asingh15/value-as-tool-rollouts

Value-as-Tool proof rollouts

Complete trajectories behind the rollout explorer Space: every model call (prompt, private reasoning, response and tool calls), every verifier verdict and value estimate, and the external judge's score. The trajectories cover olympiad proof problems from IMO-ProofBench (60) and ProofBench (145).

id	experiment	model	methods	seeds
q9_base	Qwen3.5-9B · harness methods	Qwen/Qwen3.5-9B	9	0–7
q9_attempt	Qwen3.5-9B · attempt-conditioned verifiers	Qwen/Qwen3.5-9B	24	8–15
q27_base	Qwen3.8-27B · harness methods	Qwen/Qwen3.8-27B	9	0–2

Methods include Direct generation; Generator/Verifier/Reviser loops (GVR), GVR with subagents, reference-assisted GVR and rationale+score feedback; solver-invoked value tools; a plan-work-review harness; and 24 attempt-conditioned verifier arms. In those arms the verifier sees 8 labeled prior Direct attempts (seeds 0–7 of the 9B harness run) as full solutions, solution summaries or solution+thinking summaries, with or without the reference proof. The external judge is openai/gpt-oss-20b, scoring 0–7, and strict success is 7/7.

Layout

• index/experiments.json: experiments, methods (with descriptions) and code legends.
• index/summary.json: per experiment, benchmark scope and method: 7/7 rate with 95% bootstrap intervals, tokens, how runs end, GVR verifier calls, value-tool queries, and fix and break rates against Direct. Built from each artifact's official report/rows.jsonl.
• index/<benchmark>.json.gz: per problem, one cell per rollout: [seed, judge score, status code, verdict codes, generated tokens, calls].
• problems/<benchmark>/<problem>.json.gz: statement, solver prompt, reference proof, judge grading prompts (the candidate's text replaced by ⟦final⟧), and the three prior-attempt evidence packs.
• rollouts/<experiment>/<benchmark>/<problem>/<method>.json.gz: all seeds of one method on one problem.

Trajectories are de-duplicated. Continuation calls list only the messages added after the call they extend (cont). Assistant turns that repeat an earlier response are {"ref": call}. Text shown elsewhere is replaced by markers:

• ⟦prompt⟧, ⟦problem⟧, ⟦reference⟧ and ⟦evidence⟧;
• ⟦c<i>⟧ and ⟦r<i>⟧: output and reasoning of call i;
• ⟦a<i>.<field>⟧: a tool-call argument of call i.

Files are gzipped JSON.

Provenance and terms

Generated with the value_as_tool harness from the artifacts listed in index/experiments.json. Problem statements, reference proofs and grading guidelines come from lm-provers/IMOProofBench and lm-provers/ProofBench and keep those datasets' terms. Model outputs are from Qwen models and are subject to their licenses. Scores come from an automatic judge and are noisy, so treat individual 7/7 or 0/7 grades as evidence, not ground truth.