Self‑play is a powerful paradigm for language‑model self‑evolution, yet without guidance the solver often plateaus or degrades after several rounds. Unguided approaches steer question generation using signals such as difficulty, learnability or diversity; these keep the questions challenging but do not indicate which reasoning weaknesses should be targeted later. Guided methods rely on external resources—human examples, corpora, or preset difficulty targets—thus requiring information outside the self‑play loop.
DiagEvo addresses this by deriving direction from the solver’s own failure history. The diagnostician extracts recurring error causes and stores them in a hierarchical error‑cause memory. This memory groups related causes under skill nodes and marks each as Active or Mastered according to self‑consistency on targeted questions.
The challenger uses these states and recurrence counts to balance cause‑targeted generation with free exploration. A double‑confidence filter retains only intermediate‑difficulty questions, i.e., when the most common solver answer has a clear vote lead. Consequently the curriculum is built entirely from information produced during self‑play, without external task resources.
With the default 4B diagnostician, DiagEvo outperforms every baseline in mean accuracy across nine benchmarks for three solvers (Qwen3‑4B, Qwen3‑8B, OctoThinker‑8B). On Qwen3‑8B it reaches 72.3% mean accuracy on five mathematical‑reasoning benchmarks, 4.5 percentage points above R‑Zero. Its overall mean accuracy across all nine benchmarks is 57.4%, 1.1 points above DARC. Ablation studies confirm that both the hierarchical error‑cause memory and double‑confidence filtering contribute to these gains.
Review: DiagEvo leverages internal error memory to create a self‑consistent evolution path, eliminating the need for external annotations and delivering robust improvements across models and tasks.