RobustRouteBench

RobustRouteBench evaluates how model routers balance quality and cost when requests change in wording, language, context, or visual content. It covers General Text, Multimodal, and Chinese, each with its own fixed candidate pool.

This benchmark accompanies Section 6.3 of the SinapisAI technical report Toward Collective Model Intelligence: Characterizing and Coordinating Models. Reported results cover offline, fixed-pool, single-request model selection ($T=1$) under monetary budgets.

Dataset overview

Scenario / config	Models	Train	Standard test (final)	Robust test (final)
General Text / text	63	3,000	4,188 (3,688)	3,000 (2,894)
Multimodal / multimodal	45	4,000	5,000 (4,600)	2,000 (1,900)
Chinese / chinese	10	7,590	4,400 (3,900)	2,100 (2,000)
Total	—	14,590	13,588 (12,188)	7,100 (6,794)

Parentheses show the final evaluation counts after excluding development source groups: 18,982 requests in total. Candidate pools are scenario-specific; the model counts do not form one combined pool.

Quick start

from datasets import load_dataset

text = load_dataset("SinapisAI/RobustRouteBench", "text")
multimodal = load_dataset("SinapisAI/RobustRouteBench", "multimodal")
chinese = load_dataset("SinapisAI/RobustRouteBench", "chinese")

Each config provides train, validation, and test splits. Rows contain a request and aligned per-model outcomes:

• model_ids indexes the performance, cost_cny, and latency_ms arrays. Unsuccessful source outcomes have performance zero.
• evaluation_type separates standard and robust test rows; variant_type identifies the perturbation.
• Costs in the dataset use CNY. Released BGE-M3 arrays provide optional text-only baseline features; other encoders are allowed.

For the reported Multimodal experiment, all compared methods share Qwen3-VL-Embedding-8B representations, jointly encoding each request's text and images (technical report). Image and combined perturbations therefore enter the selector through the multimodal request representation.

All multimodal training and evaluation media are included under media/multimodal/. Field definitions are in normalized-schema.json.

Evaluation protocol

Separate development and final evaluation

Validation and the published test split overlap by design. Each scenario's 500 validation rows contain 400 standard and 100 robust requests. Before final scoring, exclude every source group represented in validation, including linked base requests and variants; dropping only is_validation=True rows is insufficient.

Group General Text and Chinese by base_query_id, falling back to query_id; group Multimodal by source_sample_id. This leaves 6,582 / 6,500 / 5,900 final requests, respectively, with no development-group overlap. Training groups are also disjoint from development and final-test groups under these identifiers. Benchmark sources and task domains can still be shared, so this is base-query/source-sample separation rather than a document-level or cross-benchmark independence claim. Development used a public subset of the same benchmark release; the final evaluation is group-disjoint, not an independently administered hidden benchmark.

Example: construct the final evaluation splits

def source_group(row, scenario):
    if scenario == "multimodal":
        key = row["source_sample_id"]
        if key is None or key == "":
            raise ValueError("Missing multimodal source group")
        return key
    return row["base_query_id"] or row["query_id"]

final_tests = {}
for scenario, ds in [("text", text), ("multimodal", multimodal), ("chinese", chinese)]:
    dev_groups = {source_group(row, scenario) for row in ds["validation"]}
    final_tests[scenario] = ds["test"].filter(
        lambda row: source_group(row, scenario) not in dev_groups
    )

assert [len(final_tests[s]) for s in ("text", "multimodal", "chinese")] == [6582, 6500, 5900]

Train and tune on training/development data only. Methods share the same candidate pool and final requests within each scenario; held-out utilities, references, and realized costs are evaluation labels, unavailable to the router at decision time.

General Text and Chinese contain paired base requests and variants. Multimodal standard and robust sets are independently sampled, so their difference is a distribution-level comparison. Even in the paired scenarios, an overall standard–robust difference is not a paired effect: the standard set also contains requests outside the robust base set.

Metrics

Evaluate each method on 100 evenly spaced nominal budgets, from the cheapest to the most expensive fixed-model mean cost on the reporting population. The endpoints use held-out cost aggregates for evaluator-side budget construction and are shared across methods; they are not training/development statistics. Predicted cost determines feasibility; predicted quality or preference selects the model. Report quality and realized cost from held-out outcomes.

When no candidate meets the predicted cost budget, the offline sweep falls back to the model with the lowest predicted cost and retains its measured outcome. This fallback can exceed the nominal budget even in predicted cost. It is an evaluation convention; Section 4.2's deployment rule reports unmet constraints when its feasible set is empty.

Metric	Definition
Peak quality	Highest mean utility reached on the budget grid.
nAUC	Trapezoidal area under the cumulative-best quality curve in increasing nominal-budget order, divided by the full nominal-budget interval. Use actual router points only; no fixed-model references, added endpoints, or extrapolation.
Cost@90 / 95 / 100	Lowest realized mean cost among grid points reaching 90% / 95% / 100% of the strongest fixed model's quality on the same split. N/R means the target was not reached.

Macro nAUC is the unweighted mean over the six scenario/split results. Latency is auxiliary: it does not enter the selection objective, impose a hard constraint, or determine the ranking. A nominal budget does not guarantee a realized-cost bound. The toolkit's realized-cost Pareto AUC is a different metric. To reproduce the report, align repeated runs on the same nominal-budget grid, average their utility and realized-cost curves pointwise, and then compute the cumulative maximum, peak quality, nAUC, and target costs.

Pooled Robust is evaluated directly on the combined robust request population. Each robustness type is evaluated separately on its own cost interval; averaging type-level nAUC values does not generally recover pooled Robust nAUC. Both budget grids and the nonlinear cumulative-maximum envelope can differ. This also applies to the four equally sized Multimodal robustness types.

Units and runs. Dataset costs are CNY; the report tables retain USD for General Text and Multimodal and CNY for Chinese. Convert the first two using the fixed benchmark constant 1 USD = 7.2 CNY, including budgets. Multimodal cost proxies are fixed, not live prices. Multimodal metrics are computed from the pointwise mean of three run curves; ± denotes exact run-bootstrap standard error, obtained from all $3^3$ ordered resamples of the run curves with the complete curve statistic recomputed. General Text and Chinese use one run. Threshold costs come from the aggregate curve; (r/3) records how many individual runs reach the target. N/R can coexist with a nonzero reach count when the aggregate curve does not reach the target. These errors describe fitting variation, not request-sampling confidence intervals.

Results

Mio selector achieves macro nAUC 0.6375, followed by MLP at 0.6151, across 13 distinct implementations. The best baseline below is selected independently for each split.

Scenario	Split	Mio selector nAUC	Best baseline	Baseline nAUC
General Text	Standard	0.5908	Eagle	0.6045
General Text	Robust	0.5299	MLP	0.5052
Multimodal	Standard	0.7109 ± 0.0083	Causal RM-Softmax	0.6703 ± 0.0402
Multimodal	Robust	0.6228 ± 0.0064	EmbedLLM-RouteJudge	0.5958 ± 0.0125
Chinese	Standard	0.7653	GraphRouter	0.7688
Chinese	Robust	0.6051	GraphRouter	0.6167

Mio selector leads both Multimodal splits and General Text Robust by nAUC; Eagle leads General Text Standard and GraphRouter leads both Chinese splits. For Mio selector, General Text Robust Cost@90/95/100 is 0.2492 USD and Chinese Robust does not reach Cost@100. The aggregate Multimodal Mio curve does not reach Cost@100 on either split; each records a 1/3 individual-run diagnostic rather than an aggregate target cost.

Green: Mio selector frontier and peak. Hollow circles: other baseline peaks. Blue diamond: best baseline by nAUC. Multimodal curves are pointwise means of three runs; both the displayed peaks and nAUC are computed from those mean curves. The log x-axis shows realized cost (USD for General Text/Multimodal; CNY for Chinese); the plotted area is not nominal-budget nAUC. This supplementary visualization uses the same frozen results as the report tables.

Full results: all 13 implementations and robustness-type breakdown

Costs are per request in the stated currency. Rankings use nAUC. Multimodal values are aggregate-curve statistics with exact run-bootstrap standard errors; threshold-cost reach counts are individual-run diagnostics. Values below reproduce the report's frozen results.

Macro nAUC

Rank	Implementation	Macro nAUC
1	Mio selector	0.6375
2	MLP	0.6151
3	EmbedLLM-RouteJudge	0.6053
4	Eagle	0.5552
5	Avengers-Pro	0.5503
6	Avengers	0.5442
7	RouteLLM-MF	0.5333
8	HybridLLM	0.5245
9	kNN	0.5190
10	Causal RM-Softmax	0.5088
11	SVM Router	0.5031
12	GraphRouter	0.4629
13	MIRT	0.4315

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	0.5949	0.5908	0.0369	0.0553	0.2471
5	MLP	0.5949	0.5891	0.0296	0.0391	0.2475
6	EmbedLLM-RouteJudge	0.5957	0.5863	0.2476	0.2476	0.2476
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	0.2687	0.2616	N/R	N/R	N/R
13	GraphRouter	0.2393	0.2379	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	0.5299	0.2492	0.2492	0.2492
2	MLP	0.5596	0.5052	0.2492	0.2492	0.2492
3	EmbedLLM-RouteJudge	0.5399	0.4990	0.2488	0.2488	N/R
4	Eagle	0.2592	0.2428	N/R	N/R	N/R
5	Avengers	0.1653	0.1625	N/R	N/R	N/R
6	RouteLLM-MF	0.1683	0.1593	N/R	N/R	N/R
7	Avengers-Pro	0.1599	0.1562	N/R	N/R	N/R
8	kNN	0.1373	0.1266	N/R	N/R	N/R
9	HybridLLM	0.1255	0.1092	N/R	N/R	N/R
10	SVM Router	0.1141	0.1030	N/R	N/R	N/R
11	GraphRouter	0.0233	0.0233	N/R	N/R	N/R
12	MIRT	0.0218	0.0210	N/R	N/R	N/R
13	Causal RM-Softmax	0.0000	0.0000	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	0.7116 ± 0.0080	0.7109 ± 0.0083	3.00e-6 (3/3)	3.00e-6 (3/3)	N/R (1/3 runs)
2	Causal RM-Softmax	0.6703 ± 0.0402	0.6703 ± 0.0402	2.94e-6 (2/3)	N/R (2/3 runs)	N/R (2/3 runs)
3	MLP	0.6609 ± 0.0024	0.6293 ± 0.0042	0.0041 (3/3)	N/R	N/R
4	Avengers	0.6335 ± 0.0136	0.6230 ± 0.0174	N/R (2/3 runs)	N/R (1/3 runs)	N/R
5	Avengers-Pro	0.6291 ± 0.0067	0.6059 ± 0.0014	N/R	N/R	N/R
6	GraphRouter	0.6021 ± 0.0234	0.6020 ± 0.0242	N/R (1/3 runs)	N/R	N/R
7	RouteLLM-MF	0.6298 ± 0.0000	0.5844 ± 0.0000	N/R	N/R	N/R
8	HybridLLM	0.5843 ± 0.0000	0.5842 ± 0.0000	N/R	N/R	N/R
9	EmbedLLM-RouteJudge	0.6772 ± 0.0107	0.5782 ± 0.0508	0.0032 (3/3)	N/R (2/3 runs)	N/R (1/3 runs)
10	Eagle	0.5881 ± 0.0000	0.5778 ± 0.0000	N/R	N/R	N/R
11	MIRT	0.5770 ± 0.0024	0.5765 ± 0.0021	N/R	N/R	N/R
12	kNN	0.5502 ± 0.0000	0.5502 ± 0.0000	N/R	N/R	N/R
13	SVM Router	0.5049 ± 0.0048	0.5049 ± 0.0048	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	0.6261 ± 0.0083	0.6228 ± 0.0064	3.44e-6 (3/3)	3.44e-6 (3/3)	N/R (1/3 runs)
2	EmbedLLM-RouteJudge	0.6013 ± 0.0102	0.5958 ± 0.0125	0.0002 (3/3)	N/R (3/3 runs)	N/R
3	MLP	0.5965 ± 0.0079	0.5929 ± 0.0089	0.0002 (3/3)	N/R (2/3 runs)	N/R
4	Causal RM-Softmax	0.5824 ± 0.0291	0.5824 ± 0.0291	3.41e-6 (2/3)	N/R (2/3 runs)	N/R
5	Avengers-Pro	0.5802 ± 0.0041	0.5659 ± 0.0051	0.0031 (3/3)	N/R	N/R
6	Avengers	0.5891 ± 0.0082	0.5629 ± 0.0096	0.0027 (2/3)	N/R	N/R
7	HybridLLM	0.5579 ± 0.0000	0.5579 ± 0.0000	N/R	N/R	N/R
8	Eagle	0.5663 ± 0.0000	0.5434 ± 0.0000	N/R	N/R	N/R
9	SVM Router	0.5430 ± 0.0018	0.5424 ± 0.0024	N/R	N/R	N/R
10	RouteLLM-MF	0.5506 ± 0.0000	0.5366 ± 0.0000	N/R	N/R	N/R
11	GraphRouter	0.5290 ± 0.0256	0.5289 ± 0.0260	N/R (1/3 runs)	N/R	N/R
12	kNN	0.5354 ± 0.0000	0.5229 ± 0.0000	N/R	N/R	N/R
13	MIRT	0.4881 ± 0.0015	0.4880 ± 0.0012	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	0.7688	0.7688	0.0666	0.0666	0.0666
2	Avengers-Pro	0.7688	0.7660	0.0123	0.0123	0.0606
3	Mio selector	0.7711	0.7653	0.0127	0.0127	0.0348
4	MLP	0.7678	0.7642	0.0132	0.0132	N/R
5	Eagle	0.7719	0.7642	0.0122	0.0122	0.0516
6	EmbedLLM-RouteJudge	0.7679	0.7632	0.0117	0.0147	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	0.7418	0.0201	0.0296	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	0.6167	0.6167	0.0792	0.0792	0.0792
2	MLP	0.6176	0.6098	0.0062	0.0100	0.0639
3	EmbedLLM-RouteJudge	0.6167	0.6094	0.0182	0.0182	0.0778
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	0.4999	0.0725	0.0725	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	0.4636	1/13
General Text	Language variation	0.4913	1/13	0.4360	1/13
General Text	Long-context noise	0.6747	1/13	0.6556	1/13
Multimodal	Text rewrite	0.6559 ± 0.0116	1/13	0.6550 ± 0.0115	1/13
Multimodal	Context noise	0.6534 ± 0.0037	1/13	0.6380 ± 0.0030	1/13
Multimodal	Image variation	0.6375 ± 0.0068	1/13	0.6367 ± 0.0068	1/13
Multimodal	Combined perturbation	0.6265 ± 0.0029	1/13	0.6059 ± 0.0070	1/13
Chinese	Semantic rewrite	0.6156	4/13	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

Benchmark details

Source datasets and robustness transformations

General Text

Domain	Train	Published standard	Published robust
Code	1,386	859	268
General Reasoning	435	649	474
Knowledge QA	388	616	506
Long Context	0	1,000	1,000
Mathematical Reasoning	438	574	410
Reading Understanding	353	490	342

The 3,000 robust instances derive from 2,907 base queries retained in standard test. All 1,000 Long Context requests are robustness bases; none are in training. Long-context noise adds one to three removable background notes.

Multimodal

MMR-4000 training examples come from 18 MMR-Bench-V2 sub-benchmarks. VisionQA-5000 standard evaluation contains:

Capability group	Source benchmark	Published standard instances
General VQA	VQAv2	1,500
Science / Diagram	AI2D	676
Science / Diagram	ScienceQA	667
Science / Diagram	ScienceQA-IMG	657
Document / OCR	DocVQA	183
Document / OCR	OCR-VQA	1,317
Total		5,000

Requests may contain multiple images, text segments, and answer options. The 2,000 robust items are sampled independently of VisionQA-5000; both sets use the same 45 candidates.

Chinese

Coverage	Source benchmark
Alignment	AlignBench
Knowledge and examinations	C-Eval
Instruction following	CIF-Bench
Language, NLI, and reading	CLUE benchmark family
Clinical medicine	CMB-Clin from CMB
Factual knowledge	Chinese SimpleQA
Financial knowledge	FinEval
Constraint following	FollowBench-zh
Standardized examinations	GAOKAO-Bench
Legal knowledge	LawBench
Long-context tasks	Chinese subsets of LongBench

Each source retains its task and evaluator semantics; quality is not a single shared accuracy measure across tasks.

Robustness groups

Scenario	Robustness type	Published instances	Final-test instances	Relation to standard population
General Text	Semantic rewrite	1,000	963	Paired base retained
General Text	Language variation	1,000	964	Paired base retained
General Text	Long-context noise	1,000	967	Paired base retained
Multimodal	Text rewrite	500	475	Independently sampled
Multimodal	Context noise	500	475	Independently sampled
Multimodal	Image variation	500	475	Independently sampled
Multimodal	Combined perturbation	500	475	Independently sampled
Chinese	Semantic rewrite	700	667	Paired, disjoint bases
Chinese	English translation	700	666	Paired, disjoint bases
Chinese	Prefix noise	700	667	Paired, disjoint bases

Transformations must preserve meaning, required evidence, the answer contract, and scorer validity. Image variation uses 100 instances each of JPEG degradation, controlled cutout, Gaussian noise, Gaussian blur, and contrast reduction. Combined multimodal perturbations change text, context, and image together. Chinese robustness types use disjoint base requests retained in standard test.

Compared methods, implementation aliases, and run settings

The report compares Mio selector and 12 baselines through ORBIT's common interfaces.

Reported implementation	Method source
Avengers	The Avengers
Avengers-Pro	Performance-Efficiency Optimized Routing
Causal RM-Softmax	Causal LLM Routing, RM-Softmax
Eagle	Eagle
EmbedLLM-RouteJudge	EmbedLLM
Mio selector	Request-oriented selection, technical report Section 4.2; scenario-specific offline training
GraphRouter	GraphRouter
HybridLLM	Hybrid LLM
kNN	TensorOpera / ORBIT baseline
MLP	TensorOpera / ORBIT baseline
MIRT	IRT-Router, MIRT variant
RouteLLM-MF	RouteLLM, matrix-factorization variant
SVM Router	TensorOpera / ORBIT baseline

Mio selector is trained separately on each scenario's training split, with hyperparameters and operating points selected on development data. Model-side evidence comes from benchmark training observations; the Section 3 Model Eval profile-construction pipeline is not jointly evaluated here. These are selector-level results, not a hosted Mio 1.0 Lite API or end-to-end platform evaluation. They do not isolate profile encoding, pairwise supervision, or near-tie selection, and do not evaluate candidate-pool organization, multi-step handoff, or Section 4.2's rank-fusion rule.

The following labels produced identical fitted-adapter predictions and selection arrays in this experiment and are counted once:

Ranked implementation	Additional implementation aliases in this experiment
Causal RM-Softmax	RM-Classification, RM-Interval (source)
Eagle	RouteLLM-SW Ranking (RouteLLM)
EmbedLLM-RouteJudge	Model-SAT (Capability Instruction Tuning), RouterDC
MLP	OmniRouter, RouteLLM-BERT (RouteLLM)
MIRT	NIRT (IRT-Router)

These are implementation aliases within this experiment, not a claim that the original research methods are equivalent or independently reproduced.

Multimodal uses seeds 20260917, 20260918, and 20260919; General Text and Chinese use 20260917. The Multimodal runs share 4,000 training requests, 45 candidates, Qwen3-VL-Embedding-8B text-and-image representations, identical final evaluation requests and source-group exclusions, and the same 100-point budget grid for each reporting population. Only the fitting seed changes. The reported uncertainty uses exact run-bootstrap standard errors, not paired request-bootstrap intervals; no significance tests are reported.

Release scope and license

The release includes scored per-model outcomes, baseline embeddings, and multimodal media. It does not include raw model responses, source diagnostic payloads, Mio's internal training implementation, or per-run prediction arrays. Dynamic model-pool changes and live-service recovery are outside the reported evaluation.

The numerical source is the technical report's frozen RobustRouteBench results, archive format version 3, generated 2026-10-12. It records the pointwise-curve aggregation protocol, exclusion counts, seeds, budget points, metrics, and the alias audit.

The original data revision is fcf2e3607c1c3ad303c202e2175a837ae4fbfede. Pin it through revision= when reproducing the data. This documentation update preserves the published data, split layout, schema, model metadata, and original manifest. The manifest's README checksum refers to the original card at that data revision.

See manifest.json for the data inventory and LICENSES.md for upstream attribution and terms. This compilation uses license: other; upstream questions and images retain their respective terms.