README.md

---
license: cc-by-4.0
language:
- en
pretty_name: RoutingBench V2-beta
size_categories:
- n<1K
tags:
- multi-agent
- agent-routing
- orchestration
- decision-making
- deterministic-simulation
configs:
- config_name: decision
  default: true
  data_files:
  - split: development
    path: data/decision/development.parquet
  - split: validation
    path: data/decision/validation.parquet
  - split: test
    path: data/decision/test.parquet
- config_name: interactive
  data_files:
  - split: development
    path: data/interactive/development.parquet
  - split: validation
    path: data/interactive/validation.parquet
  - split: test
    path: data/interactive/test.parquet
- config_name: agent_catalog
  data_files:
  - split: train
    path: data/agent_catalog/train.parquet
---

# RoutingBench V2-beta

RoutingBench V2-beta is a deterministic benchmark for evaluating whether an
orchestrator can select the next appropriate agent from the task, current
observable state, execution memory, budget, and qualitative role cards.

This release is a **public research preview**. It evaluates routing independently
from actor-language-model quality: actors are deterministic state transitions,
and agent profiles contain neither backbone identities nor capability scores.

## What is included

- 120 interactive workflows across 8 workflow families.
- 809 isolated next-action decisions.
- 24 qualitative agent profiles.
- Set-valued admissible actions where multiple routes are reasonable.
- Scheduled actor failures, pre-execution routing rejection, recovery, budget,
  parallel joins, terminal control, and semantic-loop cases.
- Public oracle labels and deterministic simulator specifications.

## Configurations and splits

| Configuration | Development | Validation | Test | Purpose |
|---|---:|---:|---:|---|
| `decision` | 165 | 211 | 433 | Track A: isolated next-agent selection |
| `interactive` | 24 | 32 | 64 | Track B: complete deterministic trajectories |
| `agent_catalog` | 24 (`train`) | — | — | Qualitative role contracts |

All instances sharing a `split_group_id` remain in one split. The test split and
its labels are intentionally public in V2-beta. It is a diagnostic split, not a
sealed leaderboard test.

## Loading

```python
from datasets import load_dataset

decision = load_dataset("TrangBui/RoutingBench-V2", "decision")
interactive = load_dataset("TrangBui/RoutingBench-V2", "interactive")
agents = load_dataset("TrangBui/RoutingBench-V2", "agent_catalog")
```

The default configuration is `decision`.

## Track A: isolated routing decisions

Input to a router consists of public fields such as `task`, `observed_state`,
`task_context`, `execution_memory`, `recent_events`, `budget`, `agent_pool`, and
`candidate_actions`.

Primary label:

- `admissible_actions`: actions justified by the observable state and role
  contracts.

Diagnostic labels:

- `progress_actions`: actions that the hidden deterministic simulator knows can
  still reach the goal;
- `oracle_efficient_actions`: actions on the hidden step/cost Pareto frontier;
- `distractor_actions`, `action_outcomes`, and related validity fields.

**Do not provide diagnostic or oracle fields to the router.** The reference
evaluator constructs a public observation before calling a routing policy.

Primary metrics include admissible next-action accuracy, macro accuracy by
workflow template, ambiguous-state accuracy, post-failure and post-rejection
accuracy, counterfactual-pair accuracy, STOP precision/recall, distractor rate,
efficiency, and calibration.

## Track B: interactive trajectory evaluation

The `interactive` configuration contains deterministic transition specifications.
The evaluator starts from the initial public state, asks the router for one agent,
applies the corresponding transition, and repeats until completion or a terminal
failure such as step limit, cost limit, premature STOP, or semantic loop.

Primary metrics include goal completion, admissible/efficient action rates,
macro completion by workflow template, recovery success, action rejection,
stagnant repetitions, loop termination, and step/cost regret.

Transition tables are evaluator data and must not be exposed to the router.

## Workflow families

1. Sequential dependency
2. Parallel join
3. Conditional branch
4. Conflict resolution
5. Tool/artifact routing
6. Failure recovery
7. Terminal control
8. Budget-sensitive routing

Pool variants include base, irrelevant noise, adjacent-role noise, order
permutation, and stress pools.

## Failure observability

V2-beta distinguishes:

- `ACTION_REJECTED`: the role was rejected before actor execution;
- `EXECUTED_NO_PROGRESS`: the actor ran but made no substantive progress;
- `EXECUTED_RETRYABLE_FAILURE`: the actor ran and its method should be revised;
- `EXECUTED_INVALID_OUTPUT`: execution returned an unusable artifact;
- `EXECUTED_SUCCESS`: execution updated the workflow state.

These statuses are observable runtime telemetry, not recommended next actions.

## Baselines and orchestration ablation

The reference code supports:

- random, fixed-sequence, oracle, and decision-model routers;
- `naive`: minimal prompt, name/goal role cards, no routing guard;
- `structured`: full qualitative role contracts, failure-aware prompting, and an
  observable guard against repeating a rejected action without a state change.

The naive/structured comparison measures the combined effect of orchestration
engineering. It does not separately identify the causal contribution of prompt,
role-card detail, and harness guard.

Reference implementation:
[BuiThiThanhTrang/RACO_MAS](https://github.com/BuiThiThanhTrang/RACO_MAS)

## Construction and quality control

All task text and workflow specifications are synthetic/original. MuSiQue, GAIA,
and prior project runs informed abstract workflow and failure patterns; no source
questions, answers, passages, or copyrighted attachments were copied.

Generation uses seed 42. Release checks include deterministic generation, JSON
schema validation, grouped splits, zero observable-state label collisions, quota
checks, checksums, and oracle completion.

See `validation_report.json`, `quota_report.json`, `split_manifest.json`,
`release_manifest.json`, and `checksums.sha256`.

## Intended use

- Evaluate state-aware next-agent routing.
- Compare routing prompts, role-card designs, and observable harness rules.
- Study robustness to overlapping roles, irrelevant agents, order changes, and
  deterministic failures.
- Prototype benchmark contracts before larger hidden-test releases.

## Out-of-scope use and limitations

- V2-beta is not a complete measure of real-world multi-agent quality.
- It does not measure actor reasoning quality, factuality, tool reliability, or
  natural-language collaboration quality.
- Workflows are synthetic and deterministic and currently cover abstract workflow
  families rather than broad real-world domains.
- Test labels are public and have been used for diagnostic analysis; do not use
  this release for sealed-test or leaderboard claims.
- Decision-model performance may depend on provider behavior and model version.

More detail is available in `docs/DATASHEET.md`, `docs/EVALUATION.md`,
`docs/LIMITATIONS.md`, and `docs/REPRODUCIBILITY.md`.

## License

The dataset, task text, profiles, and deterministic workflow specifications are
released under CC BY 4.0. See `LICENSE_DATA`.

## Citation

```bibtex
@dataset{bui2026routingbench,
  author    = {Trang Bui},
  title     = {RoutingBench V2-beta: A Deterministic Benchmark for State-Aware Agent Routing},
  year      = {2026},
  version   = {2.0.0-beta.1},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/TrangBui/RoutingBench-V2}
}
```