Research2026-09-25
Researchers proposed a recursive self-improvement framework for reasoning models. It combines Dynamic Co-Evolution and Self-Refined Concise Learning. DCE lets the teacher model co-evolve with the student. SRCL trains on shorter, verified rewrites of the model’s own responses. The approach was tested on Qwen3-8B. It outperformed prior on-policy self-distillation by 35.62 percentage points. It also reduced output length by 7.80%. The paper was submitted to arXiv on 25 September 2026.
What changed
On-policy self-distillation froze the teacher model throughout training.
What it unlocks
A student model can now co-evolve with its teacher and learn from shorter, verified rewrites of its own responses.
- 65.97% Average@12 on Qwen3-8B
- 35.62 percentage points improvement over OPSD
- 7.80% reduction in mean output length
Sources