Glonce

DiagEvo turns solver failure history into a self-play curriculum

researchThe Glonce editorsHugging Face🔥71Glonce rating71Impact65Novelty75Relevance95Surprise45Players30Credibility80
DiagEvo turns solver failure history into a self-play curriculum

Self-play is a common way to make a language model improve itself without new labeled data, but performance often plateaus or even declines across rounds if question generation has no guidance. Existing unguided methods steer generation with signals like difficulty, learnability, or diversity, which keep questions varied but do not point at which specific reasoning weaknesses the next round should fix. Guided methods do target weaknesses, but only by pulling in external task resources such as human-written examples, document corpora, or preset difficulty targets, meaning they depend on information from outside the self-play loop.

The researchers introduce DiagEvo, which gets that missing direction from the solver's own failure history instead. A diagnostician component extracts recurring causes of error from past failures and stores them in a hierarchical error-cause memory: related causes are grouped under skill nodes, and each node is tracked as Active or Mastered based on the solver's self-consistency on questions that target it. A challenger component then uses these node states and how often each cause recurs to balance two things: generating new questions aimed at unresolved causes, and free exploration of new territory. A filtering step called double-confidence filtering keeps only intermediate-difficulty questions, defined as those where the solver's most common answer has a clear lead in votes over other answers. The result is a curriculum built entirely from information produced during self-play, with no external task resources at any point.

With its default 4B-parameter diagnostician, DiagEvo outperforms every baseline in mean accuracy across all nine evaluation benchmarks, for each of three different solver models: Qwen3-4B, Qwen3-8B, and OctoThinker-8B. Using Qwen3-8B as the solver, it reaches 72.3% mean accuracy across five mathematical reasoning benchmarks, 4.5 percentage points above the R-Zero baseline. Across the full set of nine benchmarks, its mean accuracy is 57.4%, 1.1 percentage points above the DARC baseline. Ablation experiments show that both core components, the hierarchical error-cause memory and the double-confidence filtering, each contribute to these accuracy gains.

Key facts

  • DiagEvo derives its self-play curriculum entirely from the solver's own failure history, without relying on external task resources such as human examples or document corpora.
  • A diagnostician extracts recurring error causes into a hierarchical error-cause memory, grouping them under skill nodes tracked as Active or Mastered by self-consistency.
  • With a default 4B diagnostician, DiagEvo beats every baseline in mean accuracy across all nine benchmarks for three solvers: Qwen3-4B, Qwen3-8B, and OctoThinker-8B.
  • On Qwen3-8B, DiagEvo reaches 72.3% mean accuracy across five mathematical reasoning benchmarks, 4.5 percentage points above R-Zero.
  • Across all nine benchmarks DiagEvo's mean accuracy is 57.4%, 1.1 percentage points above DARC, with ablations confirming both the error memory and the double-confidence filtering contribute to the gains.

Why it matters

Self-play training for language models tends to stall without guidance: signals like difficulty or diversity keep questions varied but never say which specific reasoning weakness to fix next, and the alternative, guided methods, has always meant importing external task resources such as human examples or document corpora. DiagEvo shows that the missing direction can instead come from the solver's own record of mistakes, closing the self-play loop without any outside task data.

Who it affects

The result is aimed at researchers and teams building self-improving training pipelines for reasoning models, particularly anyone relying on self-play to push math or reasoning accuracy without a supply of external training data or human-curated examples.

How to use it

DiagEvo is presented as a research method rather than a released product; no pricing, licensing, or deployment details are given. Its default configuration pairs a solver model (tested with Qwen3-4B, Qwen3-8B, or OctoThinker-8B) with a 4B-parameter diagnostician model that maintains the hierarchical error-cause memory driving question generation.

How solid is it

The evaluation covers three solver models across nine benchmarks. With the default 4B diagnostician, DiagEvo outperforms every baseline in mean accuracy across all nine benchmarks for each solver. On Qwen3-8B it posts 72.3% mean accuracy on five mathematical reasoning benchmarks, 4.5 percentage points above the R-Zero baseline, and 57.4% mean accuracy across all nine benchmarks, 1.1 percentage points above the DARC baseline. Ablation experiments isolate the contribution of both the hierarchical error-cause memory and the double-confidence filtering step, and both are shown to matter.

Risks and caveats

The identities of the nine benchmarks are not specified beyond the five mathematical reasoning benchmarks used for the 72.3% figure, and no description is given of what the R-Zero or DARC baselines actually are, which limits how the gains can be independently checked. The gains are also measured against those two specific baselines and three specific solver backbones; how the method generalizes beyond them is not addressed in the available text.