ObjRun AIME 2025 - oracle correctness feedback
30/30 (100%) on AIME 2025 I + II, with gold-answer correctness feedback during solving.
This repository identifies the ObjRun F evaluation system and publishes retained results. It contains no model weights and does not claim a new pretrained checkpoint.
Scope and protocol
The source is the completed historical aime-f-ns run on 2026-09-26, which covered 60 AIME 2024-2025 problems. This submission includes all 30 AIME 2025 problems only: 15 from AIME I and 15 from AIME II. All 30 have a retained passing final grader verdict.
The solver could query gold-answer correctness and revise answers during solving. The observed tools were Python execution, verification, submission, and task-tree inspection; no web-search calls were observed in the retained run. This is an oracle-assisted system success rate, not ordinary blind pass@1, zero-shot model accuracy, or a claim of superiority over standard leaderboard entries.
- Configured API model order:
deepseek/deepseek-v4-flash-0731, thendeepseek-v4-flash. These are retained configuration identifiers; per-call served checkpoint identities are not independently certified. - Temperature 0, seed 42, medium reasoning effort, maximum 16,384 tokens per request.
- 1,800 seconds per task, up to 4 branches, objective checks setting 10.
- Historical local textual renderings were used. Some wording and diagram representations differ from the MathArena dataset. This was not a run of the MathArena harness.
Retained answers and verification
On 2026-10-26, all 30 last retained submit-call answers were compared offline with the corresponding MathArena/aime_2025 reference integers. All 30 match. No new model calls were made. The mapping is AIME I question n to problem_idx=n, and AIME II question n to problem_idx=15+n.
Per-task results provides the problem identifiers, retained boxed integer answers, match outcomes, and candidate hashes. Machine-readable report records the original settings and provenance hashes. The publication contains only this report, the score table, and evaluation metadata.
This is a self-reported Hub evaluation entry without a Hugging Face verification token. Submission to the Hugging Face evaluation-results system does not imply acceptance on the MathArena website or endorsement by the benchmark authors.