benchgen/router-pilot

Benchgen Router Pilot: per-question agent rewards

A routing dataset: for each question, how every agent in a fixed pool actually performed. It is the input a model-selection policy trains on, not a question-answering dataset. Published by Benchgen, the benchmarking infrastructure for AI agents.

Each row is one task. mean_reward[i] is the accuracy of agent agent_order[i] over 3 independent attempts. Ties are preserved explicitly rather than collapsed by argmax, because on easy questions several agents are genuinely equal and pretending otherwise teaches a router a coin flip.

Companion dataset with the prompts: benchgen/router-pilot-tasks. Join on task_id.

Contents

46 tasks x 5 agents x 3 repetitions = 690 graded calls.

Facet	Breakdown
Source	HuggingFaceH4/MATH-500 14, TIGER-Lab/MMLU-Pro 10, allenai/ai2_arc 6, cais/mmlu 10, openai/gsm8k 4, opencompass/AIME2025 2
Domain	knowledge 20, math 20, reasoning 6
Difficulty	easy 9, medium 19, hard 18

Schema (rewards.jsonl)

Field	Type	Notes
task_id	string	Join key with benchgen/router-pilot-tasks, format {source}/{split}/{index}
agent_order	list of string	The 5 agent ids, in the order used by every array below
mean_reward	list of float	Accuracy of each agent over 3 attempts, 0.0 to 1.0

The pool

#	Agent	Model	Mode	$/1M out
0	frontier_a	openai/gpt-oss-120b	reasoning	0.17
1	frontier_b	deepseek/deepseek-v4-flash-0731	reasoning	0.18
2	frontier_c	google/gemma-3-27b-it	direct	0.45
4	open_mid	mistralai/mistral-nemo	direct	0.03
7	open_cheap_reasoning	inclusionai/ling-3.0-flash	reasoning	0.063

Provenance and licence

This dataset contains our own measurements (graded correctness per agent per task). It contains no question text and no raw model output.

Part	Licence	Note
Reward labels (rewards.jsonl)	CC-BY-4.0	Measurements by Benchgen. Attribute "Benchgen"
Task identity (task_id)	n/a	Identifies an upstream row, no upstream text is included

Upstream sources the tasks come from (licence recorded 2026-10-15 from each dataset's Hugging Face metadata): openbmb/RLPR-Train-Dataset Apache-2.0 (excluded, see below), HuggingFaceH4/MATH-500 MIT (inherited from MATH, the mirror lists no licence field), TIGER-Lab/MMLU-Pro MIT, cais/mmlu MIT, allenai/ai2_arc CC-BY-SA-4.0, openai/gsm8k MIT, opencompass/AIME2025 MIT.

Collected but excluded

Source	Why
rlpr	Answers are free-form prose and tables. Our string-match grader scores correct answers as wrong, so these labels are not trustworthy.

What is deliberately not here

• No question text. Upstream licences differ per source, and ARC-Challenge is share-alike. task_id identifies the upstream row so prompts can be rehydrated from the original datasets.
• No raw completions. Provider terms on republishing model outputs were not cleared. The routing signal (correctness, ties, empties, errors) survives without them.

Limitations, stated plainly

• High tie rate (43 of 46 tasks). Most agents solve most of these questions, so the per-question best agent is often arbitrary. A router trained on this alone learns little. The dataset is most useful as evidence that routing does not pay on saturated benchmarks.
• One agent dominates. frontier_a or frontier_b is the best agent on almost every dataset here, so the pool behaves as a ranking rather than a set of specialists. That is a property of the pool, and no amount of extra data changes it.
• The pool was hand-picked, not selected by measurement. Trinity's Appendix A.6 chooses agents and datasets jointly from a full screening matrix. That screening was skipped for cost.
• No direct-vs-reasoning contrast. The intended pair of one base model in both modes was dropped: every candidate -thinking slug ignored max_tokens (24k to 75k tokens against a 4,096 cap), which breaks protocol comparability.
• No code domain. Execution-based grading needs a sandbox that does not exist here.
• Small. This is a pilot, sized to a strict budget.

Protocol

Identical for every agent, because E(D, M) is not comparable otherwise: max_tokens 4096, temperature 0.1, top_p 1.0, minimal reasoning effort, 3 repetitions. Any response exceeding the token cap aborts collection rather than being recorded, so no row mixes protocols.

Grading

Answers are extracted by tagged answer, boxed LaTeX, labelled answer, then a last-number fallback that strips markup first. An empty response is never scored correct.

Run the underlying benchmarks on Benchgen

These are the public benchmarks the tasks were drawn from. Each has a page on Benchgen with a leaderboard and evaluations:

Source	Benchgen page
MATH-500	benchgen.com/benchmarks/lightman-et-al-openai/math-500
MMLU-Pro	benchgen.com/benchmarks/tiger-ai-lab-university-of-waterloo/mmlu-pro
MMLU	benchgen.com/benchmarks/hendrycks-et-al/mmlu
GSM8K	benchgen.com/benchmarks/cobbe-et-al-openai/gsm8k
ARC-Challenge	benchgen.com/benchmarks/clark-et-al-allen-institute-for-ai/arc-c
AIME 2025	benchgen.com/benchmarks/mathematical-association-of-america/aime-2025

To evaluate a router built from this data, see the Router Fidelity Benchmark.

Citation

@misc{benchgen_router_pilot_2026,
  title        = {Benchgen Router Pilot: per-question agent rewards},
  author       = {{Benchgen}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/benchgen/router-pilot}}
}