ReSI-Math
A reinforcement learning training framework for mathematical reasoning in large language models. It identifies recoverable failure cases through self-judge-guided evolutionary search, and combines this with adaptive frontier problem selection for serialized RLVR training.
Timeline (last 90 days)
The ReSI-Math framework reveals that some models' failure to solve problems is not a lack of capability but rather insufficient elicitation of capability, and that a recoverable capability frontier region exists.
The core pipeline of ReSI-Math consists of two steps: self-judging-guided evolutionary search and adaptive frontier problem selection.
The approach of ReSI-Math is expected to generalize to reasoning tasks beyond mathematics.
All Facts (3)
The ReSI-Math framework reveals that some models' failure to solve problems is not a lack of capability but rather insufficient elicitation of capability, and that a recoverable capability frontier region exists.
50%UnverifiedThe core pipeline of ReSI-Math consists of two steps: self-judging-guided evolutionary search and adaptive frontier problem selection.
50%UnverifiedThe approach of ReSI-Math is expected to generalize to reasoning tasks beyond mathematics.
50%