DiagEvo: Diagnosis-Guided Self-Evolution via Hierarchical Error Memory
DiagEvo turns a model’s own failure history into the curriculum for self-play.
The paper says unguided self-play can stall because it does not aim future questions at unresolved reasoning weaknesses. DiagEvo adds a diagnostician that clusters recurring error causes into a hierarchical memory, marks them Active or Mastered, and uses that state to guide new question generation. With a 4B diagnostician, it reports higher mean accuracy than all baselines across nine benchmarks for Qwen3-4B, Qwen3-8B, and OctoThinker-8B. HF Daily Papers' note
The paper says unguided self-play can stall because it does not aim future questions at unresolved reasoning weaknesses. DiagEvo adds a diagnostician that clusters recurring error causes into a hierarchical memory, marks them Active or Mastered, and uses that state to guide new question generation. With a 4B diagnostician, it reports higher mean accuracy than all baselines across nine benchmarks for Qwen3-4B, Qwen3-8B, and OctoThinker-8B. HF Daily Papers' note
score 5