TrangBui/RoutingBench-V3

RoutingBench V3-beta

RoutingBench V3-beta is a public research preview for evaluating whether an orchestrator can dynamically select the next appropriate agent from the task, current observable state, execution memory, budget, and qualitative role cards.

Human review and adjudication remain pending for all 116 workflows. Formal validation has passed, but this release must not be described as human-validated or as the final RoutingBench V3 benchmark.

What is included

• 116 workflows across 10 workflow families.
• 1,156 isolated next-action decisions with set-valued valid actions.
• 52 workflows derived from licensed real tasks: 24 MuSiQue and 28 MMLU-Pro.
• 39 qualitative agent profiles without capability scores or backbone identity.
• Track A for isolated next-agent selection.
• Track B specifications for live actor execution and separate routing, execution, completion, semantic-accuracy, and failure-origin metrics.
• Public provenance, source registry, review queue, validation report, schemas, checksums, and reproducible baseline configurations.

Configurations and splits

Configuration	Development	Validation	Test	Purpose
decision	320	322	514	Track A routing states and public oracle labels
interactive	32	32	52	Track B workflow specifications
actor_inputs	16	16	20	Inputs exposed only after an actor is selected
evaluation	16	16	20	Gold sidecar; never a router or actor input
provenance	32	32	52	Source ID, hash, license, and derivation
human_review	32	32	52	Pending review checklist
agent_catalog	39 (train)	—	—	Qualitative role contracts

The test split and labels are public in this beta. It is not a sealed leaderboard split.

Loading

from datasets import load_dataset

decisions = load_dataset("TrangBui/RoutingBench-V3", "decision")
workflows = load_dataset("TrangBui/RoutingBench-V3", "interactive")
agents = load_dataset("TrangBui/RoutingBench-V3", "agent_catalog")

The default configuration is decision.

Nested state fields

Dynamic nested objects are stored as canonical UTF-8 JSON strings in columns ending in _json. This keeps the Parquet schemas stable across workflow families and compatible with the Dataset Viewer without changing their contents. Decode them after loading, for example:

import json

state = json.loads(decisions["development"][0]["observed_state_json"])
transitions = json.loads(workflows["development"][0]["transitions_json"])

Track A: isolated routing

The router receives the task, observable workflow state, recent events, execution memory, qualitative agent profiles, candidate actions, and remaining call budget. The primary metric is whether the selected action belongs to the set-valued valid_actions label.

Do not provide action_rationales, valid_actions, evaluation gold, transition tables, or provenance-only fields to the router.

Track B: live trajectory evaluation

Track B starts from the initial state, calls the selected actor, updates public state from the observed result, and routes again. Actors are real model calls; their outputs are not simulated. Report these dimensions separately:

• routing quality;
• actor contract success;
• workflow completion;
• semantic end-task accuracy;
• router/executor/mixed failure origin.

STOP is a reserved zero-cost terminal decision outside the actor-call budget. If a final artifact is committed on the last allowed actor call, the router still receives one explicit terminal decision. Selecting a non-STOP action in that state produces TERMINAL_ACTION_ERROR.

Reference baselines

GPT-6 Luna Decisions was evaluated through OpenRouter. Track B used Gemini 2.5 Flash for every actor.

Track/split	Router profile	Primary routing score	Completion	End-task accuracy
A/development	naive	81.56% valid action	—	—
A/development	structured	80.94% valid action	—	—
B/validation real tasks	naive	78.05% valid transition	62.50%	62.50%
B/validation real tasks	structured	66.33% valid transition	68.75%	68.75%

Structured routing substantially improves STOP precision on Track A (94.12% versus 71.11%) but does not improve overall valid-action accuracy. These results are intentionally reported without a subjective weighted score. Exact summaries and resolved configurations are included under baselines/.

Construction and validation

• Formal transition preconditions derive set-valued valid actions.
• No preferred actions, numeric action utilities, capability scores, or actor backbone identities are exposed to the router.
• Official answers are isolated in the evaluation configuration.
• Split groups do not cross partitions; selected MMLU-Pro domains are disjoint across splits.
• Formal validation reports zero observable-state label collisions.
• Oracle development Track A reaches 100%; random reaches 11.875%.
• All 116 review records remain pending.

See validation_report.json, quota_report.json, source_registry.json, release_manifest.json, and checksums.sha256.

Known limitations

• This is a small beta intended to validate benchmark meaning and mechanics, not a conference-ready final benchmark.
• Human review and adjudication are pending.
• v3b-musique-3hop-06 is a priority review item because the supplied corpus may not express one bridge fact directly enough.
• Live Track B scores depend on both the router and actor model. Gemini may occasionally fail structured JSON serialization on LaTeX-heavy tasks.
• The public test labels make this release unsuitable for a sealed leaderboard.

Licensing and attribution

This is a mixed-license release:

• project-authored workflow specifications, annotations, schemas, and documentation: CC BY 4.0;
• MuSiQue-derived records: upstream CC BY 4.0;
• MMLU-Pro-derived records: upstream MIT.

Per-record source identifiers and licenses are provided in the provenance configuration. See LICENSE_DATA and source_registry.json.

Reference implementation: BuiThiThanhTrang/RACO_MAS

Citation

Until a paper citation is available, cite this versioned dataset release and the applicable upstream datasets. A machine-readable citation is included in CITATION.cff.