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.

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.

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, hard 18, medium 19

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

Collected but excluded

These were collected and are not in this release, because a label we cannot trust is worse than no label.

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: ARC-Challenge is share-alike, and MMLU-Pro's licence was never verified. 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 (43/46 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-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.