RobustRouteBench

RobustRouteBench evaluates request-level model selection under quality and cost trade-offs across General Text, Multimodal, and Chinese. It accompanies Section 6.3 of Toward Collective Model Intelligence: Characterizing and Coordinating Models.

Technical report · Code · Data fields · Multimodal images · Source attribution and terms

Dataset overview

Scenario / config	Models	Train	Validation	Standard test	Robust test
General Text / text	63	3,000	500	3,688	2,894
Multimodal / multimodal	45	4,000	500	4,600	1,900
Chinese / chinese	10	7,590	500	3,900	2,000
Total	Scenario-specific	14,590	1,500	12,188	6,794

The dataset contains 35,072 request instances, including 18,982 test requests. Each scenario has its own candidate pool. The test populations above are exactly those used in the technical report.

train is used for fitting selectors, validation for configuration selection, and test for reporting results. Each validation split contains 400 standard and 100 robust requests. The supplied test split is ready for evaluation: use all its rows directly. Standard and robust results are slices of that same test split, identified by evaluation_type.

Quick start

from collections import Counter
from datasets import load_dataset

ds = load_dataset(
    "SinapisAI/RobustRouteBench", "text",
    revision="03bd3da2c3bb4a998f73d285c8853c721697e3cb",
)
print({split: len(rows) for split, rows in ds.items()})
# {'train': 3000, 'validation': 500, 'test': 6582}
print(Counter(ds["test"]["evaluation_type"]))
# Counter({'standard': 3688, 'robust': 2894})

standard = ds["test"].filter(lambda row: row["evaluation_type"] == "standard")
robust = ds["test"].filter(lambda row: row["evaluation_type"] == "robust")

Use "multimodal" or "chinese" for the other scenarios. Their test sizes are 6,500 and 5,900, respectively.

Data fields and units

Each row contains one request and aligned outcomes for its candidate models.

Field	Meaning
query_id	Request identifier, shared with the prediction and media files.
scenario, published_split	Scenario and split membership (train, validation, or test).
query_text, input_json	Full request text and structured input.
evaluation_type, variant_type	Standard/robust designation and transformation type.
domain, source_dataset, source_sample_id, base_query_id	Task and source identifiers.
model_ids	Candidate-model order for all outcome arrays.
performance	Source-task utility in [0, 1]; unsuccessful source outcomes are zero.
cost_cny	Cost in CNY per request.
latency_ms	Recorded latency in milliseconds.
embedding_model, embedding	Encoder identity and the shared 1,024-dimensional feature vector.
media_json	Media archive, image paths, and image provenance.
provenance_json, source_split, is_validation	Source metadata and split annotations.

performance[i], cost_cny[i], and latency_ms[i] correspond to model_ids[i]. USD cost proxies use the fixed benchmark conversion 1 USD = 7.2 CNY. Candidate identities are listed in metadata/.

Request features

General Text and Chinese use BGE-M3 features. Multimodal uses CLIP ViT-B/32 (openai/clip-vit-base-patch32, checkpoint 3d74acf9a28c67741b2f4f2ea7635f0aaf6f0268). Full text is split into chunks of up to 75 payload tokens. Normalized text vectors are pooled by payload-token count; normalized vectors for all request images are pooled equally. Each modality pool is normalized, then the two 512-vectors are concatenated and divided by the square root of two. The result is a unit-norm 1,024-dimensional vector. All methods and fitting seeds share these features. Reference answers are not inputs to the encoder.

Router fitting and prediction replay use the bundled features without downloading encoder weights or images. Original images and reference annotations are supplied in the multimodal media archives, organized as train.zip, validation.zip, standard_test.zip, and robust_test.zip.

Benchmark composition

General Text covers code, general reasoning, knowledge QA, mathematical reasoning, reading understanding, and long context. Its robust requests apply semantic rewrites, language variation, and long-context noise to source requests.

Multimodal training draws from 18 MMR-Bench-V2 sub-benchmarks. Standard evaluation covers VQAv2, AI2D, ScienceQA and ScienceQA-IMG, DocVQA, and OCR-VQA. Robust evaluation includes text rewriting, context noise, image variation, and combined perturbations. The standard and robust populations are independently constructed; their comparison is at the distribution level.

Chinese covers AlignBench, C-Eval, CIF-Bench, CLUE tasks, CMB-Clin, Chinese SimpleQA, FinEval, FollowBench-zh, GAOKAO-Bench, LawBench, and Chinese LongBench tasks. Its robust requests use semantic rewriting, English translation, and prefix noise.

Robust test populations

Scenario	Transformation	Test requests
General Text	Semantic rewrite	963
General Text	Language variation	964
General Text	Long-context noise	967
Multimodal	Text rewrite	475
Multimodal	Context noise	475
Multimodal	Image variation	475
Multimodal	Combined perturbation	475
Chinese	Semantic rewrite	667
Chinese	English translation	666
Chinese	Prefix noise	667

Transformations preserve the answer contract and the source scorer. Image variation includes JPEG degradation, cutout, Gaussian noise, Gaussian blur, and contrast reduction. Text-scenario variants retain their source identifiers.

Compared methods

The report compares 13 implementations using ORBIT interfaces, with the same candidate pool, request features, and test requests within each scenario. Mio selector is fitted separately per scenario and configured using validation requests.

Implementation	Method reference
Mio selector	Technical report Section 4.2
Avengers	The Avengers
Avengers-Pro	Performance-Efficiency Optimized Routing
Causal RM-Softmax	Causal LLM Routing
Eagle	Eagle
EmbedLLM-RouteJudge	EmbedLLM
GraphRouter	GraphRouter
HybridLLM	Hybrid LLM
kNN	TensorOpera / ORBIT
MLP	TensorOpera / ORBIT
MIRT	IRT-Router
RouteLLM-MF	RouteLLM
SVM Router	TensorOpera / ORBIT

The comparison evaluates the offline request-level selector with $T=1T=1$ under monetary budgets, as in Section 6.3 of the technical report. Implementation settings are provided with the prediction files and in the code.

Metrics

Section 6.3 uses nAUC and peak quality. Threshold costs and savings are supplementary measures.

Evaluate 100 evenly spaced nominal budgets between the cheapest and most expensive fixed-model mean costs on each reporting population. These endpoints use held-out costs and are shared across methods. Predicted cost determines feasibility; predicted utility selects among candidates. A request with no feasible candidate uses the model with the lowest predicted cost. Observed outcomes determine mean utility $Q_{\pi }(C)Q\_\backslash pi(C)$ and realized mean cost ${\bar{c}}_{\pi }(C)\backslash bar\; c\_\backslash pi(C)$.

For the common nominal-budget grid $\mathcal{G}\backslash mathcal\; G$, peak quality and threshold cost are:

Cost@90, Cost@95, and Cost@100 use $\tau =0.90,0.95,1.00\backslash tau=0.90,0.95,1.00$, respectively. An unattained target is N/R. The strongest fixed model maximizes mean quality on the same reporting population and candidate pool; its mean cost is the savings reference. Savings are:

$${\mathrm{Saving}}_{\pi }\mathrm{@}\tau =1-\frac{{\bar{c}}_{\pi }\mathrm{@}\tau }{{\bar{c}}_{\mathrm{s}\mathrm{t}\mathrm{r}\mathrm{o}\mathrm{n}\mathrm{g}\mathrm{e}\mathrm{s}\mathrm{t}}}\mathrm{.}\backslash operatorname\{Saving\}\_\{\backslash pi\}\backslash mathbin\{@\}\backslash tau=1-\backslash frac\{\backslash bar\{c\}\_\{\backslash pi\}\backslash mathbin\{@\}\backslash tau\}\{\backslash bar\; c\_\{\backslash mathrm\{strongest\}\}\}.$$

nAUC integrates the cumulative-maximum quality envelope over increasing nominal budget and divides by the full budget interval:

Curves contain evaluated selector points only. The quality-cost figure uses realized cost on a logarithmic axis; its geometric area is not nAUC. Macro nAUC equally averages the six scenario/standard-robust scores. Robust split metrics use all robust requests pooled within a scenario. Each robustness type is evaluated separately with its own cost endpoints; averaging type-level nAUC does not generally recover pooled Robust nAUC. Latency is auxiliary and is not used in the selection objective or as a hard constraint in this evaluation.

Multimodal uses fitting seeds 20260917, 20260918, and 20260919. Curves are averaged pointwise on the common budget grid before computing their metrics. Reported uncertainties are exact run-bootstrap standard errors over all 27 ordered resamples. General Text and Chinese each use seed 20260917.

Dataset costs use CNY. The supplementary cost tables below use USD for General Text and Multimodal, and CNY for Chinese; divide the first two scenarios' cost_cny values by 7.2 for those display units. Multimodal costs are fixed proxies.

Results

Mio selector achieves the highest macro nAUC, 0.6360, followed by MLP at 0.6183. The nAUC values and robustness-type scores and ranks below match Tables 6 and 7 of the technical report. The full split-level tables also provide peak quality, threshold costs, and fitting uncertainty.

Supplementary figure. Green: Mio selector frontier and peak. Hollow circles: baseline peaks. Blue diamond: best baseline by nAUC. Multimodal points use the mean curve of three fits. Costs are USD for General Text and Multimodal, and CNY for Chinese.

Costs below are per request. N/R denotes an unattained aggregate-curve target; (r/3) records the number of individual Multimodal runs reaching that target.

Macro nAUC

Rank	Implementation	Macro nAUC
1	Mio selector	0.6360
2	MLP	0.6183
3	EmbedLLM-RouteJudge	0.6044
4	Avengers-Pro	0.5753
5	Eagle	0.5731
6	Avengers	0.5603
7	RouteLLM-MF	0.5573
8	kNN	0.5443
9	HybridLLM	0.5344
10	Causal RM-Softmax	0.5245
11	SVM Router	0.5146
12	MIRT	0.4733
13	GraphRouter	0.4669

General Text: Standard

Costs in USD per request.

nAUC rank	Implementation	Peak quality	nAUC	Cost@90	Cost@95	Cost@100
1	Eagle	0.6104	0.6045	0.0253	0.0416	0.1131
2	Avengers-Pro	0.6071	0.6022	0.0442	0.0625	0.1067
3	RouteLLM-MF	0.6016	0.5939	0.0476	0.0874	0.2490
4	Mio selector	0.5979	0.5924	0.0321	0.0609	0.2285
5	MLP	0.5949	0.5884	0.0238	0.0409	0.2475
6	EmbedLLM-RouteJudge	0.5949	0.5863	0.1674	0.2475	0.2475
7	kNN	0.5905	0.5853	0.0360	0.0691	N/R
8	SVM Router	0.5854	0.5789	0.0725	0.1643	N/R
9	Avengers	0.5817	0.5770	0.0653	0.1267	N/R
10	HybridLLM	0.5778	0.5684	0.0722	0.2039	N/R
11	Causal RM-Softmax	0.4969	0.4969	N/R	N/R	N/R
12	MIRT	0.2853	0.2771	N/R	N/R	N/R
13	GraphRouter	0.2501	0.2468	N/R	N/R	N/R

General Text: Robust

Costs in USD per request.

nAUC rank	Implementation	Peak quality	nAUC	Cost@90	Cost@95	Cost@100
1	Mio selector	0.5596	0.5328	0.2258	0.2389	0.2492
2	EmbedLLM-RouteJudge	0.5596	0.5304	0.2492	0.2492	0.2492
3	MLP	0.5596	0.5231	0.2492	0.2492	0.2492
4	Eagle	0.3483	0.3359	N/R	N/R	N/R
5	RouteLLM-MF	0.2974	0.2891	N/R	N/R	N/R
6	Avengers-Pro	0.2894	0.2865	N/R	N/R	N/R
7	Avengers	0.2854	0.2832	N/R	N/R	N/R
8	kNN	0.2773	0.2671	N/R	N/R	N/R
9	HybridLLM	0.2764	0.2616	N/R	N/R	N/R
10	SVM Router	0.2706	0.2611	N/R	N/R	N/R
11	Causal RM-Softmax	0.1774	0.1774	N/R	N/R	N/R
12	MIRT	0.1185	0.1133	N/R	N/R	N/R
13	GraphRouter	0.1154	0.1110	N/R	N/R	N/R

Multimodal: Standard

Costs in USD per request.

nAUC rank	Implementation	Peak quality	nAUC	Cost@90	Cost@95	Cost@100
1	Mio selector	0.7084 ± 0.0008	0.7077 ± 0.0007	2.91e-6 (3/3)	2.91e-6 (3/3)	N/R
2	MIRT	0.6359 ± 0.0155	0.6359 ± 0.0155	N/R (1/3 runs)	N/R	N/R
3	MLP	0.6663 ± 0.0027	0.6299 ± 0.0032	0.0035 (3/3)	N/R	N/R
4	Avengers-Pro	0.6464 ± 0.0010	0.6239 ± 0.0019	N/R (2/3 runs)	N/R	N/R
5	Causal RM-Softmax	0.6217 ± 0.0393	0.6217 ± 0.0393	N/R (1/3 runs)	N/R (1/3 runs)	N/R
6	Eagle	0.6507 ± 0.0000	0.6095 ± 0.0000	0.0039 (3/3)	N/R	N/R
7	Avengers	0.6358 ± 0.0075	0.6064 ± 0.0096	N/R	N/R	N/R
8	RouteLLM-MF	0.6236 ± 0.0000	0.5914 ± 0.0000	N/R	N/R	N/R
9	kNN	0.6000 ± 0.0000	0.5846 ± 0.0000	N/R	N/R	N/R
10	GraphRouter	0.5558 ± 0.0006	0.5558 ± 0.0005	N/R	N/R	N/R
11	EmbedLLM-RouteJudge	0.6193 ± 0.0291	0.5510 ± 0.0524	N/R (3/3 runs)	N/R (2/3 runs)	N/R
12	HybridLLM	0.5555 ± 0.0000	0.5385 ± 0.0000	N/R	N/R	N/R
13	SVM Router	0.4869 ± 0.0222	0.4869 ± 0.0222	N/R	N/R	N/R

Multimodal: Robust

Costs in USD per request.

nAUC rank	Implementation	Peak quality	nAUC	Cost@90	Cost@95	Cost@100
1	Mio selector	0.6266 ± 0.0008	0.6128 ± 0.0005	3.41e-6 (3/3)	3.41e-6 (3/3)	N/R
2	MLP	0.5932 ± 0.0006	0.5918 ± 0.0007	0.0001 (3/3)	N/R	N/R
3	EmbedLLM-RouteJudge	0.5869 ± 0.0145	0.5869 ± 0.0138	0.0002 (3/3)	N/R (3/3 runs)	N/R
4	Avengers-Pro	0.5979 ± 0.0025	0.5680 ± 0.0028	0.0031 (3/3)	N/R	N/R
5	MIRT	0.5641 ± 0.0128	0.5638 ± 0.0128	N/R (1/3 runs)	N/R	N/R
6	Avengers	0.5913 ± 0.0019	0.5553 ± 0.0204	0.0025 (3/3)	N/R	N/R
7	Causal RM-Softmax	0.5477 ± 0.0286	0.5477 ± 0.0287	N/R (1/3 runs)	N/R (1/3 runs)	N/R
8	RouteLLM-MF	0.5572 ± 0.0000	0.5439 ± 0.0000	N/R	N/R	N/R
9	Eagle	0.5609 ± 0.0000	0.5264 ± 0.0000	N/R	N/R	N/R
10	HybridLLM	0.5234 ± 0.0000	0.5106 ± 0.0000	N/R	N/R	N/R
11	GraphRouter	0.5038 ± 0.0003	0.5038 ± 0.0002	N/R	N/R	N/R
12	kNN	0.4995 ± 0.0000	0.4995 ± 0.0000	N/R	N/R	N/R
13	SVM Router	0.4714 ± 0.0204	0.4714 ± 0.0204	N/R	N/R	N/R

Chinese: Standard

Costs in CNY per request.

nAUC rank	Implementation	Peak quality	nAUC	Cost@90	Cost@95	Cost@100
1	GraphRouter	0.7675	0.7675	0.0666	0.0666	N/R
2	Avengers-Pro	0.7688	0.7660	0.0123	0.0123	0.0606
3	MLP	0.7687	0.7656	0.0132	0.0132	N/R
4	Mio selector	0.7711	0.7653	0.0127	0.0127	0.0348
5	Eagle	0.7719	0.7642	0.0122	0.0122	0.0516
6	EmbedLLM-RouteJudge	0.7675	0.7629	0.0117	0.0148	N/R
7	kNN	0.7663	0.7598	0.0122	0.0122	N/R
8	Avengers	0.7604	0.7597	0.0132	0.0132	N/R
9	HybridLLM	0.7623	0.7555	0.0137	0.0137	N/R
10	RouteLLM-MF	0.7585	0.7532	0.0140	0.0140	N/R
11	MIRT	0.7688	0.7437	0.0206	0.0266	0.0666
12	Causal RM-Softmax	0.7406	0.7404	0.0143	0.0143	N/R
13	SVM Router	0.7434	0.7394	0.0137	0.0171	N/R

Chinese: Robust

Costs in CNY per request.

nAUC rank	Implementation	Peak quality	nAUC	Cost@90	Cost@95	Cost@100
1	GraphRouter	0.6163	0.6163	0.0791	0.0791	0.0791
2	MLP	0.6178	0.6108	0.0062	0.0098	0.0642
3	EmbedLLM-RouteJudge	0.6167	0.6087	0.0182	0.0182	0.0775
4	Avengers-Pro	0.6161	0.6054	0.0173	0.0186	N/R
5	Mio selector	0.6137	0.6051	0.0061	0.0089	N/R
6	Eagle	0.6142	0.5982	0.0122	0.0165	N/R
7	Avengers	0.5932	0.5801	0.0257	0.0308	N/R
8	RouteLLM-MF	0.5869	0.5724	0.0123	0.0604	N/R
9	HybridLLM	0.5897	0.5721	0.0189	0.0561	N/R
10	kNN	0.5919	0.5694	0.0224	0.0521	N/R
11	Causal RM-Softmax	0.5632	0.5628	0.0078	N/R	N/R
12	SVM Router	0.5607	0.5499	0.0479	N/R	N/R
13	MIRT	0.6162	0.5060	0.0728	0.0728	0.0792

Mio selector by robustness type

Ranks are among 13 distinct implementations; Multimodal metrics use the pointwise mean curve over three runs and ± is exact run-bootstrap standard error.

Scenario	Robustness type	Peak quality	Peak rank	nAUC	nAUC rank
General Text	Semantic rewrite	0.5124	1/13	0.4829	1/13
General Text	Language variation	0.4913	1/13	0.4487	2/13
General Text	Long-context noise	0.6747	1/13	0.6529	1/13
Multimodal	Text rewrite	0.6455 ± 0.0010	1/13	0.6438 ± 0.0003	1/13
Multimodal	Context noise	0.6219 ± 0.0028	1/13	0.6021 ± 0.0018	1/13
Multimodal	Image variation	0.6289 ± 0.0009	1/13	0.6247 ± 0.0010	1/13
Multimodal	Combined perturbation	0.6099 ± 0.0006	2/13	0.5827 ± 0.0013	3/13
Chinese	Semantic rewrite	0.6156	5/13	0.6086	4/13
Chinese	English translation	0.5893	7/13	0.5802	5/13
Chinese	Prefix noise	0.6375	7/13	0.6287	5/13

Reproduce the report

From the GitHub repository:

python -m pip install -e '.[report]'
python -m examples.reproduce_report --mode replay --output runs/report-replay
python -m examples.reproduce_report --mode fit --device cpu --output runs/report-fit

Replay evaluates the 65 supplied prediction arrays across 13 methods and checks the aggregate against report/aggregate-results.json. Fit trains the 12 baseline implementations; Mio is supplied through its prediction arrays. Predictions are aligned by request and model identifiers. All requests in the official test split contribute to evaluation.

report/reproduction-manifest.json records prediction checksums and fitting seeds. manifest.json records data and media checksums. The original benchmark questions and images retain their upstream attribution and terms in LICENSES.md.