SPLICE: repairing failed math reasoning with solution hints boosts AIME and proofs
behrouz_ali · x · 2026-10-16
The paper introduces SPLICE, which learns from a model's failed math attempts: the model is given hints from a reference solution, repairs its reasoning, and a smaller model is then trained on the repaired reasoning traces. This improves both AIME accuracy and Olympiad-level proofs. The authors note that getting a reasoning model to actually follow hints took nontrivial effort.
More from Research
- UCSD Researchers Propose Dataset Watermarking Detectable via Black-Box API Access — yuxiangw_cs · 2026-10-16
- EmbodiedEvo tops RoboDojo as agents inherit reusable skills and lessons — Aiden_Tech_Ai · 2026-10-16
- CHI reviewer laments: papers' Discussion sections now cite nothing — IanArawjo · 2026-10-16
- Casper: Understanding 90% of neurons ≠ understanding the network — StephenLCasper · 2026-10-16
- NVIDIA Launches cuPotential in CUDA-X to Bring AI Agents to Semiconductor Materials Simulation — tomaszbednarz · 2026-10-16
- Agentic Search Systems Are Rewarded Only for Final Answers, Raising Faithfulness Concerns — burkov · 2026-10-16