question twice, as in re-reading (RE2). However, these approaches either inject task-agnostic tokens or mandate heuristics that do not explain, and often ignore, the spontaneous repetition that many LRMs exhibit at the head of their long chain-of-thought (CoT). In this work, we investigate the model's tendency to restate the question, which we dub Echo of Prompt (EOP), as a mechanism that shapes computation from the start and can be viewed as prompt repetition inside the reasoning trace rather than in the input. We formalize its probabilistic cost by casting echo removal as rejection sampling, and define the Echo Likelihood Gap as a computable proxy linking early repetition to likelihood gains and downstream accuracy. We find that (1) in a causal intervention, reinserting an echo into failed GSM8K traces raises exact match by 7.9-10.4 points for two reasoning models but not for a non-reasoning base model; and (2) likelihood analyses controlling for length and suffix indicate that EOP's benefits go beyond verbosity, while layer-wise attention studies show that EOP increases attention from answer tokens to the answer prefix in middle layers, consistent with an attention refocusing mechanism. Building on these findings, we develop Echo-Distilled SFT (ED-SFT) to instill echoing before reasoning via supervised fine-tuning, and training-free Echoic Prompting (EP) to ground the model on the prompt midway through the trace. We evaluate on GSM8K, MathQA, Hendrycks-MATH, AIME24, and MATH-500 under identical decoding settings and budgets, and observe gains over baselines, including the insertion of thinking tokens, in most settings. Our code is publicly released at https://github.com/hhh2210/echoes-as-anchors.

Block-diffusion drafters have recently emerged as a powerful alternative for speculative decoding by predicting multiple future-token distributions in a single parallel step. However, since these parallel predictions are sampled from position-wise marginals rather than fully conditioned sequences, committing to a single greedy path often fails to capture the target model's preferred trajectory. To address this, we propose BASTION, a budget-aware speculative decoding framework with tree-based diffusion drafting. Unlike existing methods that rely on static tree topologies, BASTION dynamically constructs query-dependent trees by balancing draft quality against hardware constraints. Our framework integrates three synergistic components: (1) an acceptance surrogate that estimates expected accepted length via path confidence, (2) an online latency estimator that calibrates a hardware-aware roofline model, and (3) an adaptive best-first expansion that grows the tree until marginal gains no longer justify incremental verification costs. BASTION is training-free, preserves the target model's distribution, and requires no per-setting tuning. Across diverse benchmarks and GPU architectures, BASTION achieves up to a 6.61times speedup over standard autoregressive decoding, outperforming state-of-the-art block-diffusion baselines by 39%.

Neural speech codecs provide discrete tokens for speech generation. Lower frame rates shorten token sequences, but each token must then carry more linguistic and acoustic information. Variable-frame-rate codecs address this trade-off by adapting token allocation to speech content, but threshold-based allocation lets content determine the token count, so the sequence length a speech generator must produce varies across utterances and datasets at a fixed setting. We present RubatoCodec, a variable-frame-rate speech codec built on X-Codec-2 with an exact utterance-level event budget and streaming decoding. Its Adaptive Firing mechanism uses changes in codec features and waveform spectra to divide speech into variable-duration events without a learned boundary predictor. Discrete codes and an explicit duration represent each event, allowing a speech generator to predict and decode one event at a time. Controlled experiments show that adaptive allocation outperforms fixed-rate coding in reconstruction fidelity and intelligibility at low event rates. A model trained on large-scale multilingual speech achieves competitive reconstruction quality across English, Japanese, and Mandarin, with the best intelligibility preservation among the compared codecs on the evaluated English and Japanese benchmarks at comparable bitrates. In a matched text-to-speech comparison, the adaptive representation yields lower WER and higher UTMOS and UTMOSv2 with comparable speaker similarity. The speech generator also learns the codec's utterance-level event budget without an explicit generation-time rate constraint. These results demonstrate that budgeted adaptive speech representations support both high-quality reconstruction and streaming text-to-speech generation. Code and weights are available at https://github.com/third-intelligence/RubatoCodec and https://huggingface.co/third-intelligence/RubatoCodec.

Speculative decoding accelerates Large Language Model (LLM) inference by verifying multiple drafted tokens in parallel. However, for Mixture-of-Experts (MoE) models, this parallelism introduces a severe bottleneck: large draft trees activate many unique experts, significantly increasing memory pressure and diminishing speedups from speculative decoding relative to autoregressive decoding. Prior methods reduce speculation depth when MoE verification becomes expensive. We propose MoE-Spec, a training-free verification-time expert budgeting method that decouples speculation depth from memory cost by enforcing a fixed expert capacity limit at each layer, loading only the experts that contribute most to verification and dropping the long tail of rarely used experts that drive bandwidth overhead. Experiments across multiple model scales and datasets show that this method yields 10--30\% higher throughput than state-of-the-art speculative decoding baselines (EAGLE-3) at comparable quality, with flexibility to trade accuracy for further latency reductions through tighter budgets.

Most image generation models rely on uniform tokenization, allocating the exact same computational budget to equally-sized image patches. This static paradigm cannot adapt to different resource constraints at inference time, and yields suboptimal quality-cost tradeoff by devoting the same effort to both plain backgrounds and intricate details. We propose BudgetPix, an adaptive tokenization framework that dynamically allocates compute based on visual complexity and spatial layout, enabling flexible computational budgeting at inference time. BudgetPix comprises three key components: (1) an adaptive encoder that maps a fixed-size image to a variable-length token sequence using an entropy-guided quadtree alongside a multi-scale patch embedder; (2) a scale-aware decoder reconstructs fixed-resolution images from multi-scale token sets; and (3) a flexible training and sampling schedule that enables pixel-space denoisers to operate across variable token counts. BudgetPix seamlessly integrates with existing pixel-space diffusion architectures, enabling a single checkpoint to be operated at a wide range of compute budgets. Evaluated on text-to-image generation, BudgetPix matches the fidelity of MiniT2I-L at 512^2 and PixelDiT at 1024^2 using just 25% of the original compute budget. In class-conditional generation using a MeanFlow backbone, BudgetPix requires merely 60% of the full compute budget to produce images with near-zero quality degradation, observing a marginal 0.8-point increase in FID. Comprehensive assessments by human and VLM judges confirm that BudgetPix establishes a significantly improved quality-efficiency tradeoff over prior budget-adaptive baselines. More details are available at our project page: https://karaozgur.com/BudgetPix

Tree-search decoding is an effective form of test-time scaling for large language models (LLMs), but real-world deployment imposes a fixed per-query token budget that varies across settings. Existing tree-search policies are largely budget-agnostic, treating the budget as a termination condition, which can lead to late-stage over-branching or premature termination. We propose {Budget-Guided MCTS} (BG-MCTS), a tree-search decoding algorithm that aligns its search policy with the remaining token budget: it starts with broad exploration, then prioritizes refinement and answer completion as the budget depletes while reducing late-stage branching from shallow nodes. BG-MCTS consistently outperforms budget-agnostic tree-search baselines across different budgets on MATH500 and AIME24/25 with open-weight LLMs.

Inference-time scaling has emerged as a powerful alternative to parameter scaling for improving language model performance on complex reasoning tasks. While existing methods have shown strong performance gains under fixed compute budgets, there has been little focus on optimally utilizing that budget during inference. In this work, we introduce A*-decoding, a search-based inference-time strategy that builds on the A* search algorithm to optimally utilize a fixed compute budget by prioritizing high-quality reasoning paths during generation. We frame language model decoding as a structured search in a state space of partial solutions, applying the A* transition model to identify promising continuations guided by an external process supervision signal. In our experiments, A*-decoding reaches the performance levels of strong inference scaling baselines like best-of-N and particle filtering while using up to 3x fewer tokens and 30% fewer PRM passes under equivalent compute budgets. On the MATH500 and AIME 2024 benchmarks, A*-decoding enables Llama-3.2-1B-Instruct to match the performance of the 70x larger Llama-3.1-70B-Instruct, and allows Qwen3-1.7B to reach o1-like reasoning accuracy. These results highlight the power of structured search in decoding, offering an alternative to brute-force sampling or scale-driven gains. Our work demonstrates how thoughtful inference-time strategies can enhance reasoning in SLMs, pointing toward future advances in more efficient and scalable language model deployment.

Speculative decoding (SD) is a widely adopted approach for accelerating inference in large language models (LLMs), particularly when the draft and target models are well aligned. However, state-of-the-art SD methods typically rely on tightly coupled, self-attention-based Transformer decoders, often augmented with auxiliary pooling or fusion layers. This coupling makes them increasingly complex and harder to generalize across different models. We present Budget EAGLE (Beagle), the first, to our knowledge, cross-attention-based Transformer decoder SD model that achieves performance on par with leading self-attention SD models (EAGLE-v2) while eliminating the need for pooling or auxiliary components, simplifying the architecture, improving training efficiency, and maintaining stable memory usage during training-time simulation. To enable effective training of this novel architecture, we propose Two-Stage Block-Attention Training, a new method that achieves training stability and convergence efficiency in block-level attention scenarios. Extensive experiments across multiple LLMs and datasets show that Beagle achieves competitive inference speedups and higher training efficiency than EAGLE-v2, offering a strong alternative for architectures in speculative decoding.

Modern language models (LMs) tend to memorize portions of their training data and emit verbatim spans. When the underlying sources are sensitive or copyright-protected, such reproduction raises issues of consent and compensation for creators and compliance risks for developers. We propose Anchored Decoding, a plug-and-play inference-time method for suppressing verbatim copying: it enables decoding from any risky LM trained on mixed-license data by keeping generation in bounded proximity to a permissively trained safe LM. Anchored Decoding adaptively allocates a user-chosen information budget over the generation trajectory and enforces per-step constraints that yield a sequence-level guarantee, enabling a tunable risk-utility trade-off. To make Anchored Decoding practically useful, we introduce a new permissively trained safe model (TinyComma 1.8B), as well as Anchored_{Byte} Decoding, a byte-level variant of our method that enables cross-vocabulary fusion via the ByteSampler framework (Hayase et al., 2025). We evaluate our methods across six model pairs on long-form evaluations of copyright risk and utility. Anchored and Anchored_{Byte} Decoding define a new Pareto frontier, preserving near-original fluency and factuality while eliminating up to 75% of the measurable copying gap (averaged over six copying metrics) between the risky baseline and a safe reference, at a modest inference overhead.

Speculative decoding accelerates autoregressive language models by using a lightweight drafter to propose multiple future tokens, which the target model then verifies in parallel. DFlash shows that a block diffusion drafter can generate an entire draft block in a single forward pass and achieve state-of-the-art speculative decoding performance, outperforming strong autoregressive drafters such as EAGLE-3. Vanilla DFlash, however, still verifies only a single drafted trajectory per round, potentially limiting its acceptance length. We introduce DDTree (Diffusion Draft Tree), a method that constructs a draft tree directly from the per-position distributions of a block diffusion drafter. Under a fixed node budget, DDTree uses a simple best-first heap algorithm to select the continuations that are most likely to match the target model according to a surrogate defined by the draft model's output. The resulting tree is verified efficiently in a single target model forward pass using an ancestor-only attention mask. Because DDTree builds on DFlash, a leading draft model for speculative decoding, these gains place DDTree among the leading approaches to speculative decoding.

Speculative decoding accelerates LLM inference by drafting a tree of candidate continuations and verifying it in one target forward. Existing drafters fall into two camps with opposite weaknesses. Autoregressive drafters such as EAGLE-3 preserve dependence along each draft path but call the drafter once per tree depth, making drafting a non-trivial share of per-iteration latency. Parallel drafters cut drafter calls by predicting multiple future positions in one forward, but each position is predicted without seeing the others, producing paths the verifier rejects. In this paper, we propose SpecBlock, a block-iterative drafter that combines path dependence with cheap drafting. Each drafter forward produces K dependent positions and we call this a block. The draft tree grows through repeated block expansions. Two mechanisms explicitly carry path dependence to keep later draft positions accurate. Within each block, a layer-wise shift carries the previous position's hidden state into every decoder layer. Across blocks, each new block can start from any position of the previous block, inheriting its hidden state to extend the path. To spend verifier budget where acceptance is likely, a co-trained rank head replaces the fixed top-k tree by allocating per-position branching during drafting. To avoid training the drafter on prefixes it never produces at inference, a valid-prefix mask drops the loss at later positions once an earlier one is wrong. Beyond static drafting, a cost-aware bandit at deployment uses free verifier feedback to update the drafter selectively, only when the expected throughput gain exceeds the update cost. Experiments show that SpecBlock improves mean speedup by 8-13% over EAGLE-3 at 44-52% of its drafting cost, and cost-aware adaptation extends this lead to 11-19%.

Decoding sits between a language model and everything we do with it, yet it is still treated as a heuristic knob-tuning exercise. We argue decoding should be understood as a principled optimisation layer: at each token, we solve a regularised problem over the probability simplex that trades off model score against structural preferences and constraints. This single template recovers greedy decoding, Softmax sampling, Top-K, Top-P, and Sparsemax-style sparsity as special cases, and explains their common structure through optimality conditions. More importantly, the framework makes it easy to invent new decoders without folklore. We demonstrate this by designing Best-of-K (BoK), a KL-anchored coverage objective aimed at multi-sample pipelines (self-consistency, reranking, verifier selection). BoK targets the probability of covering good alternatives within a fixed K-sample budget and improves empirical performance. We show that such samples can improve accuracy by, for example, +18.6% for Qwen2.5-Math-7B on MATH500 at high sampling temperatures.

Large Language Models (LLMs) are increasingly deployed in scenarios demanding ultra-long context reasoning, such as agentic workflows and deep research understanding. However, long-context inference is constrained by the KV cache, a transient memory structure that grows linearly with sequence length and batch size, quickly dominating GPU memory usage. Existing memory reduction techniques, including eviction and quantization, often rely on static heuristics and suffer from degraded quality under tight budgets. In this paper, we propose ARKV, a lightweight and adaptive framework that dynamically allocates precision levels to cached tokens based on per-layer attention dynamics and token-level importance. During a short prefill phase, ARKV estimates the original quantization (OQ) ratio of each layer by computing statistical scores such as attention entropy, variance and kurtosis. During decoding, tokens are assigned to one of three states, Original (full precision), Quantization (low precision), or Eviction, according to a fast heavy-hitter scoring strategy. Our experiments on LLaMA3 and Qwen3 models across diverse long- and short-context tasks demonstrate that ARKV preserves ~97% of baseline accuracy on long-context benchmarks while reducing KV memory usage by 4x, with minimal throughput loss. On short-context tasks, ARKV matches full-precision baselines; on GSM8K math reasoning, it significantly outperforms uniform quantization. These results highlight the practical viability of ARKV for scalable LLM deployment, offering fine-grained, data-driven memory control without retraining or architectural modifications. The source code and artifacts can be found in: https://github.com/Large-scale-Sustainable-Computing-LSC/ARKV

Mixture-of-Experts (MoE) models are bottlenecked by loading routed-expert weights from memory during interactive decoding at small batch sizes. Because most intermediate channels in a routed expert contribute little to any single token, channel selectors predict which channels matter and read only their weights, typically keeping channels above a threshold calibrated on offline data. We show that channel selection needs no calibration data. First, a calibrated threshold does not fix how much a selector reads, because the number of channels passing it varies across inputs; Prox with thresholds calibrated for an 18% read budget actually reads only 12.9% on average across our benchmarks. Second, top-k keeps each token's k highest-scoring channels and needs no calibration data. At the calibrated budget, a selector's top-k version matches or beats its threshold version on the same score in 40 of 43 comparisons at 13% and 18% budgets, including all 23 on generation tasks. The top-k versions also avoid the repetition-loop collapse that calibrated thresholds cause on Gemma 4. Calibrating a threshold thus provides neither of the two things selectors use it for: a known read budget and a consistent accuracy gain at that budget. We furthermore introduce a reader built entirely from the model's weights, which screens channels with a ternary sketch and recovers the skipped ones with a low-rank map. Across three post-trained MoE models, our reader retains 98.1% of full-model accuracy at 17.8% of routed-expert reads, with Triton kernels running MoE blocks 1.60 to 1.90x faster than optimized vLLM at batch one and up to 2.69x at larger batches on the same three models.

Autoregressive language models demonstrate excellent performance in various scenarios. However, the inference efficiency is limited by its one-step-one-word generation mode, which has become a pressing problem recently as the models become increasingly larger. Speculative decoding employs a "draft and then verify" mechanism to allow multiple tokens to be generated in one step, realizing lossless acceleration. Existing methods mainly adopt fixed heuristic draft structures, which fail to adapt to different situations to maximize the acceptance length during verification. To alleviate this dilemma, we proposed OPT-Tree, an algorithm to construct adaptive and scalable draft trees. It searches the optimal tree structure that maximizes the mathematical expectation of the acceptance length in each decoding step. Experimental results reveal that OPT-Tree outperforms the existing draft structures and achieves a speed-up ratio of up to 3.2 compared with autoregressive decoding. If the draft model is powerful enough and the node budget is sufficient, it can generate more than ten tokens in a single step. Our code is available at https://github.com/Jikai0Wang/OPT-Tree.

Large language models (LLMs) exhibit exceptional performance across a wide range of tasks; however, their token-by-token autoregressive generation process significantly hinders inference speed. Speculative decoding presents a promising draft-then-verify framework that reduces generation latency while maintaining output distribution fidelity. Nevertheless, the draft model introduces additional computational overhead, becoming a performance bottleneck and increasing the time to first token (TTFT). Previous approaches to mitigate draft model overhead have primarily relied on heuristics and generally failed to match the quality of the draft language models. To address these challenges, we propose DuoDecoding, a novel approach that strategically deploys the draft and target models on the CPU and GPU respectively, enabling parallel decoding while preserving draft quality. Our method incorporates a hardware-aware optimal draft budget to minimize idle times and employs dynamic multi-sequence drafting to enhance draft quality. Extensive experiments across seven tasks show that DuoDecoding achieves up to 2.61x speedup in generation latency, while reducing TTFT to 83% of that in conventional speculative decoding. The Code is available at https://github.com/KaiLv69/DuoDecoding.

We present Jina-OCR-v1, an end-to-end document parsing model built to serve on low-budget GPUs. It combines the compressed-vision encoder and the 3B mixture-of-experts decoder of DeepSeek-OCR, which activates about 570M parameters per token, with a FastMTP speculative decoding head that shares a single draft block recursively across K=3 prediction steps. Greedy verification makes decoding lossless. Post-training combines instruction alignment, robustness fine-tuning on difficult documents, and GRPO under dense verifiable rewards: deterministic formula, table, and structural checks that award partial credit. The training data mixes cleaned public corpora with targeted synthetic pages. At the default dynamic-resolution setting, Jina-OCR-v1 scores 91.14 on OmniDocBench v1.6 and 83.4 on olmOCR-Bench, and reaches the highest page throughput in our comparison at 2.57 pages per second. On a low-budget GPU such as the NVIDIA L4, FastMTP doubles decoding speed over greedy autoregressive decoding. The model is publicly available at https://huggingface.co/jinaai/jina-ocr-v1.

Serving large language models at long context is bottlenecked by the key-value (KV) cache, which is read in full at every decode step. Attention keys are locally low-rank though globally high-rank: a fixed low-rank sketch shared across pages is provably blind to page-specific directions, while at the same summary size a page's own basis ranks pages and keeps carriers far better. LOCKS gives every page its own rank-r spectral summary (resident, a tenth of the cache at r{=}8 and a twenty-fifth at r{=}2), reconstructs within-page logits, estimates each page's attention mass by log-sum-exp, and attends only the top pages; selection itself reads no candidate keys or values. Selecting on this summary alone stays within about a point of the full cache on long-document QA (LongBench-v1; Llama-3.1-8B), tracks the read-every-key exact-LSE oracle on retrieval-dense RULER down to the smallest budgets, and holds quality furthest into the small-budget regime on long-form reasoning (AIME26, MATH-500; Qwen3-4B), where selectors and eviction-based reasoning compressors both fall away. At a 2048-token budget LOCKS matches FullKV aggregate quality at 100K+ context (GLM-4-9B-Chat-1M) while attending about 2% of the tokens; since the summary is scanned in full each step, the per-step KV read falls by about 10--25times across that rank range, and this halves per-token decode latency (2.0times at 1M tokens on one H200 NVL, measured at r{=}8) against dense attention. LOCKS ships as a drop-in plugin for unmodified vLLM, with batched decode running in full CUDA graphs.

Large language model (LLM)-based agent applications often incur high response time. Speculative decoding is a promising solution to improve the inference efficiency of LLM agents without impacting generation quality. However, state-of-the-art speculative decoding algorithms exhibit substantial speed degradation under large batch sizes, limiting their effectiveness to deploy in real-world agent applications. In this work, we first present a systematic analysis of speculative decoding for LLM agents and identify two dominant factors of speedup degradation: high rejection rate of speculative tokens, and under-utilization of dynamic token budgets.B ased on these observations, we propose AgentSpec, a speculative decoding algorithm that addresses the limitations of existing methods for LLM agents. AgentSpec incorporates structure-isolated drafting that constrains speculation to semantically coherent segments of the agent workflow, reducing the drafts of irrelevant semantic paths and achieving an extremely low rejection rate. Moreover, AgentSpec adopts redundancy-aware budget allocation that exploits agent-level information to better utilize the dynamically-free token budget during the agent inference. We implement and evaluate AgentSpec on five different workloads and four different models from four different LLM families in vLLM. Our results demonstrate the superiority of AgentSpec over state-of-the-arts.

Feed-forward 3D Gaussian Splatting (FF3DGS) reconstructs 3D scenes from images in a single forward pass, making it well-suited to applications that must operate within limited resource budgets. To further improve its practicality, recent budget-controllable FF3DGS methods allow the number of Gaussians to be adjusted at inference, so that the output meets a given storage limit or rendering-latency constraint. However, these methods overlook the cost of inference itself: to produce a representation with the target number of Gaussians, they first predict a dense set of Gaussians and then prune or select from it, requiring nearly constant latency and peak memory regardless of the budget. We therefore propose RLSplat, a resource-linear budget-controllable FF3DGS framework whose representation size, inference latency, and peak memory all decrease linearly as the Gaussian budget is reduced. Instead of pruning or selection, RLSplat learns a Gaussian allocation policy that distributes the budget across local scene regions, so that exactly the requested number of Gaussians is generated. These Gaussians are then refined through resource-linear Gaussian decoding, in which each Gaussian attends only to nearby Gaussians found via an efficient hierarchical neighbor search, so that the decoding cost decreases linearly with fewer Gaussians. In zero-shot evaluation, RLSplat outperforms prior budget-controllable methods at every budget, surpassing the strongest baseline even with 4times fewer Gaussians, while running up to 4.1times faster with 9.6times less peak memory at low budgets.

Speculative decoding (SD) accelerates autoregressive Large Language Models (LLMs) by drafting multiple tokens and verifying them in parallel, but it faces a scaling limitation: increasing the draft budget improves speed only when acceptance remains high and drafting overhead stays low. This ceiling has been difficult to break because prior head-based SD methods face a causality-efficiency dilemma. Autoregressive drafters produce path-conditioned candidates that are effective for tree speculative decoding with higher acceptance length, but their drafting cost grows with tree depth. Bidirectional block-diffusion drafters generate all positions in one pass, but their branch-agnostic marginals can form individually plausible yet mutually inconsistent trees, wasting budget and reducing acceptance. We propose JetSpec, a head-based SD framework that combines one-forward drafting efficiency with branch-wise causal conditioning. JetSpec trains a causal parallel draft head over fused hidden states from the frozen target model, producing candidate trees whose scores align with the target model's autoregressive factorization. This enables JetSpec to convert larger draft budgets into longer accepted prefixes and higher end-to-end speedup. Across math, coding, and chat benchmarks on dense and MoE Qwen3 models, JetSpec consistently outperforms bidirectional-head and tree-based SD baselines. On H100 GPUs, JetSpec achieves up to 9.64x speedup on MATH-500 and 4.58x on open-ended conversational workloads, with further latency gains demonstrated through vLLM integration under realistic serving loads. Our code and models are available at https://github.com/hao-ai-lab/JetSpec.

Reinforcement learning with verifiable rewards (RLVR) improves language-model reasoning, but how these gains relate to inference-time decoding and search remains unclear. Does RL create reasoning the base model lacks, or shift the rollout distribution toward trajectories it can already reach but rarely samples? We study this behaviorally with a Unified Decoding Framework (UDF), which expresses token-level sampling, beam-like search, tree search, and sequence-level resampling as executable policies over a shared budgeted operating space, scored post hoc with pass@k, self-consistency, best-of-N, and first-finish success. Using paired Base/RL checkpoints from SimpleRL-Zoo, we ask whether an RL default-policy curve can be approximated by a structured path of Base operating points. On Math500, AIME, GPQA, and IFEval, the pass@k recovery path follows a Budgeted Operating-Point Transition Rule (BOPTR), N_{Base} approx αN_{RL}^β, with benchmark-conditioned exponents. On Qwen2.5-7B, BOPTR gives the lowest transfer error among the non-oracle rules we test, 3.41 pp (95% CI [2.32, 5.53]); a three-seed replication gives 3.07 pm 0.39 pp. The rule extends to ten models across four families (3.28 to 4.87 pp on checkpoints added after fitting), to four benchmarks it was never fitted on (5.03 pp vs. 4.44 pp in fit), and holds without an RL checkpoint for the target model (4.19 pp) or without RL supervision of any kind (5.08 pp). These results support a qualified internalized-search reading: under the recipe we test, much of the measured RL gain corresponds to a change in sampling efficiency toward operating points the base model can already reach under search. We treat the scaling patterns as descriptive of this recipe and cohort, report where they break down, and use UDF and BOPTR as behavioral diagnostics rather than evidence of parameter-level equivalence.

At each step, decoding one sequence with a large language model rereads the projection weights, whose traffic is fixed, and the key-value (KV) cache, whose traffic grows with context. Activation sparsity trims the first term and KV-cache sparsity the second, yet their reported speedups are hard to compare because each depends on context length and on the dense attention kernel it is measured against. We derive a byte crossover, the context length at which the two savings are equal, together with ideal speedup bounds for each branch and for their composition, from model dimensions and keep ratios alone. We then time both branches and their composition from 2K to 128K tokens on two GPUs after a dense prefill of real text, with dense and sparse modes reading the cache through the same split-K attention kernel. The projection branch leads at short context and the KV branch at long context, with speedups that follow their byte bounds up to fixed kernel costs. Adding these costs, measured in separate sweeps, lets the byte account predict the measured crossings of three keep-ratio pairs, a second model, and a second GPU to within 4.1K tokens. Timing the dense baseline with masked instead of split-K attention inflates the apparent speedup of the same KV policy about fivefold. An attention-scored KV selection answers the same passkey and multi-key placements as dense decoding up to 127K tokens, whereas a KV window misses most of them. Under matched perplexity budgets, activation sparsity composed with this selection decodes 14 to 26% faster than the best single branch on both GPUs. Code is available at https://github.com/js-lee-AI/ByteCross.

Speculative decoding accelerates generation without changing its output, but on vision-language models (VLMs) a self-reinforcing cycle holds it back. Because an autoregressive drafter pays a sequential pass for each drafted token, it must stay small and can ill afford to attend to the image at each pass. Prior work therefore compresses or hides the image, leaving the drafter weakest on the text the image determines. We present GLANCE, a one-pass block drafter that breaks this cycle on an unmodified VLM target. Its block-diffusion head drafts a whole block in one forward pass over the target's already fused vision-language states, reading the multimodal context once, however deep the draft. The target verifies a wide candidate tree in one pass and commits exactly its greedy output. In one production engine at a fixed round budget, GLANCE decodes up to 3.05 times faster than autoregressive decoding and outpaces the production EAGLE3-VL head on average and by about 11% on grounded tasks. An entropy law explains when drafting pays, predicting the longest accepted blocks on grounded tasks, where the target's next-token entropy is lowest. Our code is available at https://github.com/js-lee-AI/GLANCE.

Speculative decoding is an inference-acceleration method for large language models (LLMs) where a small language model generates a draft-token sequence which is further verified by the target LLM in parallel. Recent works have advanced this method by establishing a draft-token tree, achieving superior performance over a single-sequence speculative decoding. However, those works independently generate tokens at each level of the tree, not leveraging the tree's entire diversifiability. Besides, their empirical superiority has been shown for fixed length of sequences, implicitly granting more computational resource to LLM for the tree-based methods. None of the existing works has conducted empirical studies with fixed target computational budgets despite its importance to resource-bounded devices. We present Recursive Speculative Decoding (RSD), a novel tree-based method that samples draft tokens without replacement and maximizes the diversity of the tree. During RSD's drafting, the tree is built by either Gumbel-Top-k trick that draws tokens without replacement in parallel or Stochastic Beam Search that samples sequences without replacement while early-truncating unlikely draft sequences and reducing the computational cost of LLM. We empirically evaluate RSD with Llama 2 and OPT models, showing that RSD outperforms the baseline methods, consistently for fixed draft sequence length and in most cases for fixed computational budgets at LLM.

Inference-time optimization scales computation to derive deliberate reasoning steps for effective performance. While previous search-based strategies address the short-sightedness of auto-regressive generation, the vast search space leads to excessive exploration and insufficient exploitation. To strike an efficient balance to derive the optimal step, we frame the decoding strategy as foresight sampling, leveraging simulated future steps to obtain globally optimal step estimation. Built on it, we propose a novel decoding strategy, named phi-Decoding. To provide a precise and expressive estimation of step value, phi-Decoding approximates two distributions via foresight and clustering. Sampling from the joint distribution, the optimal steps can be selected for exploitation. To support adaptive computation allocation, we propose in-width and in-depth pruning strategies, featuring a light-weight solution to achieve inference efficiency. Extensive experiments across seven benchmarks show phi-Decoding outperforms strong baselines in both performance and efficiency. Additional analysis demonstrates its generalization across various LLMs and scalability across a wide range of computing budgets. The code will be released at https://github.com/xufangzhi/phi-Decoding, and the open-source PyPI package is coming soon.

Speculative decoding accelerates autoregressive generation by verifying a draft block with a target model in parallel. Under standard greedy verification, decoding stops at the first draft token that differs from the target argmax, discarding the remaining target-scored suffix. Although accepting such a mismatch changes the decoding trajectory, it can make a contiguous suffix reusable when its tokens remain target-greedy under the realized prefix. In this paper, we introduce Approximate Speculative Decoding (ASD), a training-free verifier that replaces binary first-mismatch truncation with budgeted longest-prefix selection. ASD accepts selected mismatches subject to a local target-logit regret gate, a per-block exception cap, and a persistent request-level regret budget, then reuses the contiguous target-greedy suffix without additional approximate decisions or target-model forward passes. ASD requires neither a new draft model nor fine-tuning, and exactly reduces to standard greedy verification when the budget is zero. Experiments show that ASD improves fixed-workload throughput by 3.05%--15.26% over matched strict verification and averages a 7.78% gain across seven Qwen3-14B + DSpark-14B tasks. On DeepSeek-V4-Flash (284B) with DSpark it also raises verifier-side acceptance by roughly 10%--16% on GSM8K and MATH-500 in an FP4-to-FP8 compatibility setting. The source code is publicly available at: https://github.com/Kissmetothemoon/ASD

Recent advancements in speculative decoding have demonstrated considerable speedup across a wide array of large language model (LLM) tasks. Speculative decoding inherently relies on sacrificing extra memory allocations to generate several candidate tokens, of which acceptance rate drives the speedup. However, deploying speculative decoding on memory-constrained devices, such as mobile GPUs, remains as a significant challenge in real-world scenarios. In this work, we present a device-aware inference engine named SpecMemo that can smartly control memory allocations at finer levels to enable multi-turn chatbots with speculative decoding on such limited memory devices. Our methodology stems from theoretically modeling memory footprint of speculative decoding to determine a lower bound on the required memory budget while retaining speedup. SpecMemo empirically acquires a careful balance between minimizing redundant memory allocations for rejected candidate tokens and maintaining competitive performance gains from speculation. Notably, with SpecMemo's memory management, we maintain 96% of overall throughput from speculative decoding on MT-Bench, with reduced generation-memory by 65% on single Nvidia Titan RTX. Given multiple constrained GPUs, we build on top of previous speculative decoding architectures to facilitate big-model inference by distributing Llama-2-70B-Chat model, on which we provide novel batched speculative decoding to increase usability of multiple small server GPUs. This novel framework demonstrates 2x speedup over distributed and batched vanilla decoding with the base model on eight AMD MI250 GPUs. Moreover, inference throughput increases remarkably 8x with batch size 10. Our work contributes to democratized LLM applications in resource-constrained environments, providing a pathway for faster and cheaper deployment of real-world LLM applications with robust performance.

Self-taught reasoners (STaRs) enhance the mathematical reasoning abilities of large language models (LLMs) by leveraging self-generated responses for self-training. Recent studies have incorporated reward models to guide response selection or decoding, aiming to obtain higher-quality data. However, they typically allocate a uniform sampling budget across all problems, overlooking the varying utility of problems at different difficulty levels. In this work, we conduct an empirical study and find that problems near the boundary of the LLM's reasoning capability offer significantly greater learning utility than both easy and overly difficult ones. To identify and exploit such problems, we propose HS-STaR, a Hierarchical Sampling framework for Self-Taught Reasoners. Given a fixed sampling budget, HS-STaR first performs lightweight pre-sampling with a reward-guided difficulty estimation strategy to efficiently identify boundary-level problems. Subsequently, it dynamically reallocates the remaining budget toward these high-utility problems during a re-sampling phase, maximizing the generation of valuable training data. Extensive experiments across multiple reasoning benchmarks and backbone LLMs demonstrate that HS-STaR significantly outperforms other baselines without requiring additional sampling budget.

Speculative decoding (SD) accelerates large language model inference by leveraging a draft-then-verify paradigm. To maximize the acceptance rate, recent methods construct expansive draft trees, which unfortunately incur severe VRAM bandwidth and computational overheads that bottleneck end-to-end speedups. While dynamic-depth pruning can reduce this latency by removing marginal branches, it also discards potentially valid candidates, preventing the acceptance rate from reaching the upper bound of dense trees. In this paper, we identify a critical opportunity in resource allocation: the transition from dense to pruned drafting frees up significant computational budget. To break this Pareto tradeoff, we introduce Graft, a compensation framework that couples pruning and retrieval as mutually reinforcing operations. Pruning supplies sufficient budget for retrieval, while retrieval compensates for pruning-induced coverage loss and recovers accepted length. By employing a sequential `prune-then-graft' mechanism, Graft attaches highly predictive retrieved tokens into positions opened by pruning, filling the topological gaps with near-zero overhead. Graft is entirely training-free and lossless. Comprehensive evaluations show that Graft establishes a new Pareto frontier across practical deployment settings, including short-context generation, long-context generation, and large-scale models. On short-context benchmarks, it achieves up to 5.41times speedup and improves average speedup over EAGLE-3 by up to 21.8% on the large-scale Qwen3-235B. We also provide a preliminary exploration of applying Graft to the DFlash-style block drafting paradigm, offering initial evidence and insights for extending grafting beyond autoregressive draft trees.

Using a diffusion model for parallel drafting is a promising approach for speculative decoding. By predicting tokens at multiple future positions in a single forward pass, diffusion drafters substantially reduce drafting latency. However, this shifts the bottleneck to verification: verifying a single sequence limits acceptance length, while verifying large draft trees incurs excessive target-model latency. We identify a key mismatch in existing draft-tree methods: existing diffusion-tree methods rank nodes by the marginal probability, ignoring that verification is prefix-conditioned. As a result, they may verify unreachable descendants of rejected prefixes, increasing latency with limited acceptance gains. To address this, we propose TAPS, a target-aware prefix selection method that turns diffusion marginals into path-conditioned acceptance estimates. TAPS then selects a compact prefix-closed subtree under a fixed verification budget, improving the acceptance-cost tradeoff rather than simply expanding the draft tree. Experiments across diverse datasets and model families demonstrate that TAPS achieves up to 7.9x lossless end-to-end speedup over vanilla autoregressive decoding, outperforming state-of-the-art DFlash and DDTree by 1.36x and 1.74x respectively. Our work is available at https://anonymous.4open.science/r/TAPS-EMNLP2026-53DD

Speculative decoding typically characterizes speedup by draft tokens accepted per verification round. Lossless verification preserves the target model's output distribution and thus expected completion length, making this metric a meaningful efficiency proxy. Relaxed (lossy) verification accepts tokens that lossless verification would reject, potentially altering that distribution. We identify a length inflation phenomenon: relaxed verification lengthens completions, primarily by extending reasoning traces, largely offsetting apparent gains in accepted tokens per round. We evaluate existing training-free relaxed verification rules on twelve target-drafter pairs against lossless verification on identical prompts, measuring completion length, verifier rounds, wall-clock time, and task accuracy. At each rule's most relaxed evaluated setting, relaxed verification on average accepts 1.34times as many tokens per round but produces 1.54times longer completions. For rules with higher acceptance lengths, tokens accepted only under relaxed verification tend to have low target-model probability, contributing to length inflation. For rules that reduce verifier rounds, accuracy typically drops substantially. Further analysis traces some failures to continued reasoning that exhausts the budget needed to deliver a final answer. Overall, accepted tokens per round alone is an unreliable measure of decoding speed under relaxed verification. Completion length and verifier rounds must also be reported to assess actual efficiency.

Pretrained visual representations support image generation, but may not fully preserve the fine-grained details needed for faithful reconstruction. Meanwhile, intermediate encoder layers contain complementary visual details, but learning to fuse them for reconstruction can produce a latent distribution that is difficult to model. Existing fusion methods require empirical tuning of layer selection or staged optimization of fusion and decoding, increasing configuration effort or training complexity. We introduce HiRAE (Hierarchical Representation Autoencoder), which learns a hierarchical fusion framework over the full encoder hierarchy to improve reconstruction fidelity while maintaining compatibility with generative modeling. HiRAE groups encoder layers by depth and learns residual corrections to the deepest representation. Group-wise norm caps bound these corrections relative to the deep anchor, with tighter budgets for shallower groups. Our HiRAE-24 preserves the latent token count and channel dimension. On ImageNet-256, HiRAE-24 reduces reconstruction FID from 0.299 to 0.209 relative to RAEv2 while maintaining competitive guided generation quality. For text-to-image generation, HiRAE-24 improves alignment over RAEv2 on GenEval, DPG-Bench, and GenAI-Bench both before and after supervised fine-tuning. Under the same generator-training and evaluation protocol, post-fine-tuning GenEval increases from 84.86 to 87.70.

Large Language Models (LLMs) have demonstrated strong capabilities across various domains, with recent advancements in challenging reasoning tasks such as mathematics and programming. However, solving reasoning tasks often requires long decoding chains (of thoughts), which incur O(N) time and memory consumption, where N is the chain length. To mitigate O(N) time and memory consumption, existing sparsity-based algorithms propose retaining only the most critical token's intermediate data (i.e., key-value cache) and discarding the rest. However, these existing algorithms struggle with the ``impossible trinity'' of accuracy, time, and memory. For example, the state-of-the-art algorithm, Quest, achieves high accuracy with O(L) time but O(N) memory (L is the cache budget, L ll N). To address this issue, in this paper, we identify a new attention pattern during the decode stage of reasoning tasks, where milestone tokens (analogous to lemmas in mathematical proofs) emerge, are utilized, and then become unimportant afterward. Based on this pattern, we propose a new algorithm named RaaS that identifies and retains milestone tokens only until they are no longer needed, achieving high accuracy with O(L) time and O(L) memory complexity.

As the usage of large language models (LLMs) grows, performing efficient inference with these models becomes increasingly important. While speculative decoding has recently emerged as a promising direction for speeding up inference, existing methods are limited in their ability to scale to larger speculation budgets, and adapt to different hyperparameters and hardware. This paper introduces Sequoia, a scalable, robust, and hardware-aware algorithm for speculative decoding. To attain better scalability, Sequoia introduces a dynamic programming algorithm to find the optimal tree structure for the speculated tokens. To achieve robust speculative performance, Sequoia uses a novel sampling and verification method that outperforms prior work across different decoding temperatures. Finally, Sequoia introduces a hardware-aware tree optimizer that maximizes speculative performance by automatically selecting the token tree size and depth for a given hardware platform. Evaluation shows that Sequoia improves the decoding speed of Llama2-7B, Llama2-13B, and Vicuna-33B on an A100 by up to 4.04times, 3.84times, and 2.37times, and Llama2-70B offloading by up to 10.33times on L40.

Test-time scaling (TTS) boosts LLM reasoning via long chain-of-thought (CoT), but the linear KV-cache growth amplifies the memory-bound bottleneck of LLM decoding. Query-aware page-level sparse decoding can achieve state-of-the-art performance under constrained FLOPs budgets, but is limited by both sequential-dependent page filtering and coarse-grained token selection, hampering serving efficiency and model performance on TTS tasks under high concurrency and long CoT scenarios (consuming even higher runtime than the forward pipeline itself). In this paper, we first find that the current-step query state can be accurately approximated in a unified manner from a short window of recent queries, enabling training-free query-aware sparsity without waiting in the decoding loop. We propose AsyncSpade, an asynchronous framework for efficient TTS built on two core components: (1) a novel light-weight temporal-regressive module that predicts the next-token query state; (2) an asynchronous and disaggregated framework that decouples the KV cache filtering from the auto-regressive decoding loop, overlapping the token-level KV selection with the forward inference computation through asynchronism. To our knowledge, AsyncSpade is the first to eliminate the sequential dependence without sacrificing model performance. We validate the effectiveness of AsyncSpade on common LLM serving setups with an A100 node, where AsyncSpade fully overlaps KV-cache operations with the inference pipeline, achieving theoretical optimal time-per-output-token (TPOT). Specifically, AsyncSpade delivers over 20% reduction on TPOT compared to SoTA baseline (i.e. Quest) and at least 50% TPOT reduction compared to full attention on Qwen3-8B and Qwen3-32B models, while matching or surpassing their accuracy on various TTS benchmarks (AIME-24/25, GPQA-Diamond, MATH-500).

Diffusion Large Language Models (DLLMs) have emerged as a new paradigm of language modeling beyond autoregressive next-token prediction. Thanks to their bidirectional attention mechanism, DLLMs are more capable of capturing the connection of context, and thus show unique advantages in challenges like the famous "reversal curse" or learning under data-constrained scenarios. In addition, taking advantage of their inherent modeling foundations, DLLMs have the great potential of efficient inference with parallel decoding algorithms, which enable multi-token prediction per step. However, the high generation quality often requires the number of decoding steps equal to the sequence length, which performs a one-token-per-step decoding, and existing parallel decoding algorithms, which yield suboptimal decoding paths, bring inference speedup at the cost of non-negligible performance degradation. To overcome this challenge, we introduce Free Draft-and-Verification (FreeDave), a novel fast decoding algorithm tailored for DLLMs that achieves lossless parallel decoding without any model modification or extra modules. Specifically, we propose an algorithm of parallel-decoded candidate generation and verification, which is theoretically guaranteed to use the fewest model forward calls to reproduce the same sequence generated by static decoding when enough computation and memory budget is provided. By extensive evaluations on math reasoning and code generation benchmarks across different DLLMs, FreeDave is proven to boost the inference throughput up to 3.78times without performance degradation.

Diffusion Language Models (DLMs) have recently emerged as a strong alternative to autoregressive language models (LMs). DLMs offer comparable accuracy with faster inference speed via parallel decoding. However, standard DLM decoding strategies relying on high-confidence tokens encounter an inherent information-theoretic bottleneck that restricts decoding progress and ultimately slows generation. We demonstrate both theoretically and empirically that prioritizing high-confidence tokens is inherently inefficient. High-probability tokens carry negligible information and strictly relying on them limits the effective progress made in each decoding round. We prove that the number of decoding rounds must grow linearly with the sample's total information (negative log-likelihood) and inversely with the per-round information budget, establishing a bits-to-rounds principle. We also propose Explore-Then-Exploit (ETE), a training-free decoding strategy that maximizes information throughput and decoding efficiency. ETE combines cross-block decoding with targeted exploration of high-uncertainty tokens to reshape the conditional distribution and trigger cascades of confident predictions. Experiments verify our theoretical bounds and demonstrate that ETE consistently reduces the required number of decoding rounds compared to confidence-only baselines without compromising generation quality.

Reinforcement learning (RL) has become a representative post-training paradigm for LLMs, enabling strong reasoning and agentic capabilities. However, rollout generation remains a dominant latency bottleneck because autoregressive sampling decodes responses sequentially and a small number of long-tailed generations often determine completion time. Speculative decoding (SD) offers a natural way to address this bottleneck, as it is a well-established technique for serving fixed LLMs that reduces latency by rapidly drafting tokens and accepting them through parallel verification while preserving the target-model distribution. However, its practical speedups do not directly carry over to RL rollouts: (i) the evolving target policy makes any fixed drafter increasingly mismatched with the policy's output distribution; and (ii) active batch sizes shrink throughout rollout decoding, shifting decoding from compute-bound to memory-bound regimes where parallel verification can exploit underutilized compute. Therefore, accelerating RL rollouts requires both a drafter that remains effective under long, high-temperature generations from an evolving policy and system-aware use of SD that avoids compute-bound regimes. We present EfficientRollout, a system-aware self-SD framework designed to address this gap for RL rollouts. EfficientRollout induces a quantized drafter from the target model (i.e. self-speculative decoding), keeping it coupled to the evolving policy without separate drafter pretraining or online adaptation. It further coordinates a system-aware SD toggle policy with acceptance-aware draft-length adaptation, enabling speculation only in beneficial regimes while matching the drafting budget to evolving drafter quality. EfficientRollout reduces rollout and end-to-end latency by up to 19.6% and 12.7%, respectively, over an accelerated AR rollout baseline, while preserving final model quality.

When agentic sessions run to a million tokens with many sessions resident at once, the KV cache and the index that ranks it live in host memory, and the scan that ranks all n keys for a top-k step becomes the traffic that bounds decoding. We present Fathom, a key scan in which each query decides how many bits of each key channel to read. The 4-bit K cache is stored channel-major as bit planes, so a prefix of t planes is exactly the channel's t-bit quantizer, and the query spends its bit budget by reverse water-filling over the variance-weighted importance of its channels. At one million tokens on Qwen3-8B a decode step is 1.67x faster in GPU time than with the 136-bit scans of Double Sparsity, Loki and SparQ r=32, and in the same GPU time as SparQ's 68-bit read (r=16) Fathom reads 18% fewer bytes with lower attention error on six of seven model and context settings. On RULER-style tasks every per-token scan matches exact top-k decoding, and on real coding-agent sessions Fathom reaches the step agreement of the most accurate 136-bit scan at 92 bits. The store is the 4-bit K copy a quantized serving stack already holds, and the method is not faster when the index is resident in GPU memory.

Diffusion language models (DLMs) promise parallel text generation, but in practice face an unresolved speed-quality trade-off. Exact decoding fills in only one masked position per model pass, while faster multi-token decoding schemes are lossy. Speculative decoding is the standard lossless acceleration for autoregressive models. However, it does not transfer directly to DLMs, which must choose both what to generate and where. We show that this order is the missing piece: the model's own predictions from the previous pass already give the right token at over 90% of positions, but not which position comes next. We introduce Speculating Positions Order Trajectory (SPOT), a 0.62M-parameter model that reads only the target's output distributions, predicts this order and proposes a tree of candidate continuations, which the target verifies in a single pass by state equality. Coupled-noise match, which shares Gumbel noise between drafter and target, keeps this verification exact at any temperature. This approach also inspired us to introduce SPOT-Sets for dynamic threshold decoding, where a pass commits several positions: SPOT-Sets drafts whole commit sets together with their tokens. Our comprehensive experiments show that SPOT cuts target passes by up to 6.8times and reaches up to 3.4times the throughput of standard decoding on SDAR-8B, matching or beating the calibrated and training-free drafting baselines in every task and budget at T=0 and T=1. In addition, we show that SPOT's learned drafter transfers across models, demonstrating successful transfer across six different DLMs; with a light calibration, the SPOT-Sets drafter transfers as well. On top of dynamic threshold decoding, SPOT-Sets improves the throughput of SDAR-8B by 59% on average over the threshold decoder alone. Project page: https://eliezer318.github.io/SPOT-dLLM/

The Key-Value (KV) cache is crucial for efficient Large Language Models (LLMs) inference, but excessively long contexts drastically increase KV cache memory footprint. Existing KV cache compression methods typically rely on input-side attention patterns within a prompt observation window to estimate token importance during the prefill stage. They fail to preserve critical tokens for future generation since these assessments are not derived from the decoding process. Intuitively, an effective observation window should mirror the decoding-stage queries to accurately reflect which tokens the generation process will attend to. However, ground-truth decoding queries are inherently unavailable during inference. For constructing pseudo queries to approximate them, we find that positional information plays a more critical role than semantic content. Motivated by this insight, we propose decoding-aligned KV cache compression via position-aware pseudo queries (DapQ), a novel and lightweight eviction framework that leverages position-aware pseudo queries to simulate the output tokens, thereby establishing an effective observation window for importance assessment. It aligns closely with the actual generation context and enables precise token eviction. Extensive evaluations across multiple benchmarks and LLMs demonstrate that DapQ achieves superior performance, particularly under strict memory constraints (e.g., up to nearly lossless performance 99.5% on NIAH with 3% KV cache budgets).

Autoregressive decoding remains a primary bottleneck in large language model (LLM) serving, motivating speculative decoding methods that reduce expensive teacher-model invocations by verifying multiple candidate tokens per step. Tree-structured speculation further increases parallelism, but is often brittle when ported across heterogeneous backends and accelerator stacks, where attention masking, KV-cache layouts, and indexing semantics are not interchangeable. We present EAGLE-Pangu, a reproducible system that ports EAGLE-3-style tree speculative decoding to a Pangu teacher backend on Ascend NPUs. EAGLE-Pangu contributes (i) an explicit branch/commit cache manager built on the Cache API, (ii) accelerator-safe tree tensorization that removes undefined negative indices by construction and validates structural invariants, and (iii) a fused-kernel-compatible teacher verification path with a debuggable eager fallback. On 240 turns from MT-Bench and HumanEval-style prompts, EAGLE-Pangu improves end-to-end decoding throughput by 1.27x on average, up to 2.46x at p99, over teacher-only greedy decoding in the fused-kernel performance path. We also provide a fused-kernel-free reference path with structured traces and invariant checks to support reproducible debugging and ablation across execution modes and tree budgets.

Deploying a language model under different computing and latency budgets calls for compact models with different quality-cost trade-offs. Towards this end, elastic spectral state space models provide ordered and temporally decomposed channels that can be truncated, but they are associated with linear time-invariant filters that cannot selectively preserve relevant past information or forget irrelevant information as the context evolves. To address this, we introduce the Elastic Selective Spectral Hybrid (ESSH), which realizes each Hankel spectral channel as an independent recurrent unit using fitted damped rotation modes. It also features an input-dependent decay and write/read gates that make temporal retention and state update input-dependent while preserving channel-wise truncation and a structured recurrence for efficient execution. ESSH combines these selective spectral mixers with sliding-window attention and jointly trains multiple capacities by reducing spectral-channel count and feed-forward width at different rates through a two-rate capacity map with full-model distillation. The resulting models support chunked parallel training and fused recurrent decoding while avoiding computation for discarded channels. At full capacity, ESSH achieves language-modeling quality comparable to similarly sized independently trained models, while smaller exports exhibit a smooth quality-cost trade-off. We validate the effectiveness of the proposed framework using language understanding, retrieval, cross-domain text, and DNA experiments, by assessing quality retention and the trade-off against independently trained and elastic baselines. At the 1.53B model configuration, fused batch-one decoding takes 1.37 ms per token on a B300, providing a 2.14-2.80x speedup over the tested Mamba-2 and Mamba-3 implementations and 3.03x over Transformer++ at matched parameter counts.

Parallel diffusion decoding can accelerate diffusion language model inference by unmasking multiple tokens per step, but aggressive parallelism often harms quality. Revocable decoding mitigates this by rechecking earlier tokens, yet we observe that existing verification schemes frequently trigger flip-flop oscillations, where tokens are remasked and later restored unchanged. This behaviour slows inference in two ways: remasking verified positions weakens the conditioning context for parallel drafting, and repeated remask cycles consume the revision budget with little net progress. We propose COVER (Cache Override Verification for Efficient Revision), which performs leave-one-out verification and stable drafting within a single forward pass. COVER constructs two attention views via KV cache override: selected seeds are masked for verification, while their cached key value states are injected for all other queries to preserve contextual information, with a closed form diagonal correction preventing self leakage at the seed positions. COVER further prioritises seeds using a stability aware score that balances uncertainty, downstream influence, and cache drift, and it adapts the number of verified seeds per step. Across benchmarks, COVER markedly reduces unnecessary revisions and yields faster decoding while preserving output quality.

Speculative decoding accelerates large language model (LLM) inference without compromising output quality. Recent parallel drafting methods further improve single-request performance by decoupling draft length from drafting latency, enabling longer drafts and higher mean accepted tokens (MAT). However, under high request concurrency, long drafts waste substantial computation on rejected tokens, increasing verification cost and potentially making speculative decoding slower than autoregressive decoding. We present D-Cut, an adaptive pruning method that selects draft tokens jointly across the batch and concentrates the verification budget on tokens most likely to be accepted. D-Cut is motivated by two observations. First, acceptance lengths vary considerably across concurrent requests; D-Cut therefore performs cross-request pruning, allocating the verification budget adaptively according to draft confidence. Second, verification cost depends strongly on the deployment environment, including GPU architecture and parallelism strategy; D-Cut incorporates a runtime cost model to adapt its pruning depth to the target environment. Experiments on dense and mixture-of-experts (MoE) models show that, under high concurrency, D-Cut improves the average speedup from \(1.26\times\) to \(1.65\times\), restores acceleration in dense-model configurations where long-draft baselines are slower than autoregressive decoding, and achieves up to \(3.0\times\) speedup over autoregressive decoding on MoE models.

Speculative decoding mitigates the latency of sequential generation in autoregressive Large Language Models (LLMs) by interleaving draft generation with target verification. However, existing parallel drafting backends often suffer from rapid accuracy degradation over long horizons, leading to high rejection rates during verification and suboptimal wall-clock speedups. We observe that drafting errors are not uniformly distributed but typically stem from localized high-uncertainty tokens that destabilize downstream generation trajectories. Motivated by this token error pattern, we propose CURE, a budget-aware dynamic repair tree designed to repair errors at uncertainty focal points without incurring prohibitive tree-verification overheads. Specifically, our method uses predictive confidence margins to dynamically locate candidate error tokens within a block-parallel draft, expands bounded repair paths only at these fragile nodes, and employs a novel repair resynchronization mechanism to realign draft states post-verification. Evaluations on code-generation benchmarks (HumanEval, MBPP, and LiveCodeBench-lite) and mathematical reasoning benchmark (GSM8K) demonstrate that CURE increases the average accepted length by 4.2-7.5% over parallel baselines without repair, translating to an end-to-end speedup of 2.66-3.49times over target-only decoding. Furthermore, we provide a plug-and-play repair module compatible with standard parallel drafting frameworks. We also characterize the trade-off between draft compute and verification efficiency.

As LLMs scale to multi-million-token KV histories, real-time autoregressive decoding under tight Token-to-Token Latency (TTL) constraints faces growing pressure. Two core bottlenecks dominate: accessing Feed-Forward Network (FFN) weights and reading long KV caches. While Tensor Parallelism (TP) helps mitigate the cost of FFN weight reads, it does not scale well for attention. When TP width exceeds the number of KV heads, it leads to inefficient KV duplication, limits parallelism, and constrains batch size. Simultaneously, DRAM reads for long KV histories scale linearly with batch size, further capping efficiency. We introduce Helix Parallelism, a hybrid execution strategy that applies KV parallelism during attention to shard KV caches across GPUs, then reuses the same GPUs for TP in dense LLMs or TPxExpert Parallel (EP) in MoEs during FFN computation. To preserve exact attention behavior, Helix includes a lightweight communication step. To minimize the exposed communication cost, we introduce Helix HOP-B. Helix HOP-B effectively minimizes communication overhead through batchwise overlap, preserving low TTL while improving GPU efficiency. Compared to conventional parallelism approaches, Helix reduces TTL by up to 1.5x at fixed batch sizes and supports up to 32x larger batches under the same latency budget for DeepSeek-R1, pushing forward the throughput-latency Pareto on Blackwell and making real-time inference with ultra-long-sequence practical.

Block drafters such as DFlash draft a block of future tokens in one forward pass, and draft trees such as DDTree verify the prefixes ranked highest by the product of the drafter's per-position distributions. Under tree verification, the expected number of accepted tokens is the probability mass the tree covers. We show that a product misallocates it: when the continuation can follow several futures, it splices them into chimeras, prefixes stitched together from different futures that the target never produces, and gives them as much budget as the futures themselves. Barring ties, no product-ranked tree can hold two futures that disagree in m positions without 2^{m-1}-1 chimeras, however well the drafter is trained, and in practice chimeras make up 27-33% of a 64-node tree. We propose Multi-Future Drafting (MFD), which restores the coupling with a latent variable instead of the chain rule: K learned mode embeddings at the drafter's input draft K coherent blocks in the same pass, and a mixture DDTree lets the verifier pick the future the target takes, losslessly and with K+1 extra vectors. On four Qwen targets from 4B to 27B parameters, at batch size 1, MFD reaches up to 7.8times speedup, 5-8% faster than the best single-mode tree trained identically and, on Qwen3-4B and Qwen3-8B, 9-16% faster than the released DARTree trained on the same prompts.

Speculative decoding accelerates large language model inference by drafting multiple candidate tokens and verifying them in a single forward pass. Candidates are organized as a tree: deeper trees accept more tokens per step, but adding depth requires sacrificing breadth (fallback options) under a fixed verification budget. Existing training-free methods draft from a single token source and shape their trees without distinguishing candidate quality across origins. We observe that two common training-free token sources - n-gram matches copied from the input context, and statistical predictions from prior forward passes - differ dramatically in acceptance rate (~6x median gap, range 2-18x across five models and five benchmarks). We prove that when such a quality gap exists, the optimal tree is anisotropic (asymmetric): reliable tokens should form a deep chain while unreliable tokens spread as wide branches, breaking through the depth limit of balanced trees. We realize this structure in GOOSE, a training-free framework that builds an adaptive spine tree - a deep chain of high-acceptance context-matched tokens with wide branches of low-acceptance alternatives at each node. We prove that the number of tokens accepted per step is at least as large as that of either source used alone. On five LLMs (7B-33B) and five benchmarks, GOOSE achieves 1.9-4.3x lossless speedup, outperforming balanced-tree baselines by 12-33% under the same budget.

Parallel drafters such as DFlash and DSpark reduce proposal latency, but early mismatches limit how much of each block can be accepted. Retaining alternatives introduces parent states and candidate choices that reference-conditioned training does not directly supervise. We introduce TreeSpark, a unified training and inference framework extending the semi-autoregressive DSpark architecture. It constructs parent-conditioned candidate trees from one backbone pass and replays their expansion, pruning, and budgeted packing during training. Survival negative log-likelihood weights retained edges by reference-derived path mass, while a capped coverage hinge promotes target-important alternatives near local selection boundaries. Both objectives reuse reference-chain target distributions without additional target forward passes. We evaluate greedy decoding with Qwen3-4B and Qwen3-8B on nine general benchmarks and six writing domains. Across targets and suites, TreeSpark improves mean acceptance length over DSpark by 23.4% to 31.6% and mean single-request throughput by 15.1% to 25.0%. Batched Qwen3-4B throughput improves by 12.6% to 17.2%. At fixed attention capacity and inference policy, the proposed supervision further improves mean acceptance length by approximately 6.4% and mean system throughput by 4.2% to 5.0% on Qwen3-4B.

Deploying large language models (LLMs) in mobile and edge computing environments is constrained by limited on-device resources, scarce wireless bandwidth, and frequent model evolution. Although edge-cloud collaborative inference with speculative decoding (SD) can reduce end-to-end latency by executing a lightweight draft model at the edge and verifying it with a cloud-side target model, existing frameworks fundamentally rely on tight coupling between the two models. Consequently, repeated model synchronization introduces excessive communication overhead, increasing end-to-end latency, and ultimately limiting the scalability of SD in edge environments. To address these limitations, we propose FlexSpec, a communication-efficient collaborative inference framework tailored for evolving edge-cloud systems. The core design of FlexSpec is a shared-backbone architecture that allows a single and static edge-side draft model to remain compatible with a large family of evolving cloud-side target models. By decoupling edge deployment from cloud-side model updates, FlexSpec eliminates the need for edge-side retraining or repeated model downloads, substantially reducing communication and maintenance costs. Furthermore, to accommodate time-varying wireless conditions and heterogeneous device constraints, we develop a channel-aware adaptive speculation mechanism that dynamically adjusts the speculative draft length based on real-time channel state information and device energy budgets. Extensive experiments demonstrate that FlexSpec achieves superior performance compared to conventional SD approaches in terms of inference efficiency.

Dynamic sparse attention (DSA) accelerates long-context LLM decoding by attending to only the top-K KV blocks relevant to each query, but it introduces a serialized selection-to-attention dependency that emerges as a new latency bottleneck. We present PRR, a speculate-reuse-repair runtime that exploits temporal locality in DSA selections to predict likely blocks, speculate the attention over them while selection is in flight, and incrementally repair missed blocks once the true selected set is known. PRR uses a lightweight EMA-based predictor, a profiling-guided speculation budget that keeps speculative work off the critical path, and a FlashAttention-based repair kernel that folds missed blocks into the partial attention state using online-softmax statistics. Across long-context benchmarks and representative DSA methods, PRR reduces per-token decoding latency by up to 40% while preserving downstream task accuracy. Github: https://github.com/Tianyu9748/Incremental_FlashAttention

The throughput gains of speculative decoding depend on the trade-off between acceptance length and execution latency. We ask when additional acceptance is worth the cost of obtaining it. We establish a comparison protocol spanning training, inference, and evaluation, remeasuring released checkpoints and retraining six drafter recipes on shared target-generated data across ten benchmarks and two target models. Alignment changes the picture: gaps that appear large in published comparisons shrink substantially under shared training and inference conditions. This narrowing motivates two deeper questions: how is acceptance improved at different costs, and when do these gains improve throughput? We find that acceptance can be improved through nearly free objective and context changes or more expensive token-to-token conditioning. DCore illustrates the former route as a strong baseline with little additional measured drafting overhead. The payoff is conditional: task uncertainty and serving conditions affect acceptance returns, execution costs, and the preferred drafter and budget. These findings inform future drafter design and deployment choices.

Recent advancements in Large Language Models (LLMs) have leveraged increased test-time computation to enhance reasoning capabilities, a strategy that, while effective, incurs significant latency and resource costs, limiting their applicability in real-world time-constrained or cost-sensitive scenarios. This paper introduces BudgetThinker, a novel framework designed to empower LLMs with budget-aware reasoning, enabling precise control over the length of their thought processes. We propose a methodology that periodically inserts special control tokens during inference to continuously inform the model of its remaining token budget. This approach is coupled with a comprehensive two-stage training pipeline, beginning with Supervised Fine-Tuning (SFT) to familiarize the model with budget constraints, followed by a curriculum-based Reinforcement Learning (RL) phase that utilizes a length-aware reward function to optimize for both accuracy and budget adherence. We demonstrate that BudgetThinker significantly surpasses strong baselines in maintaining performance across a variety of reasoning budgets on challenging mathematical benchmarks. Our method provides a scalable and effective solution for developing efficient and controllable LLM reasoning, making advanced models more practical for deployment in resource-constrained and real-time environments.

While agents are increasingly spending more resources, today agent cost is mostly measured only after execution. A Budget-Aware Agent (BAGEN) should treat budget as an active control signal, rather than a passive cost metric. We first systematically define budget estimation as internal budgets (from agent computation) and external budgets (from agent actions). We then formalize budget-awareness as progressive interval estimation: at each step of a plan, an agent should predict an upper and lower bound on remaining budget, and alert when completion is unlikely. Scoring with a rollout-replay protocol, we find consistent failure patterns on four environments and five frontier agents: (1) strong agents do not necessarily have strong budget-awareness, with correlation r=0.35. (2) frontier models are consistently over-optimistic, continue spending on tasks that are unlikely to succeed, instead of alerting the user early. (3) budget-aware signal is actionable and trainable. Early stop saves 28-64% tokens on failed trajectories, and SFT+RL strengthens early stop and alert behavior. (4) precise interval calibration remains challenging, with interval coverage capping at 47% after SFT+RL. Project page: https://ragen-ai.github.io/bagen/

Multi-head attention enables Transformers to capture diverse representations, but all attention heads are typically activated for every input, regardless of task complexity. For coarse-grained tasks such as text classification, where relevant information is often global, this fixed allocation can introduce unnecessary computation. We propose BudgetFormer, a Transformer architecture that dynamically allocates attention heads on a per-input basis. The model learns both a head budget and a relevance distribution to select the most informative heads. To support effective head selection, we introduce a training strategy that balances exploration and exploitation. Experiments on text classification tasks show that BudgetFormer reduces FLOPs and memory usage while matching or surpassing the performance of standard multi-head attention. These results highlight adaptive head allocation as an effective approach to improving Transformer efficiency and performance.

Speculative decoding has emerged as a powerful method to improve latency and throughput in hosting large language models. However, most existing implementations focus on generating a single sequence. Real-world generative AI applications often require multiple responses and how to perform speculative decoding in a batched setting while preserving its latency benefits poses non-trivial challenges. This paper describes a system of batched speculative decoding that sets a new state of the art in multi-sequence generation latency and that demonstrates superior GPU utilization as well as quality of generations within a time budget. For example, for a 7.8B-size model on a single A100 GPU and with a batch size of 8, each sequence is generated at an average speed of 5.8ms per token, the overall throughput being 1.1K tokens per second. These results represent state-of-the-art latency and a 2.15X speed-up over optimized regular decoding. Within a time budget that regular decoding does not finish, our system is able to generate sequences with HumanEval Pass@First of 43% and Pass@All of 61%, far exceeding what's feasible with single-sequence speculative decoding. Our peak GPU utilization during decoding reaches as high as 15.8%, more than 3X the highest of that of regular decoding and around 10X of single-sequence speculative decoding.

LLM agents often retry after external validation rejects a candidate, but the interface between validation and the next model call remains underspecified. We introduce VeriHarness, a code-controlled agent loop in which models generate candidates while external validators control acceptance, budgets, and traces. We use it to compare raw diagnostics with feedback that identifies the failure location, observed value, and admissible alternatives. Across 50 paired TextWorld games under a four-call cap, feedback containing all three fields raises terminal success from 14/50 to 36/50 for Qwen2.5-Coder-14B (+44 percentage points) and from 8/50 to 29/50 for Llama-3.1-8B (+42 points). Ablations locate most of the gain in the admissible alternatives: feedback containing only the location and observed value remains near the raw diagnostic baseline. Presenting the complete repair information in prose instead of a keyed JSON record yields nearly the same success, providing no evidence that JSON syntax itself improves repair. The ordering persists across the tested call budgets and one sampled-decoding setting.

With the emergence of wearable devices and other embedded systems, deploying large language models (LLMs) on edge platforms has become an urgent need. However, this is challenging because of their high computational and memory demands. Although recent low-bit quantization methods (e.g., BitNet, DeepSeek) compress weights to as low as 1.58~bits with minimal accuracy loss, edge deployment is still constrained by limited on-chip resources, power budgets, and the often-neglected long latency of the prefill stage. We present TeLLMe, the first table-lookup-based ternary LLM accelerator for low-power edge FPGAs that fully supports both prefill and autoregressive decoding using 1.58-bit weights and 8-bit activations. TeLLMe incorporates several novel techniques, including (1) a table-lookup-based ternary matrix multiplication (TLMM) engine utilizing grouped activations and online precomputation for low resource utilization and high throughput; (2) a fine-grained analytic URAM-based weight buffer management scheme for efficient loading and compute engine access; (3) a streaming dataflow architecture that fuses floating-point element-wise operations with linear computations to hide latency; (4) a reversed-reordered prefill stage attention with fused attention operations for high memory efficiency; and (5) a resource-efficient specialized decoding stage attention. Under a 5~W power budget, TeLLMe delivers up to 25~tokens/s decoding throughput and 0.45--0.96~s time-to-first-token (TTFT) for 64--128 token prompts, marking a significant energy-efficiency advancement in LLM inference on edge FPGAs.

Cross-Layer Transcoders (CLTs) decompose the activations of large language models into human-interpretable features and trace how these features interact across layers to form circuits, making them a key tool for mechanistic interpretability. However, training CLTs effectively and efficiently is hard. Beyond the usual trade-off between sparsity and reconstruction quality, a CLT has to model all layers at once, which adds a further challenge. The common JumpReLU recipe controls sparsity only indirectly through a penalty, while the direct extension of TopK selects a fixed number of latents in every layer. Yet reconstruction difficulty differs widely across layers, and intermediate layers are much harder than others, so giving every layer the same budget is suboptimal. We propose Cross-Layer TopK, which sets a single L0 budget for the whole model and dynamically allocates it across layers during training. In a comparison with JumpReLU and other TopK variants on Qwen3-8B-Base, this gives lower FVU while using 20% fewer active latents than JumpReLU. The better reconstruction carries over to circuit tracing, yielding more complete circuits with higher replacement and completeness scores. The fixed per-token budget of Cross-Layer TopK also makes it easy to pair with sparse decoding, which cuts both computation and communication and gives a 2.9x speedup per training step. We further show that Cross-Layer TopK follows scaling laws in both latent size and sparsity. We have open-sourced the weights of our CLT at https://huggingface.co/ByteDance-Seed/Cross-Layer-TopK-CLT.

Large language models (LLMs) are documented to struggle in settings that require complex reasoning. Nevertheless, instructing the model to break down the problem into smaller reasoning steps (Wei et al., 2022), or ensembling various generations through modifying decoding steps (Wang et al., 2023) boosts performance. Current methods assume that the input prompt is fixed and expect the decoding strategies to introduce the diversity needed for ensembling. In this work, we relax this assumption and discuss how one can create and leverage variations of the input prompt as a means to diversity of thought to improve model performance. We propose a method that automatically improves prompt diversity by soliciting feedback from the LLM to ideate approaches that fit for the problem. We then ensemble the diverse prompts in our method DIV-SE (DIVerse reasoning path Self-Ensemble) across multiple inference calls. We also propose a cost-effective alternative where diverse prompts are used within a single inference call; we call this IDIV-SE (In-call DIVerse reasoning path Self-Ensemble). Under a fixed generation budget, DIV-SE and IDIV-SE outperform the previously discussed baselines using both GPT-3.5 and GPT-4 on several reasoning benchmarks, without modifying the decoding process. Additionally, DIV-SE advances state-of-the-art performance on recent planning benchmarks (Valmeekam et al., 2023), exceeding the highest previously reported accuracy by at least 29.6 percentage points on the most challenging 4/5 Blocksworld task. Our results shed light on how to enforce prompt diversity toward LLM reasoning and thereby improve the pareto frontier of the accuracy-cost trade-off.

Spoken Dialogue Models (SDMs) typically share a speech-token space for generation and synthesis, leading to long autoregressive sequences and costly inference. Lower-rate speech tokenizers shorten sequences but may introduce compression-quality trade-offs and costly replacement without breaking this coupling. To address this limitation, we propose CoSTA, a compatibility-preserving post-hoc framework that decouples generation and synthesis token spaces while retaining the original speech tokenizer and decoder. CoSTA extends the generation token space with compressed representations via mean-initialized vocabulary extension, pronunciation adaptation, and dialogue adaptation, followed by deterministic expansion back to the original speech-token space before synthesis. On bilingual speech-dialogue benchmarks, CoSTA achieves approximately 1.7times expanded speech-token generation speed, lower full-response latency, and higher throughput while preserving response quality. Analyses show more sustained interaction under limited budgets and a practical trade-off among compression, decoding efficiency, and quality. Further experiments demonstrate cross-architecture transfer and complementarity to low-rate tokenization.

Medical language models promise expert clinical support, but the strongest open-source LLMs remain too large for private edge-device deployment, while compact medical models still trail substantially on knowledge-intensive and real-world healthcare tasks. Reasoning models that fit edge memory budgets can produce long reasoning traces, increasing sequential decoding, end-to-end latency, energy consumption, and KV-cache memory pressure. We present MedPsy, a family of text-only 1.7B and 4B medical language models designed for edge deployment, targeting this accuracy-generation-efficiency trade-off. Our recipe combines a synthetic medical data pipeline over biology, medicine, and health seeds with chain-of-thought targets from a 235B medical-focused LLM teacher; a four-stage post-training curriculum comprising two SFT stages followed by two RL stages with hard-sample mining; and an edge deployment evaluation spanning multiple GGUF quantization variants and real-device inference on laptop and smartphone hardware. On seven closed-ended medical benchmarks, MedPsy-4B surpasses MedGemma-1.5-4B by 19.34 points and matches MedGemma-27B while being 6.75x smaller; MedPsy-1.7B outperforms MedGemma-1.5-4B by 11.42 points despite being less than half its size. On HealthBench-Hard, MedPsy-4B surpasses MedGemma-27B by 15.33 points, while MedPsy-1.7B surpasses it by 11.66 points despite being ~16x smaller. Beyond accuracy, MedPsy reduces average response length by 1.7x (1.7B) and 3.2x (4B) versus its Qwen3 backbones, while Q4_K_M quantization preserves the aggregate score within 1 point of BF16 with ~69% lower disk footprint. In real-device inference, MedPsy-4B's shorter generations translate into 75% lower median end-to-end latency and estimated energy per answer than its Qwen3 backbone at matched quantization, while peak RAM remains below 5 GB across our tested laptop and smartphone configurations.

The recently proposed Conformer architecture has shown state-of-the-art performances in Automatic Speech Recognition by combining convolution with attention to model both local and global dependencies. In this paper, we study how to reduce the Conformer architecture complexity with a limited computing budget, leading to a more efficient architecture design that we call Efficient Conformer. We introduce progressive downsampling to the Conformer encoder and propose a novel attention mechanism named grouped attention, allowing us to reduce attention complexity from O(n^{2}d) to O(n^{2}d / g) for sequence length n, hidden dimension d and group size parameter g. We also experiment the use of strided multi-head self-attention as a global downsampling operation. Our experiments are performed on the LibriSpeech dataset with CTC and RNN-Transducer losses. We show that within the same computing budget, the proposed architecture achieves better performances with faster training and decoding compared to the Conformer. Our 13M parameters CTC model achieves competitive WERs of 3.6%/9.0% without using a language model and 2.7%/6.7% with an external n-gram language model on the test-clean/test-other sets while being 29% faster than our CTC Conformer baseline at inference and 36% faster to train.

Test-time scaling (TTS), which involves dynamic allocation of compute during inference, offers a promising way to improve reasoning in large language models. While existing TTS methods work well, they often rely on long decoding paths or require a large number of samples to be generated, increasing the token usage and inference latency. We observe the surprising fact that for reasoning tasks, shorter traces are much more likely to be correct than longer ones. Motivated by this, we introduce First Finish Search (FFS), a training-free parallel decoding strategy that launches n independent samples and returns as soon as any one completes. We evaluate FFS alongside simple decoding, beam search, majority voting, and budget forcing on four reasoning models (DeepSeek-R1, R1-Distill-Qwen-32B, QwQ-32B and Phi-4-Reasoning-Plus) and across four datasets (AIME24, AIME25-I, AIME25-II and GPQA Diamond). With DeepSeek-R1, FFS achieves 82.23% accuracy on the AIME datasets, a 15% improvement over DeepSeek-R1's standalone accuracy, nearly matching OpenAI's o4-mini performance. Our theoretical analysis explains why stopping at the shortest trace is likely to yield a correct answer and identifies the conditions under which early stopping may be suboptimal. The elegance and simplicity of FFS demonstrate that straightforward TTS strategies can perform remarkably well, revealing the untapped potential of simple approaches at inference time.

Large Vision-Language Models (VLMs) have emerged as powerful engines for autonomous GUI agents, yet their deployment is severely constrained by the substantial memory footprint and latency of the Key-Value (KV) cache during long-horizon interactions. While existing cache compression methods have proven effective for LLMs, we empirically demonstrate that they suffer from suboptimal performance in GUI scenarios due to a fundamental misalignment: unlike general visual tasks where attention sparsity varies across layers, GUI attention patterns exhibit uniform high-sparsity across all transformer layers. Motivated by this insight, we propose ST-Lite, a training-free KV cache compression framework tailored for efficient GUI agents that explicitly addresses the dynamic spatio-trajectory dependencies within GUI data streams. ST-Lite introduces a novel dual-branch scoring policy incorporating Component-centric Spatial Saliency (CSS) and Trajectory-aware Semantic Gating (TSG). Specifically, CSS preserves the structural integrity of interactive UI elements by evaluating local neighborhood saliency, while TSG mitigates historical redundancy by dynamically filtering visually repetitive KV pairs within the interaction trajectory. Extensive evaluations demonstrate that with only a 10-20% cache budget, ST-Lite achieves a 2.45x decoding acceleration while maintaining comparable or even superior performance compared to full-cache baselines, offering a scalable solution for resource-constrained GUI agents.

Masked Generative Models (MGMs) enable parallel decoding and achieve strong performance across modalities, but require full-sequence bidirectional transformers at every step, making training costly and degrading quality under low sampling budgets. Existing work improves efficiency via better samplers or cheaper fixed-depth denoisers, but they still allocate a fixed amount of denoiser computation to each refinement step. We introduce Fixed-Point Masked Generative Models (FP-MGMs), which replace part of the denoiser with a fixed-point solver over shared attention layers to enable adaptive depth with fewer parameters. To make it more effective for masked generation, we first introduce a cross-step consistency loss, which aligns hidden representations at neighboring denoising steps and, second, three-state reuse (3SR) which warm-starts the solver using the previous solution by treating differently unchanged, still-masked, and newly revealed tokens respectively. Together, these components define our complete training-to-inference framework for fixed-point masked generation, CoFRe. We also show that pre-trained MGMs can be converted into FP-MGMs with short fine-tuning, avoiding full retraining. Across modalities, CoFRe improves the quality and cost trade-off. On OpenWebText, CoFRe reduces parameters by 38.8\%, training time by 11.5\%, and VRAM by 16.9\%, while improving generative perplexity from 830.8 to 101.8 at a budget of 96 transformer-block forward passes, compared to MDLM. In ImageNette, CoFRe reduces training time by 48.6\% and VRAM by 50.7\%, while improving FID in all sample budgets tested. Overall, CoFRe offers a practical framework for cheaper training and stronger low-budget masked generation.

Sparse Mixture-of-Experts (SMoE) language models achieve strong capability at low per-token compute, yet deployment remains memory- and throughput-bound because the full expert pool must be stored and served. Post-training expert pruning reduces this cost, but most methods focus on which experts to prune within each layer and default to a uniform layer-wise sparsity allocation, even though the allocation can strongly affect performance. We decouple pruning into within-layer expert ranking and across-layer budget allocation, and introduce Expected Speculative Acceptance Proxy (ESAP), a speculative-decoding-inspired, teacher-forced metric that measures how well a pruned model matches the full model. ESAP is bounded and stable, enabling cheap comparison of many candidates without costly autoregressive decoding. Building on ESAP, we propose EvoESAP, an evolutionary searching framework that optimizes a non-uniform layer-wise sparsity allocation under a fixed global budget while holding the within-layer pruning order fixed, making it a plug-and-play method with criteria such as Frequency, EAN, SEER, and REAP. Across 7B--30B SMoE LLMs at 25\% and 50\% sparsity, EvoESAP consistently discovers non-uniform allocations that improve open-ended generation (up to +19.6\% on MATH-500 at 50\% sparsity) while preserving competitive multiple-choice accuracy compared with uniform pruning at the same sparsity.

LLM agents in the ReAct paradigm alternate between reasoning, acting, and observing, but deliberate reasoning is confined to the Thought phase: while the agent serializes an action and waits for the environment, its reasoning is frozen. We identify this recurring interval for Action and Observation as a reasoning idle window and ask whether it can host additional reasoning in parallel that serves future turns. Therefore, we propose Second Thought, a training-free inference framework that forks four auxiliary branches the instant each Thought phase concludes, decodes them concurrently with the main loop, and merges the generated thoughts back when the environment observation arrives. In this way, Second Thought relocates the added reasoning off the main thread's sequential decoding path. Across three agentic benchmarks and three reasoning LLMs, Second Thought lowers the average turn count in all nine (model,benchmark) pairs and reduces main thread decoding in six of them by up to 43% (roughly 20% on average among those settings), while leaving it essentially unchanged in a seventh; Pass@1 shows no significant change in seven of nine pairs and the two significant differences are +12.4 and +10.2 points. Against a compute-matched control that forces an equivalent budget onto the main thread's own reasoning, it attains strictly higher Pass@1 with 1.3 to 3.2 less sequential decoding in all four settings where the control applies.

Discrete Masked diffusion language models generate text by iterative parallel decoding, but few-step decoding suffers from a tradeoff between length and quality: with a fixed step budget, standard methods can generate a short, high-quality output, or they can produce long but repetitive text. Continuous denoising can sidestep this tradeoff by evolving all positions jointly in embedding space, but building such a model from scratch at scale remains an open problem. We show that a pretrained masked DLM can instead be lightly adapted to support continuous embedding-space denoising. Starting from LLaDA-8B-Instruct, we continue-pretrain for only 1,000 steps with Discrete Stochastic Localization (DSL), replacing binary masking with continuous per-token Gaussian noise as a soft mask. The adapted model supports continuous inference that evolves all positions jointly in embedding space and defers hard token commitment to the final step. On zero-shot summarization at low step budgets (<=16 forward passes), DSL-LLaDA-SDE achieves the best ROUGE-1 on all four benchmarks and largely avoids the premature-termination / repetition tradeoff of iterative unmasking. The same adaptation also yields selective noisy-state robustness: the model corrects corrupted tokens while preserving clean ones. Control experiments using standard masked diffusion training with the same compute demonstrate neither behavior.

Recently, dynamic computation methods have shown notable acceleration for Large Language Models (LLMs) by skipping several layers of computations through elaborate heuristics or additional predictors. However, in the decoding process of existing approaches, different samples are assigned different computational budgets, which cannot guarantee a stable and precise acceleration effect. Furthermore, existing approaches generally skip multiple contiguous layers at the bottom or top of the layers, leading to a drastic change in the model's layer-wise representations, and thus a consequent performance degeneration. Therefore, we propose a Unified Layer Skipping strategy, which selects the number of layers to skip computation based solely on the target speedup ratio, and then skips the corresponding number of intermediate layer computations in a balanced manner. Since the Unified Layer Skipping strategy is independent of input samples, it naturally supports popular acceleration techniques such as batch decoding and KV caching, thus demonstrating more practicality for real-world applications. Experimental results on two common tasks, i.e., machine translation and text summarization, indicate that given a target speedup ratio, the Unified Layer Skipping strategy significantly enhances both the inference performance and the actual model throughput over existing dynamic approaches.

Large Language Models (LLMs) that can continually improve beyond their training budgets are able to solve increasingly difficult problems by adapting at test time, a property we refer to as extrapolation. However, standard reinforcement learning (RL) operates over fixed problem distributions and training budgets, which limits extrapolation amidst distribution shift at test time. To address this, we introduce RC, an iterative decoding algorithm that replaces standard autoregressive decoding during both training and inference. RC exploits an asymmetry between the response generation and summarization capabilities of LLMs to construct reasoning chains that consistently improve across iterations. Models trained to use RC can extrapolate and continually improve over reasoning horizons more than an order of magnitude longer than those seen during training. Empirically, training a 4B model with RC using a 16k-token training budget improves performance on HMMT 2025 from 40% to nearly 70% with 0.5m tokens at test time, outperforming both comparably sized models and many larger reasoning LLMs. Finally, we also show that models trained with RC can more effectively leverage existing scaffolds to further scale test-time performance, due to the improved summary-conditioned generation abilities learned through training.

Neural image compression methods have seen increasingly strong performance in recent years. However, they suffer orders of magnitude higher computational complexity compared to traditional codecs, which stands in the way of real-world deployment. This paper takes a step forward in closing this gap in decoding complexity by adopting shallow or even linear decoding transforms. To compensate for the resulting drop in compression performance, we exploit the often asymmetrical computation budget between encoding and decoding, by adopting more powerful encoder networks and iterative encoding. We theoretically formalize the intuition behind, and our experimental results establish a new frontier in the trade-off between rate-distortion and decoding complexity for neural image compression. Specifically, we achieve rate-distortion performance competitive with the established mean-scale hyperprior architecture of Minnen et al. (2018), while reducing the overall decoding complexity by 80 %, or over 90 % for the synthesis transform alone. Our code can be found at https://github.com/mandt-lab/shallow-ntc.

Scaling test-time compute with multi-path chain-of-thought improves reasoning accuracy, but its effectiveness depends critically on the exploration-exploitation trade-off. Existing approaches address this trade-off in rigid ways: tree-structured search hard-codes exploration through brittle expansion rules that interfere with post-trained reasoning, while parallel reasoning over-explores redundant hypothesis paths and relies on weak answer selection. Motivated by the observation that the optimal balance is phase-dependent and that correct and incorrect reasoning paths often diverge only at late stages, we reformulate test-time scaling as a dynamic expand-reduce control problem over a pool of hypotheses. We propose HyPER, a training-free online control policy for multi-path decoding in mixture-of-experts models that reallocates computation under a fixed budget using lightweight path statistics. HyPER consists of an online controller that transitions from exploration to exploitation as the hypothesis pool evolves, a token-level refinement mechanism that enables efficient generation-time exploitation without full-path resampling, and a length- and confidence-aware aggregation strategy for reliable answer-time exploitation. Experiments on four mixture-of-experts language models across diverse reasoning benchmarks show that HyPER consistently achieves a superior accuracy-compute trade-off, improving accuracy by 8 to 10 percent while reducing token usage by 25 to 40 percent.

Vision-language models (VLMs) are increasingly augmented with continuous or latent non-textual tokens intended to support "visual thinking." Despite improved task accuracy, this alone does not show that models actually use these tokens for reasoning -- gains may arise from confounds such as added context length, special-token anchoring, or training-time regularization. We formalize a diagnostic principle, Ablate-to-Validate, for testing whether latent-token content is genuinely utilized, and instantiate it as the Token Replacement Test (TRT), a standardized suite of content-replacement ablations. TRT holds the prompt, image, token budget, and decoding fixed while replacing intermediate tokens with zero, random, first-repeat, or oracle alternatives, isolating whether performance depends on token content or merely on token presence. As a controlled testbed, we study relative depth reasoning with LLaVA-13B and Qwen2.5-VL-3B, training models to predict and consume continuous or discrete depth spans across multiple frozen encoders (SigLIP2, CLIP, DINOv2) and token budgets. We additionally apply TRT to three off-the-shelf visual-thinking systems (Mirage, Mull-Tokens, CoVT) on BLINK, VSP, and CV-Bench. Across all settings, accuracy gains are a misleading proxy for latent-token reasoning: VLMs retain most improvement even when token content is corrupted or replaced, revealing a persistent gap between having a latent channel and using it as an information bottleneck. We recommend TRT as a standard diagnostic alongside accuracy for any method introducing continuous thought tokens.

Leveraging attention sparsity to accelerate long-context large language models (LLMs) has been a hot research topic. However, current algorithms such as sparse attention or key-value (KV) cache compression tend to use a fixed budget, which presents a significant challenge during deployment because it fails to account for the dynamic nature of real-world scenarios, where the optimal balance between accuracy and efficiency can vary greatly. In this paper, we find that borrowing top-p sampling (nucleus sampling) to sparse attention can surprisingly achieve adaptive budgeting. Based on this, we propose Twilight, a framework to bring adaptive sparsity to any existing sparse attention algorithm without sacrificing their accuracy. Empirical results show that Twilight can adaptively prune at most 98% of redundant tokens, leading to 15.4times acceleration in self-attention operations and 3.9times acceleration in end-to-end per token latency in long context LLM decoding.

Large language models (LLMs) now exhibit strong multi-step reasoning abilities, but existing inference-time scaling methods remain computationally expensive, often relying on extensive sampling or external evaluators. We propose a principled strategy that frames reasoning as uncertainty minimisation and operates at the level of individual thoughts rather than tokens. Our method selects, at each reasoning step, the continuation that maximizes the model's self-certainty, a metric computed from its internal predictive distribution. This approach achieves significant improvement with a small number of samples, relies exclusively on model-internal signals, and applies to open-ended questions as opposed to methods like majority voting. Experiments on MATH500 and GSM8K across multiple model sizes demonstrate that thought-level self-certainty maximization consistently outperforms greedy decoding and matches or exceeds self-consistency under comparable token budgets. Cross-linguistic evaluations further indicate that the method transfers robustly beyond high-resource languages. Furthermore, analysis of self-certainty dynamics reveals that correct reasoning trajectories converge early to stable paths, suggesting that early decisions, likely associated with the planning of the reasoning process, are predictive of final accuracy. Building on this result, we show that self-certainty maximisation applied to the early steps can explain most of the performance gain and provide a simple yet efficient inference-time scaling method.

Large language models are increasingly applied to multi-document and long-form inputs, yet long-context inference remains memory- and noise-inefficient. Key-value (KV) caching scales linearly with context length, while external retrieval methods often return lexically similar but causally irrelevant passages. We present S3-Attention, a memory-first inference-time framework that treats long-context processing as attention-aligned endogenous retrieval. S3-Attention decodes transient key and query projections into top-k sparse feature identifiers using lightweight sparse autoencoders, and constructs a CPU-based inverted index mapping features to token positions or spans during a single streaming scan. This design allows the KV cache to be discarded entirely and bounds GPU memory usage by the scan chunk size. At generation time, feature co-activation is used to retrieve compact evidence spans, optionally fused with BM25 for exact lexical matching. Under a unified LongBench evaluation protocol with fixed prompting, decoding, and matched token budgets, S3-Hybrid closely matches full-context inference across multiple model families and improves robustness in several information-dense settings. We also report an engineering limitation of the current prototype, which incurs higher wall-clock latency than optimized full-KV baselines, motivating future kernel-level optimization.

We present Density-Sampled Gaussians (DeG), a novel 3D representation designed to bridge the gap between adaptive rendering primitives and scalable generative modeling. Unlike existing approaches that constrain 3D Gaussians to fixed voxel grids or arrays, DeG models Gaussian centers as samples from a learnable probability density function defined over an octree. This formulation provides a rigorous mathematical framework for adaptive density control: by jointly optimizing the spatial density and Gaussian attributes under rendering supervision, our model naturally concentrates primitives in regions of high geometric complexity. We achieve this via a new render loss contribution gradient that serves as a fully differentiable analogue to the discrete densification and pruning heuristics used in standard Gaussian Splatting. The resulting representation is highly flexible, supporting variable-resolution decoding from a single latent code by simply adjusting the sampling budget. To enable generative synthesis, we train a latent diffusion model on DeG. We identify a critical challenge in applying diffusion to unordered set-structured latents, which can significantly slow convergence, and propose VecSeq, a canonical re-indexing mechanism that anchors latent tokens to a deterministic 3D Sobol sequence. This transforms the ambiguous set-generation problem into a robust sequence modeling task. Extensive experiments demonstrate that our pipeline achieves state-of-the-art quality in single-image-to-3D generation, combining the structural adaptivity of unstructured primitives with the training stability of grid-based methods.

Vision-Language Models suffer severe KV cache pressure at inference, as a single image often encodes into thousands of tokens. Most existing methods exploit token sparsity through token pruning, but permanently discarding visual content causes substantial degradation on fine-grained perception tasks. This motivates a complementary axis, feature sparsity: under a fixed KV cache budget, compressing the channel dimension preserves more visual tokens at the same memory cost. Prior Key channel pruning methods, however, face a structural trade-off: token-wise channel pruning is expressive but unstructured and slow, while head-wise approach is hardware-friendly but less robust. We resolve this with RotateK, a rotation-based structured Key channel pruning framework. RotateK applies an online PCA-based rotation that aligns token-dependent channel importance into a shared low-dimensional subspace, enabling accurate pruning under lightweight head-wise masks; a fused Triton attention kernel operates directly on sparse-channel Keys for efficient decoding. Experiments on two representative VLM backbones show that RotateK consistently outperforms prior Key channel pruning in both accuracy and decoding latency, while joint token-channel pruning improves over token-only baselines at matched KV cache budgets.

Despite unconditional feature inversion being the foundation of many image synthesis applications, training an inverter demands a high computational budget, large decoding capacity and imposing conditions such as autoregressive priors. To address these limitations, we propose the use of adversarially robust representations as a perceptual primitive for feature inversion. We train an adversarially robust encoder to extract disentangled and perceptually-aligned image representations, making them easily invertible. By training a simple generator with the mirror architecture of the encoder, we achieve superior reconstruction quality and generalization over standard models. Based on this, we propose an adversarially robust autoencoder and demonstrate its improved performance on style transfer, image denoising and anomaly detection tasks. Compared to recent ImageNet feature inversion methods, our model attains improved performance with significantly less complexity.

Large Language Models (LLMs) have been widely deployed in a variety of applications, and the context length is rapidly increasing to handle tasks such as long-document QA and complex logical reasoning. However, long context poses significant challenges for inference efficiency, including high memory costs of key-value (KV) cache and increased latency due to extensive memory accesses. Recent works have proposed compressing KV cache to approximate computation, but these methods either evict tokens permanently, never recalling them for later inference, or recall previous tokens at the granularity of pages divided by textual positions. Both approaches degrade the model accuracy and output quality. To achieve efficient and accurate recallable KV cache compression, we introduce ClusterKV, which recalls tokens at the granularity of semantic clusters. We design and implement efficient algorithms and systems for clustering, selection, indexing and caching. Experiment results show that ClusterKV attains negligible accuracy loss across various tasks with 32k context lengths, using only a 1k to 2k KV cache budget, and achieves up to a 2times speedup in latency and a 2.5times improvement in decoding throughput. Compared to SoTA recallable KV compression methods, ClusterKV demonstrates higher model accuracy and output quality, while maintaining or exceeding inference efficiency.

As long-context inference becomes central to large language models (LLMs), attention over growing key-value caches emerges as a dominant decoding bottleneck, motivating sparse attention for scalable inference. Fixed-budget top-k sparse attention cannot adapt to heterogeneous attention distributions across heads and layers, whereas top-p sparse attention directly preserves attention mass and provides stronger accuracy guarantees. Existing top-p methods, however, fail to jointly optimize top-p accuracy, selection overhead, and sparse attention cost, which limits their overall efficiency. We present Double-P, a hierarchical sparse attention framework that optimizes all three stages. Double-P first performs coarse-grained top-p estimation at the cluster level using size-weighted centroids, then adaptively refines computation through a second top-p stage that allocates token-level attention only when needed. Across long-context benchmarks, Double-P consistently achieves near-zero accuracy drop, reducing attention computation overhead by up to 1.8x and delivers up to 1.3x end-to-end decoding speedup over state-of-the-art fixed-budget sparse attention methods.

INT2 KV-cache quantization is attractive for long-context LLM serving, but it remains difficult to make both accurate and deployable. Simple rotations such as Hadamard transforms reduce outliers, but still degrade at INT2 because they are not aligned with downstream attention. We propose OSCAR, an Ultra-low-bit KV Cache quantization method that estimates attention-aware covariance structures offline and uses them to derive fixed rotations and clipping thresholds for quantization. In this way, it aligns KV quantization with the covariance structures that attention actually consumes. More importantly, we not only provide theoretical justification but also develop a fully deployable OSCAR system with a custom INT2 attention kernel that remains compatible with paged KV-cache serving and fused kernel pipelines, enabling seamless integration into modern LLM serving frameworks such as SGLang and vLLM. We evaluate our methods on recent reasoning models with reasoning traces of up to 32k tokens across 5 tasks. On Qwen3-4B-Thinking-2507 and Qwen3-8B, OSCAR reduces the BF16 accuracy gap to 3.78 and 1.42 points, respectively, while naive rotation INT2 collapses to nearly zero. We further scale OSCAR to Qwen3-32B and GLM-4.7 (358B params), where it remains effectively on par with BF16. On long context - RULER-NIAH up to 128K, OSCAR remains robust on both Qwen3 models, while naive rotation INT2 collapses. System-wise, OSCAR reduces KV-cache memory by approximately 8x, improves throughput by up to 7x at large batch sizes under the same memory budget, and accelerates batch-size-1 decoding by up to 3x over BF16 due to reduced memory bandwidth overhead.

Reinforcement learning from verifiable rewards (RLVR) has demonstrated remarkable effectiveness in improving the reasoning capabilities of large language models. As models evolve into natively multimodal architectures, extending RLVR to video understanding becomes increasingly important yet remains largely unexplored, due to the diversity of video task types, the computational overhead of repeatedly decoding and preprocessing high-dimensional visual inputs, and the difficulty of reproducible evaluation across numerous sensitive hyperparameters. Existing open-source RL training frameworks provide solid infrastructure for text and image scenarios but lack systematic optimizations tailored for video modality. In this work, we present EasyVideoR1, a complete and efficient reinforcement learning framework specifically designed for training large vision-language models on video understanding tasks. EasyVideoR1 makes the following contributions: (1) a full video RL training pipeline with offline preprocessing and tensor caching that eliminates redundant video decoding and yields a 1.47 times throughput improvement; (2) a comprehensive, task-aware reward system covering 11 distinct video and image problem types with unified routing and modular extension; (3) a mixed offline-online data training paradigm that combines curated high-quality trajectories with on-policy exploration, benefiting the learning of more challenging tasks; (4) joint image-video training with independently configurable pixel budgets, allowing the two modalities to mutually reinforce each other; and (5) an asynchronous multi-benchmark evaluation framework covering 22 mainstream video understanding benchmarks, with reproduced accuracy closely aligned with officially reported scores.

Resource constraints increasingly determine what can be trained, fine-tuned, and deployed in large language models (LLMs), yet efficiency is often studied through isolated techniques rather than as an interacting system of limits. This survey adopts a constraint-centric perspective and organizes recent progress around three coupled bottlenecks: data efficiency (what to train on), memory efficiency (how to fit training), and compute budget awareness (when and where to spend FLOPs). On the data axis, we review selection and pruning methods that maximize learning per token, ranging from scalable proxy signals based on learning dynamics to gradient- and influence-based scoring, as well as difficulty-aware and curriculum-style strategies. We highlight emerging evidence that different notions of good data dominate in different regimes, implying that optimal subsets depend on the task objective and resource budget rather than being universal. On the systems side, we show that GPU memory, not raw compute, is often the dominant bottleneck in fine-tuning, and that effective scaling requires jointly reducing weight storage, optimizer states, and activation memory rather than optimizing any single component in isolation. Beyond memory, we frame training and inference as compute-governed processes in which optimization, data selection, and decoding must explicitly account for finite FLOP budgets. We review evidence for compute-optimal allocation and stopping rules, where computation should be halted or reallocated once marginal performance gains fall below a budget-dependent threshold. Together, these results unify compute-aware data selection, scaling laws, and adaptive inference under a common principle of resource-conditioned decision-making.

Efficient long-video understanding with vision-language models (VLMs) is often framed as selecting informative frames or visual tokens at a fixed native resolution. We show that per-frame resolution can instead be traded for denser temporal coverage, while front-end decoding latency depends on the size of the candidate pool rather than the final token budget. An empirical study across multiple VLMs and long-video benchmarks yields three findings: dense low-resolution sampling outperforms sparse native-resolution sampling at matched token budgets; resolution-sensitive tasks benefit from selected high-resolution frames; and front-end decoding dominates wall time for hour-long videos. Motivated by these findings, we introduce LoHi, a training-free, single-pass framework that combines a dense low-resolution video stream with sparse high-resolution image frames through the VLM's native video and image pathways. LoHi-Anchor selects high-resolution frames using codec I-frame metadata, while LoHi-SemDiv uses query relevance and visual diversity over CLIP features. Across three long-video benchmarks, LoHi improves average accuracy by 10.6 percentage points over the native-resolution baseline at a matched token budget and by 5.2 percentage points over the strongest prior efficiency method. It also reduces front-end decoding latency by up to 7x on hour-long videos. Project page: https://sixundong.com/projects/lohi

Large language models serve applications with diverse inference budgets, but training and maintaining separate models for each budget incurs substantial costs. Recent Looped Transformers enable flexible inference computation through parameter reuse, but maintaining strong performance across different inference budgets remains an open challenge. We introduce Accordion, a mixture-of-experts-based model that supports five inference budgets through fixed-prefix layer looping. Each budget follows a pre-defined execution path, while the minimum budget retains the unlooped model. Joint pretraining samples one budget per micro-batch and optimizes its next-token prediction loss without additional cross-budget consistency objectives. We further preserve these inference budgets through supervised fine-tuning and reinforcement learning, where parameter-group merging and budget-matched multi-teacher on-policy distillation integrate three domain experts into one model. After 3.6T tokens of pretraining, increasing inference depth improves Accordion-6B-A0.5B's five-category macro-average across 37 tasks by 3.01 percentage points, with the middle budget capturing 89% of this gain using half as many additional layer executions. The final model surpasses each domain expert on its corresponding task-and-budget average, while achieving strong instruction-following performance across all budgets and retaining depth-dependent gains in mathematical reasoning and code generation.

Autoformalization, translating natural-language mathematics into formal proof assistants, is bottlenecked not by translation fluency but by faithfulness: a formal statement can typecheck and be provable, yet still encode a different theorem than the source intended. We introduce Bidirectional Provability Fingerprinting (), a framework that certifies faithfulness by characterizing each candidate through its forward and backward consequence neighborhoods in the ambient theory and matching these against probes derived from the natural-language statement. We further introduce four novel components: (i) Counterfactual Probe Generation (), a contrastive procedure that synthesizes probes targeting specific drift directions; (ii) the Equivalence Spectrum, a continuous faithfulness score that replaces brittle binary verdicts; (iii) Adaptive Probe Budget Allocation (), an information-theoretic budget router; and (iv) Faithfulness-Guided Decoding (), which uses signals as a reward during autoformalization. We prove a drift detection theorem and a PAC-faithfulness result establishing that the equivalence class of a natural language statement is learnable from O(log(1/δ)/varepsilon) probes under mild assumptions. We release , a benchmark of 2{,}183 NL/Lean~4 pairs with controlled drift labels across six subfields of mathlib4. \,+\, detects 89.6% of drifted formalizations at a 3.0% false-positive rate-against 41.2% for typecheck and 63.3% for LLM-judge baselines, and reduces the rate at which a state-of-the-art autoformalizer emits drifted statements by 47%. https://pmlrbd.github.io/BPF/

Recent deep-thinking large language models often reason extensively to improve performance, but such lengthy reasoning is not always desirable, as it incurs excessive inference costs with disproportionate performance gains. Controlling reasoning length without sacrificing performance is therefore important, but remains challenging, especially under tight thinking budgets. We propose budget guidance, a simple yet effective method for steering the reasoning process of LLMs toward a target budget without requiring any LLM fine-tuning. Our approach introduces a lightweight predictor that models a Gamma distribution over the remaining thinking length during next-token generation. This signal is then used to guide generation in a soft, token-level manner, ensuring that the overall reasoning trace adheres to the specified thinking budget. Budget guidance enables natural control of the thinking length, along with significant token efficiency improvements over baseline methods on challenging math benchmarks. For instance, it achieves up to a 26% accuracy gain on the MATH-500 benchmark under tight budgets compared to baseline methods, while maintaining competitive accuracy with only 63% of the thinking tokens used by the full-thinking model. Budget guidance also generalizes to broader task domains and exhibits emergent capabilities, such as estimating question difficulty. The source code is available at: https://github.com/UMass-Embodied-AGI/BudgetGuidance.

Long-video language models cannot look at every frame: an hour sampled once per second is 3,600 images, and a system keeps only a small fixed slice of that pool. Which frames survive that slice is usually treated as a preprocessing detail; we test whether it should be. Published selectors make the comparison hard because they change the frame scorer, the prompt boundary, the resolution policy, and the answering model all at once. We hold each fixed and vary one decision at a time: selection, spatial compression, and reinvestment of the savings, across six training-free selection rules, three long-video benchmarks, and two answering models. Selection is the largest single lever: on LongVideoBench's hour-long bin, eight query-selected frames beat sixteen uniformly spaced ones by 6.9 points, and Orthogonal Matching Pursuit, an unmodified decades-old sparse-approximation algorithm, matches or comes within a point of every purpose-built selector we compare it against, across all three benchmarks. Compression is close to free: halving each frame's spatial budget at fixed timestamps costs at most 0.44 points. Reinvestment is where that budget turns back into accuracy: spending the freed tokens on twice as many compressed frames, at a measured cost no higher than the original eight, returns a further two to three points; compression only pays off once its savings are spent this way. Along the way, an implementation bug in our own AKS baseline and a 0.07 to 3.74 point gap between two harnesses running the same published rules at the same budget show why these comparisons need to happen inside one controlled harness rather than across papers.

Computationally intensive decoding procedures--including search, reranking, and self-critique--can improve the quality of language model (LM) outputs in problems spanning code generation, numerical reasoning, and dialog. Existing work typically applies the same decoding procedure for every input to an LM. But not all inputs require the same amount of computation to process. Can we allocate decoding computation adaptively, using more resources to answer questions whose answers will be harder to compute? We present an approach that predicts the distribution of rewards given an input and computation budget, then allocates additional computation to inputs for which it is predicted to be most useful. We apply this approach in two decoding procedures: first, an adaptive best-of-k procedure that dynamically selects the number of samples to generate as input to a reranker; second, a routing procedure that dynamically responds to a query using a decoding procedure that is expensive but accurate, or one that is cheaper but less capable. Across a suite of programming, mathematics, and dialog tasks, we show that accurate computation-allocation procedures can be learned, and reduce computation by up to 50% at no cost to response quality, or improve quality by up to 10% at a fixed computational budget.

Selecting intermediate features from a frozen diffusion model for discriminative tasks is non-trivial because the useful signal is distributed across the denoising trajectory, network depth, and channels. In practice, this requires choosing a timestep, a block, and a channel subset. Existing methods either fix these choices by hand or optimize only a few extraction hyperparameters, most often the timestep, then reuse the same rule across inputs and budgets. We introduce Policy-Mask, a budget-conditioned policy trained on the downstream objective that jointly selects all three under a prescribed channel budget. Without normalization, the channel-score contribution grows with the number of selected channels and dominates the timestep and block updates. We normalize it by subset cardinality, keeping the three policy heads on comparable scales and stabilizing joint training. At matched budgets, Policy-Mask improves over fixed extraction and feature-filter baselines across segmentation, classification, and semantic correspondence, reaching 38.9 mIoU on ADE20K, 75.4% top-1 on CIFAR-100 at 128 dimensions, and 76.3 PCK@0.1 on SPair-71k. Component-wise ablations show that block selection materially affects downstream accuracy, and the highest recorded channel-budget score occurs at 256 retained channels, identifying a practical operating point for compact diffusion representations. We further adapt Policy-Mask to diffusion transformers, including PixArt-Σ and SD3.5 MMDiT, by replacing U-Net channels with image-token hidden dimensions.

Multi-image vision-language models (VLMs) incur substantial computation when processing long visual sequences. Compressing these sequences must preserve complementary evidence across images and keep tokens from different images separate. Global selection can underrepresent question-relevant images, and pruning in early decoder layers risks removing relevant content before sufficient multimodal processing, yet existing methods spend most of the budget on these early layers. We propose SIEVE, a training-free, single-pass framework that jointly allocates visual tokens across images and decoder depth. SIEVE first pools each image into a compact spatial grid to reduce early processing cost. It then progressively selects representatives using prompt attention under equal per-image quotas and aggregates unselected tokens into similar representatives from the same image. These operations keep every image represented and its tokens separate from other images, while refining the retained representations. A closed-form schedule sets retention counts to target a prescribed average token budget across decoder layers. Across four VLM backbones, SIEVE achieves the highest average at a 6.25% budget on five multi-image benchmarks and four single-image benchmarks, and the highest mean on two video benchmarks. On LLaVA-OneVision-7B, it outperforms the strongest baseline by 9.96 and 9.03 points at 6.25% and 12.5% budgets, and at 12.5% it exceeds the uncompressed model by 2.01 points with 6.6x fewer prefill FLOPs.

Omnimodal large language models (Omni-LLMs) encode audio and visual streams into temporally interleaved token sequences for multimodal reasoning. However, processing long audio-visual token sequences incurs substantial prefill costs. Existing compression methods have made progress, but often overlook temporal changes in audio-visual semantic relevance. Motivated by temporal variation and local continuity, we propose OmniRoute, a training-free, two-stage compression framework. First, Temporal Evidence-Guided Budgeting (TEGB) derives chunk-wise modality preferences and initial leading-modality budgets from semantic relevance and local content variation. Second, Budget-Constrained Semantic Compression (BCSC) compresses the leading modality and then calibrates the follower's retention target using the actual retained fraction. For video, it combines spatiotemporal grouping with query-guided selection; for audio, it selects tokens based on encoder attention and query relevance, then merges residual tokens into context anchors under visual guidance. Experiments on four representative benchmarks demonstrate a better trade-off between inference efficiency and performance than competitive baselines. The code and interface will be released to facilitate further research.

When audio, video, and text disagree, accuracy alone does not show whether to acquire another source or abstain. We study these choices in a controlled three-score benchmark: a policy observes two signed scores, may request the third at a cost, and can abstain. The primary reward is mechanism-specific: abstention is correct only for one designated ambiguity mechanism and is penalized under mixed corruption. Matched controls show that a threshold policy matches always-request decisions with fewer requests; its advantage over always-answer fusion depends on the reward assigned to that ambiguity. On a partially held-out synthetic split, the threshold policy reaches 0.789 +/- 0.006 targeted decision accuracy and 0.481 +/- 0.014 utility across 83 seeds. A three-score majority reference reaches 0.626 +/- 0.008 and 0.252 +/- 0.016, but uses more information. In a matched-budget test, a train-only value selector improves utility over no-query and matched-random policies at 10% and 25% budgets, while pair uncertainty has higher utility at every budget. At 50% and 63.7% budgets, the selector lowers utility despite slightly higher non-ambiguous accuracy. If all abstentions are scored incorrect, majority outranks the threshold policy in utility. At a central temporal setting, full-trace controls match the neural models while position perturbations separate them. On held-out-actor emotion clips, eight-frame fusion has opposite-signed accuracy differences for two encoder pairs, with both actor intervals containing zero; matched-request routing gains are small and uncertain. These results separate full-modality accuracy from pre-request selection value and show that selection value depends on budget and the observed-pair ranking.

Memory is increasingly central to Large Language Model (LLM) agents operating beyond a single context window, yet most existing systems rely on offline, query-agnostic memory construction that can be inefficient and may discard query-critical information. Although runtime memory utilization is a natural alternative, prior work often incurs substantial overhead and offers limited explicit control over the performance-cost trade-off. In this work, we present BudgetMem, a runtime agent memory framework for explicit, query-aware performance-cost control. BudgetMem structures memory processing as a set of memory modules, each offered in three budget tiers (i.e., Low/Mid/High). A lightweight router performs budget-tier routing across modules to balance task performance and memory construction cost, which is implemented as a compact neural policy trained with reinforcement learning. Using BudgetMem as a unified testbed, we study three complementary strategies for realizing budget tiers: implementation (method complexity), reasoning (inference behavior), and capacity (module model size). Across LoCoMo, LongMemEval, and HotpotQA, BudgetMem surpasses strong baselines when performance is prioritized (i.e., high-budget setting), and delivers better accuracy-cost frontiers under tighter budgets. Moreover, our analysis disentangles the strengths and weaknesses of different tiering strategies, clarifying when each axis delivers the most favorable trade-offs under varying budget regimes.

Evaluation budgets in agentic retrieval-augmented generation span questions, search trajectories, and repeated answers. We measure allocation precision, reading efficiency, and cost boundaries using a retrieval-feedback comparison on HotpotQA and MuSiQue. At 34.14--34.39M model tokens, broader question coverage lowers standard error by 33\% versus five reads and 12.6\% versus three trajectories. Archived nested and Q-only forecasts predict these allocations within 4.0\% and 3.5\%, respectively. Depth subsets establish no clear forecasting advantage beyond the two-trajectory audit. One-read variance penalties relative to the fitted optimum at the same token budget are 0--9.9\%, with substantial Pro uncertainty. Under recorded model fees, more questions beat more trajectories at search prices of \$0--1 per 1,000 requests; question-versus-read fee rankings remain unresolved. Temperature zero cuts answer disagreement from 14.3\% to 3.4\% while comparison precision stays similar. \par\medskip\noindentKeywords: Agentic RAG; Evaluation budget; Generalizability theory; Repeated sampling.

LLM-based agents show strong potential for long-horizon reasoning, yet their context size is limited by deployment factors (e.g., memory, latency, and cost), yielding a constrained context budget. As interaction histories grow, this induces a trade-off between retaining past information and staying within the context limit. To address this challenge, we propose Budget-Aware Context Management (BACM), which formulates context management as a sequential decision problem with a context budget constraint. It enables agents to assess the available budget before incorporating new observations and decide when and how much of the interaction history to compress. We further develop BACM-RL, an end-to-end curriculum-based reinforcement learning approach that learns compression strategies under varying context budgets. Experiments on compositional multi-objective QA and long-horizon web browsing benchmarks show that BACM-RL consistently outperforms prior methods across model scales and task complexities, achieving over 1.6times gains over strong baselines in high-complexity settings, while maintaining strong advantages as budgets shrink, where most methods exhibit a downward performance trend.

Coreset selection picks a representative subset of the labeled training set to make training cheaper. However, it is usually evaluated by downstream accuracy at a fixed subset size, ignoring both the time spent selecting the subset and the training recipe behind each reported number. We introduce an end-to-end benchmark that standardizes downstream training and charges selection and training to the same auditable wall-clock budget, spanning 4 datasets from CIFAR-10 to ImageNet-1K, 11 selectors, 5 fractions, and 3 seeds, with over 1,500 released runs. Repeated-sampling work has shown that budget-aware evaluation already favors random strategies. Our two budget studies test whether that verdict survives when every selector is granted its most favorable operating point. Across eight wall-clock budget anchors on each of CIFAR-10 and Tiny ImageNet, no anchor is won by a sophisticated selector: every winner is class-balanced random sampling, repeated random sampling, or full-data training. In fixed-budget duels on ImageNet-1K, training on all data for fewer epochs beats every selection strategy we probe while also costing the least. A per-dataset cost audit shows that selection cost is dominated at every scale by a fixed full-dataset scan, so it cannot be amortized away by selecting a smaller fraction, and its absolute size does not extrapolate from one dataset to another. We further quantify when selection does pay back through subset reuse, and document 9 correctness fixes to a widely used codebase, one of which shifts a standard Herding baseline by nearly 6 points. Selection time is not free preprocessing, and an evaluation that ignores it measures the wrong quantity.

Emerging Omni-modal Large Language Models (OmniLLMs) enable unified understanding of text, audio, and video, but their long audio-video token sequences introduce substantial memory and inference costs. Existing compression methods mainly focus on selecting important tokens under fixed budgets, leaving the preceding budget-allocation problem underexplored. We show that direct query-to-audio/video similarity is unreliable for inter-modal budget allocation, and that uniform intra-modal budgets can miss key evidence while retaining redundant content. To address these limitations, we propose OmniDelta, a training-free, skill-driven framework that couples intent-aware inter-modal allocation with content-aware intra-modal allocation. OmniDelta first constructs audio and video skill pools to shift the fixed retained-token budget according to query demand, then reallocates modality budgets over audio segments and video frames using local complexity and temporal redundancy. The resulting local budgets can be combined with existing pruning strategies, preserving the total retained-token ratio while changing where the budget is spent. Experiments on four audio-video benchmarks with two Qwen2.5-Omni models show that OmniDelta establishes a new accuracy-efficiency Pareto frontier across pruning ratios. At 25% token retention on Qwen2.5-Omni-7B, OmniDelta reduces GPU memory by 22.0% and achieves a 1.64x end-to-end speedup over full-token inference.

Is the same data equally useful to every learner? The question is whether the value of a training sample belongs to the sample or arises in its interaction with the learner. We test it with a selection experiment. Under a per-class budget, two kinds of samples compete for the slots: easy samples, which a short training run learns first, and coverage samples, which spread over the class in feature space. Easy samples win at small budgets and coverage samples at large ones, and the budget at which the winner flips, the crossover m0, can be measured. We ask whether m0 is set by the data or by the learner. To separate the two, we freeze the selected subsets and change only the learner that trains on them. On low-resolution ImageNet-100, doubling the width of a ResNet-18 moves the crossover from 57 to 85 samples per class on the same subsets. Enlarging the input grid from 32 to 64 pixels at fixed image information moves it from 85 to 57, and a single stride change reproduces or removes this shift. The grid shift shrinks as width grows. Under the native ImageNet-1k protocol, a 4x range of width moves the boundary by at most a few samples per class. On a Vision Transformer, coverage wins at every measured budget, even with easy subsets that work on the convolutional learner. The budget at which one kind of sample stops being the more useful kind therefore belongs to the learner, and the preferred regime of a selection strategy has to be read against the learner that will use the data.

LLM search agents increasingly rely on tools at inference time, but their trajectories are often constrained by hard limits on both tool calls and generated tokens. Under such dual budgets, better answers require not only stronger models, but also explicit control over which search action should receive the next budget unit and when the accumulated evidence is sufficient to commit a final answer. We study this problem in multi-hop question answering (QA) and formulate it as two-stage inference-time budget control. At search time, our controller assigns each feasible action a task-level Value-of-Information (VOI) score, defined as an operational estimate of marginal task value per unit budget under the current search state and remaining dual budget, and uses this score to choose among retrieval, decomposition, and answer commitment. After search, a selective evidence-grounded finalizer compares the trajectory answer with a refined candidate and rewrites only when the residual error appears to be a low-risk answer-form error. Across four multi-hop QA benchmarks, three LLM backbones, and four budget levels, the method yields positive aggregate gains over four audited baselines under the same hard dual-budget protocol. Ablations show that search-time budget control, especially budget-dependent penalty, provides the main performance gain, while answer-time control helps mainly when the retrieval path is already adequate. These results suggest that inference-time budget control for LLM search agents should govern both how budget is spent during search and how the final answer is committed.

Test-time scaling has become a dominant paradigm for improving LLM agent reliability, yet current approaches treat compute as an abundant resource, allowing agents to exhaust token and tool budgets on redundant steps or dead-end trajectories. Existing budget-aware methods either require expensive fine-tuning or rely on coarse, trajectory-level heuristics that cannot intervene mid-execution. We propose the Budget-Aware Value Tree (BAVT), a training-free inference-time framework that models multi-hop reasoning as a dynamic search tree guided by step-level value estimation within a single LLM backbone. Another key innovation is a budget-conditioned node selection mechanism that uses the remaining resource ratio as a natural scaling exponent over node values, providing a principled, parameter-free transition from broad exploration to greedy exploitation as the budget depletes. To combat the well-known overconfidence of LLM self-evaluation, BAVT employs a residual value predictor that scores relative progress rather than absolute state quality, enabling reliable pruning of uninformative or redundant tool calls. We further provide a theoretical convergence guarantee, proving that BAVT reaches a terminal answer with probability at least 1-ε under an explicit finite budget bound. Extensive evaluations on four multi-hop QA benchmarks across two model families demonstrate that BAVT consistently outperforms parallel sampling baselines. Most notably, BAVT under strict low-budget constraints surpasses baseline performance at 4times the resource allocation, establishing that intelligent budget management fundamentally outperforms brute-force compute scaling.

Recent work has demonstrated that pre-training in-domain language models can boost performance when adapting to a new domain. However, the costs associated with pre-training raise an important question: given a fixed budget, what steps should an NLP practitioner take to maximize performance? In this paper, we view domain adaptation with a constrained budget as a consumer choice problem, where the goal is to select an optimal combination of data annotation and pre-training. We measure annotation costs of three procedural text datasets, along with the pre-training costs of several in-domain language models. The utility of different combinations of pre-training and data annotation are evaluated under varying budget constraints to assess which combination strategy works best. We find that for small budgets, spending all funds on annotation leads to the best performance; once the budget becomes large enough, however, a combination of data annotation and in-domain pre-training yields better performance. Our experiments suggest task-specific data annotation should be part of an economical strategy when adapting an NLP model to a new domain.

Large language models (LLMs) can achieve strong reasoning performance with sufficient computation, but they do not inherently know how much computation a task requires. We study budgeted inference-time reasoning for multiple tasks under a strict global token constraint and formalize it as a Ordered Stochastic Multiple-Choice Knapsack Problem(OS-MCKP). This perspective highlights a meta-cognitive requirement -- anticipating task difficulty, estimating return over investment (ROI), and allocating computation strategically. We propose ROI-Reasoning, a two-stage framework that endows LLMs with intrinsic, budget-aware rationality. In the first stage, Meta-Cognitive Fine-Tuning teaches models to predict reasoning cost and expected utility before generation, enabling explicit solve-or-skip decisions. Next, Rationality-Aware Reinforcement Learning optimizes sequential decision making under a hard token budget, allowing models to learn long-horizon allocation strategies. Across budgeted mathematical reasoning benchmarks, ROI-Reasoning consistently improves overall score while substantially reducing regret under tight computation budgets.

Weight-space model merging is usually formulated as an algebraic operation on checkpoints, yet at LLM scale the limiting resource is often the set of expert weights that must be read. We introduce MergePipe, a budget-aware execution layer that casts LLM merging as an expert access-set problem: given a merge operator and a checkpoint family in a shared weight coordinate system, choose which expert delta blocks to access under an explicit I/O budget. MergePipe indexes parameter blocks, builds deterministic access plans, and executes the induced budgeted merge with replayable manifests. The plan is budget-sound by construction and recovers the full-read merge at full budget; for fixed-coefficient additive operators, the omitted-update error is bounded by the norm of omitted deltas. Across Qwen and Llama merging workloads, MergePipe reduces expert-read I/O by up to an order of magnitude and achieves up to 11times speedups. Representative budget sweeps show O(10^{-3}) parameter deviation from full-read merges and no monotonic degradation on downstream benchmarks.

Large Language Models (LLMs) have achieved remarkable success in complex reasoning tasks, but their inference remains computationally inefficient. We observe a common failure mode in many prevalent LLMs, overthinking, where models generate verbose and tangential reasoning traces even for simple queries. Recent works have tried to mitigate this by enforcing fixed token budgets, however, this can lead to underthinking, especially on harder problems. Through empirical analysis, we identify that this inefficiency often stems from unclear problem-solving strategies. To formalize this, we develop a theoretical model, BBAM (Bayesian Budget Allocation Model), which models reasoning as a sequence of sub-questions with varying uncertainty, and introduce the E^3 metric to capture the trade-off between correctness and computation efficiency. Building on theoretical results from BBAM, we propose Plan-and-Budget, a model-agnostic, test-time framework that decomposes complex queries into sub-questions and allocates token budgets based on estimated complexity using adaptive scheduling. Plan-and-Budget improves reasoning efficiency across a range of tasks and models, achieving up to +70% accuracy gains, -39% token reduction, and +187.5% improvement in E^3. Notably, it elevates a smaller model (DS-Qwen-32B) to match the efficiency of a larger model (DS-LLaMA-70B)-demonstrating Plan-and-Budget's ability to close performance gaps without retraining. Our code is available at anonymous.4open.science/r/P-and-B-6513/.