The paper introduces a method for improving large language model reasoning through . The standard approach freezes a teacher model that has access to the ground truth solution, but the authors show this creates a mismatch as the student learns to reflect and self-correct while the teacher cannot. They propose Dynamic Co-Evolution (DCE) where the teacher model is periodically updated to the student's checkpoint, allowing it to provide better guidance on the student's current behavior. A second component, Self-Refined Concise Learning (SRCL), trains the model on shorter, verified rewrites of its own responses to counteract verbosity. On Qwen3-8B across four competition-level math benchmarks, the combined method improves average accuracy by 35.62 percentage points over the frozen teacher baseline while also reducing output length by 7.80%.
Co-evolving teacher and student models improves mathematical reasoning
A recursive self-distillation method substantially outperformed a frozen-teacher baseline across four competition-level mathematics benchmarks.
Big Tech
Meta AI · University of California, Riverside
Research Digest··2 min read
In one line
A recursive distillation method with a co-evolving privileged teacher improves reasoning accuracy more than using a frozen teacher.
Before you read this
Three ideas this paper builds onrecursive self-improvementAn AI system repeatedly uses its own outputs to enhance its capabilities, creating a feedback loop for autonomous improvement.66 papers →on-policy distillationA training method where a student model learns from a teacher using its own current policy's outputs, often improving reasoning and instruction following.131 papers →co-evolving feedbackFeedback that evolves alongside the policy or components it guides to improve performance.43 papers →
Read it as
Yin et al. modify on-policy self-distillation so that a solution-informed teacher periodically incorporates improvements learned by its student.
The authors begin with on-policy self-distillation, in which a student generates reasoning from a problem while a second copy of the model sees both the problem and a verified solution.
How it works
stage by stageThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.