RobustRouteBench

RobustRouteBench is a unified public benchmark for general-text, multimodal, and Chinese-language model routing.

from datasets import load_dataset

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

Each scenario exposes separate train, validation, and test splits. The 500-row validation split is a convenience overlay containing 400 standard and 100 robust queries. Those rows intentionally remain in the complete test split, so validation and test overlap by design.

The test split contains both standard and robustness evaluation rows. Use evaluation_type (standard or robust) and variant_type to report slices.

Each row is query-centric. performance, cost_cny, and latency_ms are aligned to model_ids. All costs use CNY; historical USD proxies use the frozen benchmark conversion 1 USD = 7.2 CNY.

Source outcomes that were not successful have performance zero. No source diagnostic payloads or raw model responses are published. General Text and Chinese use BGE-M3 features. Multimodal uses frozen CLIP ViT-B/32 features (ICML paper), checkpoint 3d74acf9a28c67741b2f4f2ea7635f0aaf6f0268. Full prompt tokens are split into 75-token payload chunks plus BOS/EOS. Normalize each 512-dimensional text chunk vector and pool with payload-token-count weights. Encode every referenced RGB image using official 224-pixel resizing/center-cropping, normalize each vector, and pool images equally. Normalize the two pooled modality vectors, concatenate, and divide by √2. This yields 1,024-dimensional late fusion, not Qwen joint encoding. All methods/seeds use the same features. Missing images are errors; labels and reference answers never enter the encoder.

The corrected release restores exactly 20,167 successful General Text scores whose model, request, and answer match the historical nonzero record, including 522 cells in the validation overlay. Known failed requests retain zero scores. Other legacy zeros without a verified scoring source remain unchanged. This is a provenance-confirmed repair, not independent regrading of every legacy outcome.

The repository is self-contained. All MMR-4000 training images and all standard and robustness evaluation media are stored under media/multimodal/; no other dataset repository is required to train or evaluate a router. See LICENSES.md for upstream attribution and terms. This compilation uses license: other; no blanket license is asserted over all upstream questions or images.

Evaluation guide

Companion to SinapisAI: A Model Intelligence System for Collective AI, Section 6.3 (October 2026).

This card documents the RobustRouteBench experiment in Section 6.3 of the SinapisAI technical report, including the protocol and complete results previously collected in its RobustRouteBench appendix. The scored benchmark evaluates offline request-level model selection under a fixed candidate pool. It covers semantic rewrites, language changes, context noise, and multimodal perturbations. Model-ecosystem events and service recovery are complementary extension protocols, with no numerical event-replay or live-service results in the reported snapshot.

Mio selector means the request-level selection component trained separately on each scenario's training split, with hyperparameters and operating points selected using development data. These results do not evaluate the hosted Mio 1.0 Lite API or the end-to-end SinapisAI platform. Mio's internal training and inference implementation is not included in this dataset release.

The reported experiment covers the fixed-pool, single-request setting ($T=1$) under monetary budgets. Model-side evidence is learned from benchmark training observations; the Section 3 capability-profile construction pipeline is not jointly evaluated here. The results do not isolate profile encoding, pairwise supervision, or near-tie selection, and do not evaluate pool organization, multi-step handoff, or the Section 4.2 rank-fusion rule.

• Evaluation populations and development exclusion
• Benchmark composition and transformations
• Compared implementations
• Metrics and reporting conventions
• Complete reported results
• Attribution and uncertainty protocol
• Model-ecosystem and service extensions
• Reproducibility and release boundaries

Evaluation populations and development exclusion

The published test split is the complete evaluation inventory, not the report's final holdout. Its overlap with the 500-row validation overlay is intentional. To reproduce the report's final evaluation, remove every source group represented in validation, including linked base requests and variants, before calculating any test metric. Removing only rows with is_validation=True is insufficient for the paired text scenarios.

Scenario / config	Fixed candidate models	Train	Published standard	Final standard	Published robust	Final robust	Unique scenario instances
General Text / text	63	3,000	4,188	3,688	3,000	2,894	10,188
Multimodal / multimodal	45	4,000	5,000	4,600	2,000	1,900	11,000
Chinese / chinese	10	7,590	4,400	3,900	2,100	2,000	14,090
Total	Scenario-specific	14,590	13,588	12,188	7,100	6,794	35,278

Validation contains 400 standard and 100 robust requests per scenario and is already counted in the published evaluation inventory. Group exclusion removes 606 General Text, 500 Multimodal, and 600 Chinese rows, leaving 18,982 final-test requests (6,582 / 6,500 / 5,900). The resulting development-to-test source-group overlap is zero. Document- and repository-level identities are not consistently available, so this is a base-query/source-sample separation claim, not a stronger document-level independence claim. Development used a public subset of the same benchmark release; this is not an independently administered hidden benchmark.

Training groups are disjoint from development and final-test groups under the same base-query/source-sample identifiers. Benchmark sources and task domains can still be shared across these roles; the group separation does not establish cross-benchmark or cross-domain generalization.

The grouping rule in the frozen result archive is base_query_id when present, otherwise query_id, for General Text and Chinese; Multimodal uses source_sample_id. The following read-only filtering example operates on the loaded datasets above and does not change any published split:

from collections import Counter

expected = {
    "text": {"standard": 3688, "robust": 2894},
    "multimodal": {"standard": 4600, "robust": 1900},
    "chinese": {"standard": 3900, "robust": 2000},
}

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")
    else:
        key = row["base_query_id"] or row["query_id"]
    return (scenario, key)

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

Each scenario has its own fixed action space, shared by its training, standard, and robust populations. The pools may overlap in model family or version; 63, 45, and 10 are not one global pool. Candidate identities, train/development roles, final requests, generation settings, and outcome matrices are held fixed across methods within a scenario. Test utilities, references, and realized costs are evaluation labels and must remain unavailable to the selector at decision time.

Benchmark composition and 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 unique base queries retained in standard test. All 1,000 Long Context queries serve as robustness bases, and none appear in training. The language-variation set contains 999 instances with Chinese characters. Long-context noise inserts one to three background notes whose removal exactly recovers the original request.

Multimodal

MMR-4000 training examples are sampled from 18 MMR-Bench-V2 sub-benchmarks. VisionQA-5000 standard evaluation has the following composition:

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

A multimodal request is an ordered sequence of text and visual segments and may contain multiple images and answer options. The 2,000 robust items are sampled independently of VisionQA-5000, using the same 45-model action space. Standard versus robust differences are distribution-level comparisons, not per-request clean-versus-perturbed losses.

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; the utility is not assumed to be the same accuracy measure across all tasks. The published performance arrays are the scored outcomes, with unsuccessful source outcomes set to zero as specified above.

Robustness groups and pairing

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, necessary evidence, the answer contract, and scorer validity. Image variation assigns 100 instances each to JPEG degradation, controlled cutout, Gaussian noise, Gaussian blur, and contrast reduction. Combined multimodal perturbations change text, context, and image together. The three Chinese types use disjoint base requests, all retained in standard test. Even for the paired text scenarios, the overall standard population contains requests outside the robust base set, so subtracting overall standard and robust metrics is not a paired transformation effect.

Compared implementations

The report ranks 13 distinct implementations: Mio selector and 12 baselines evaluated through ORBIT's common interfaces. All use the same scenario-specific candidate pool and available training/development evidence.

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

The implementation audit found five groups of additional labels with identical adapters/configurations and identical model_ids, performance-score, cost-score, latency-score, and selection-policy arrays across every evaluated scenario and seed:

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 aliases of the fitted implementations in this experiment, not a claim that the original research methods are equivalent. They receive no extra rows or rank weight and are not independent reproductions of every cited method. Each adapter must declare its input features, fitted parameters, predicted scores, selection policy, and deviations from the source method.

Metrics and reporting conventions

For each method and reporting group, evaluate 100 evenly spaced nominal monetary budgets from the cheapest to the most expensive fixed-model mean cost on that group's held-out requests. These endpoints use evaluation-population cost aggregates, not training/development costs, and are shared across methods. They set the evaluator's reporting grid without supplying request-specific held-out outcomes to selectors. Write the common grid as $\mathcal G={C^{(1)},\ldots,C^{(K)}}$ with $K=100$. Predicted cost determines candidate feasibility; predicted utility or preference selects the model. Held-out outcomes then determine mean utility $Q_\pi(C)$ and realized mean cost $\bar c_\pi(C)$. A nominal budget is not the measured expenditure, and predicted-cost feasibility does not guarantee a realized-cost bound.

For this offline monetary-budget sweep, a request with no candidate satisfying the predicted cost budget falls back to the model with the lowest predicted cost; its measured quality and cost remain in the reported point. This fallback can exceed the nominal budget even in predicted cost. It is an evaluation convention, distinct from the Section 4.2 deployment rule that reports unmet constraints when no feasible candidate exists.

Let $Q_{\mathrm{strongest}}$ and $\bar c_{\mathrm{strongest}}$ be the quality and mean cost of the strongest fixed model on the same split and candidate pool. The primary metrics are:

$$Q_{\mathrm{p}\mathrm{e}\mathrm{a}\mathrm{k}}(\pi )=\underset{C\in \mathcal{G}}{\max }{Q}_{\pi }(C),{\bar{c}}_{\text{pi@}}\tau =\min \{{\bar{c}}_{\pi }(C):C\in \mathcal{G},Q_{\pi }(C)\ge \tau Q_{\mathrm{s}\mathrm{t}\mathrm{r}\mathrm{o}\mathrm{n}\mathrm{g}\mathrm{e}\mathrm{s}\mathrm{t}}\}\mathrm{.}Q\_\{\backslash mathrm\{peak\}\}(\backslash pi)=\backslash max\_\{C\backslash in\backslash mathcal\; G\}Q\_\backslash pi(C),\backslash qquad\; \backslash bar\; c\_\backslash pi@\backslash tau=\backslash min\backslash \{\backslash bar\; c\_\backslash pi(C):C\backslash in\backslash mathcal\; G,\backslash \; Q\_\backslash pi(C)\backslash geq\backslash tau\; Q\_\{\backslash mathrm\{strongest\}\}\backslash \}.$$

Cost@90, Cost@95, and Cost@100 use $\tau=0.90,0.95,1.00$, respectively. An unattained target is N/R; no extrapolation supplies a missing operating point. For a reached target:

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

Positive savings mean lower cost than always selecting the strongest fixed model; negative savings mean higher cost. Savings are also N/R when the target is unreached.

For nAUC, first form the cumulative maximum of quality in increasing nominal-budget order:

$$Q_{\pi }^{\mathrm{e}\mathrm{n}\mathrm{v}}(C^{(g)})=\underset{1\le h\le g}{\max }{Q}_{\pi }(C^{(h)})\mathrm{.}Q\_\backslash pi^\{\backslash mathrm\{env\}\}(C^\{(g)\})=\backslash max\_\{1\backslash leq\; h\backslash leq\; g\}Q\_\backslash pi(C^\{(h)\}).$$

Then integrate by the trapezoidal rule and divide by the full nominal-budget interval:

$$\mathrm{nAUC}(\pi )=\frac{1}{C^{(K)}-C^{(1)}}\sum _{g=1}^{K-1}\frac{Q_{\pi }^{\mathrm{e}\mathrm{n}\mathrm{v}}(C^{(g)})+Q_{\pi }^{\mathrm{e}\mathrm{n}\mathrm{v}}(C^{(g+1)})}{2}(C^{(g+1)}-C^{(g)})\mathrm{.}\backslash operatorname\{nAUC\}(\backslash pi)=\backslash frac\{1\}\{C^\{(K)\}-C^\{(1)\}\}\backslash sum\_\{g=1\}^\{K-1\}\; \backslash frac\{Q\_\backslash pi^\{\backslash mathrm\{env\}\}(C^\{(g)\})+Q\_\backslash pi^\{\backslash mathrm\{env\}\}(C^\{(g+1)\})\}\{2\}\; \backslash bigl(C^\{(g+1)\}-C^\{(g)\}\backslash bigr).$$

This guarantees $\operatorname{nAUC}(\pi)\leq Q_{\mathrm{peak}}(\pi)$. Fixed-model references, artificial endpoints, and realized-cost completion points never enter the router curve. The report's quality-cost plot instead uses realized mean cost on a logarithmic x-axis; its geometric area is not nAUC. For that visualization, duplicate realized-cost points retain the highest quality, dominated points are removed, and the frontier is not extrapolated. Baselines are plotted at their least-cost peak-quality points, with the strongest baseline by nAUC highlighted separately on each split.

A diagnostic oracle selects the best observed model independently for each request:

$$Q_{\mathrm{o}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{l}\mathrm{e}}=\frac{1}{\mathrm{\mid }\mathcal{X}\mathrm{\mid }}\sum _{x\in \mathcal{X}}\underset{f_i\in \mathcal{C}}{\max }{y}_i(x),Q\_\{\backslash mathrm\{oracle\}\}=\backslash frac\{1\}\{|\backslash mathcal\; X|\}\backslash sum\_\{x\backslash in\backslash mathcal\; X\}\backslash max\_\{f\_i\backslash in\backslash mathcal\; C\}y\_i(x),$$

where $\mathcal X$ is the reporting population, $\mathcal C$ is its candidate pool, and $y_i(x)$ is the held-out utility. The oracle supplies no information to a selector. An evaluation record should retain the raw AUC, nominal interval endpoints, and strongest/cheapest fixed-model and oracle diagnostics so normalization can be reconstructed.

Currency. Published cost_cny arrays always use CNY. The frozen report tables below retain their original units: USD for General Text and Multimodal, CNY for Chinese. Divide the first two scenarios' cost_cny values by 7.2 to recover those reporting units; convert budget values consistently. This is a frozen benchmark constant, not a current exchange rate or the pricing convention of the report's separate Mio 1.0 Lite experiment. A uniform currency rescaling of costs and budgets leaves nAUC unchanged. Multimodal costs are frozen USD proxies, not live service quotes.

Evaluator compatibility. The accompanying Model Intelligence Toolkit evaluator computes a realized-cost Pareto AUC with a cheapest-fixed-model reference and endpoint completion. That quantity is not the report's nominal-budget nAUC above; its threshold costs can also include the inserted reference. To match this report, use actual router budget points only, align repeated runs on their common nominal-budget grid, average the utility and realized-cost curves pointwise, and then compute peak quality, nAUC, and target costs from that aggregate curve. A field named normalized_auc in another evaluator is not sufficient to establish metric compatibility.

Repetitions and uncertainty. Multimodal uses three fitted runs for every implementation, with seeds 20260917, 20260918, and 20260919, 4,000 training requests, 45 candidates, shared CLIP text/image fusion features, and the same evaluation requests and 100-point grid for each reporting population. Fitting seeds vary; evaluation populations, source-group exclusions, and candidate identities are fixed. The reported curve is the pointwise mean of those three aligned run curves, formed before the cumulative maximum or any curve statistic is computed. Quality entries show the aggregate-curve statistic ± its exact nonparametric run-bootstrap standard error, obtained from all $3^3$ ordered resamples of the run curves. General Text and Chinese each use one fitted run per method (seed 20260917). These errors describe fitting variation, not request-sampling uncertainty. Target costs are computed from the aggregate curve; individual-run reach counts are retained only as diagnostics. Current tables contain no paired request-bootstrap intervals or statistical-significance tests.

Aggregation and latency. Macro nAUC is the unweighted mean of the six standard/pooled-robust split values, not a pooled curve over different candidate pools, units, or budget ranges. Ties share rank. Latency is auxiliary: realized latency at each budget and at peak, total router inference time, and mean router time per request are useful diagnostics, but they neither determine the primary ranking nor impose a hard latency constraint in this experiment.

Robust split metrics are computed directly from the pooled robust request population. Each robustness type is evaluated separately with cost endpoints from that type's requests. Consequently, both the nominal grid and the cumulative-maximum envelope can differ, and averaging type-level nAUC values does not generally recover pooled Robust nAUC. The Multimodal robustness types each contain 475 final-test requests; equal group sizes do not change this aggregation order.

The wider reporting protocol calls for standard, each robust type, robust pooled, each task domain, and standard-plus-robust groups. A combined Multimodal statistic must be labeled an unpaired aggregate. A weakest-group statistic takes the minimum over groups declared before final testing and names the attaining group; an empty group is unavailable, not zero. The tables below contain the released primary and Mio variation-level results, not every diagnostic proposed by that protocol.

Complete reported results

The corrected result snapshot is dated 2026-10-15. Mio selector has macro nAUC 0.6360, followed by MLP at 0.6183. Mio leads pooled General Text Robust and both Multimodal splits, but not General Text Standard or either Chinese split. It also does not lead every perturbation type: language-variation nAUC ranks second and Multimodal combined-perturbation nAUC ranks third. Public baselines use fixed configurations; Mio uses designated development data for configuration selection, not a matched hyperparameter-search budget.

The following tables are generated from the corrected, public-data evaluations. All 12 baselines were refitted per scenario; Text Mio was refitted/recalibrated, Multimodal Mio was refitted with CLIP features, and unchanged Chinese Mio predictions were reevaluated. Costs are per request in the named currency. N/R means the aggregate curve does not reach a threshold; (r/3) records individual-run reach counts. Primary-table rank is by nAUC.

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

Attribution and uncertainty protocol

This section specifies analyses supported by the benchmark design. The current primary tables do not report the paired decomposition, bootstrap confidence intervals, or a complete weakest-group analysis.

For a complete routed execution, let $y_\pi(x)$ be utility and let nonnegative analysis weights $w(x)$ sum to one over a declared population $\mathcal X$. A paired policy comparison uses:

$${\bar{y}}_{\pi }=\sum _{x\in \mathcal{X}}w(x)y_{\pi }(x),{\mathrm{\Delta }}_y=\sum _{x\in \mathcal{X}}w(x)[y_{{\pi }_1}(x)-y_{{\pi }_2}(x)]\mathrm{.}\backslash bar\; y\_\backslash pi=\backslash sum\_\{x\backslash in\backslash mathcal\; X\}w(x)y\_\backslash pi(x),\backslash qquad\; \backslash Delta\_y=\backslash sum\_\{x\backslash in\backslash mathcal\; X\}w(x)\backslash bigl[y\_\{\backslash pi\_1\}(x)-y\_\{\backslash pi\_2\}(x)\backslash bigr].$$

Weights encode the source-group and within-group estimand and must be fixed independently of final outcomes. A cluster bootstrap resamples source groups within task strata while keeping variants and competing policies' outcomes together. Repeated training seeds quantify fitting variation separately. Methods, thresholds, group definitions, and interval procedures should be frozen before final evaluation.

For paired General Text or Chinese requests, let $\mathsf T$ be a validated transformation and $\pi(x)$ the selected model index with context, service snapshot, and allowance fixed. Total paired change decomposes as:

The first term measures changed performance of the originally selected model; the second measures the additional effect of changing the choice. Their sum is the total paired change. A changed selection is therefore not itself a failure. The identity requires matched outcomes and does not apply to the independently sampled Multimodal sets. Attributing candidate-organization effects requires the staged ecosystem controls below.

Model-ecosystem and service extensions

These are controlled experiment specifications, not completed numerical evaluations. Each starts from a frozen snapshot and replays an ordered event stream, recording both effective time and information-visibility time. Conditions receive the same events and may use only evidence visible before the next decision. Preparation/update costs remain separate from subsequent execution costs.

Condition	Permitted update	Intended attribution
Frozen	None	Exposure of the original system to the event
Capability refresh	New measurements or reconstructed capability profiles	Benefit of updated capability evidence
Candidate repair	Capability refresh plus candidate-pool reconstruction	Additional effect of changing the retained action set
Selector adaptation	Prior updates plus fitting or calibration	Additional effect of adapting request-level choices
Execution recovery	Endpoint rescheduling; explicit selector reselection stated separately	Restored service versus a changed model choice

The ecosystem stream covers model arrival, withdrawal, version replacement, profile refresh, and evaluation-scope expansion. It tests capability characterization, candidate organization, and selector adaptation cumulatively.

The service stream controls load, prices, rate limits, endpoint interruption/withdrawal, SLOs, and budgets. Compare frozen endpoint assignment, same-model rescheduling with bounded recovery, and a condition explicitly permitting a new Mio selector decision. Over a common post-event horizon, report task utility, expenditure, SLO attainment, throughput, tail latency, retries, endpoint switches, reselections, and recovery time. Unrecovered episodes retain failure or censoring status. Stored outcome matrices can replay selection alternatives; they cannot establish live queueing, failover delay, cancellation, or readiness.

Reproducibility and release boundaries

Use the corrected v1.1.0-20261015 release, rather than earlier score snapshots. Resolve and pin its Git revision through revision= when loading data for a fixed experiment. manifest.json records current file checksums; release-audit.json records structural/range/group/feature validation and the limits of score verification. Split inventories and media are unchanged; Text scores and Multimodal features are corrected.

The numerical source is report/aggregate-results.json, schema routestressbench-report-v3, with the same content used to generate the report tables and figure. report/reproduction-manifest.json identifies every ID-aligned, SHA256-checked per-run prediction file. Public baseline configurations and fitting code are in the toolkit.

The toolkit's python -m examples.reproduce_report --mode replay --output runs/replay recomputes report metrics from public data and prediction arrays. --mode fit re-trains the 12 public baselines; Mio is explicitly replay-only. Library/hardware differences can affect fitted neural outputs. Mio's internal training implementation, raw responses, private source diagnostics, and completed live-service traces are not supplied. Offline score replay does not reproduce live-service experiments.

For a fully auditable new run, retain split/pair identifiers, source groups, the development-group exclusion list, model/endpoint versions and candidate membership, transformations, seeds, generation/scoring settings, price tables, response status and available latency fields, method configurations, the ORBIT/toolkit commit, source-method deviations, task-level outputs, and prediction-array hashes for alias detection. Dynamic studies additionally need event ordering, visibility times, update costs, and execution traces. This is the required experimental record, not a claim that every listed item is present in the published dataset manifest.

Upstream question and media attribution and terms remain in LICENSES.md.