ReSI-Math: Exploring Recoverable Capability Frontiers for Efficient Reasoning Improvement
| Source: arXiv AI
Tags: ReSI-Math, RLVR, Qwen3-VL-8B, MATH500, evolutionary search
ReSI-Math recovers math problems a model fails directly by searching its own reasoning trajectories, then uses them for data-efficient RLVR, lifting Qwen3-VL-8B's MATH500/WE-MATH average from 78.71 to 80.16.
Details
The paper argues that when a model fails a math problem, it may still have the capability: some failures can be recovered by searching the model's own trajectories, defining a recoverable capability frontier. ReSI-Math uses Self-Judge-guided evolutionary search to find such failures under a fixed policy and picks frontier problems for sequential RLVR as the policy changes. On Qwen3-VL-8B, crossover between correct and incorrect trajectories produced 272 verifier-confirmed correct offspring from 621 attempts (43.8%). Three one-shot RLVR stages raised the MATH500 and WE-MATH average from 78.71 to 80.16, and frontier-guided selection beat random selection 54.16 to 53.31.