DiagEvo: Diagnosis-Guided Self-Evolution via Hierarchical Error Memory

DiagEvo. Diagnosis-Guided Self-Evolution via Hierarchical Error Memory. By Xincheng Wei and seven others, from the Chinese University of Hong Kong, Shenzhen, the LongCat Team at Meituan, and three other institutions. Published on the first of September. No code repository is linked to it.

Self-play is an effective way for a language model to evolve itself, but left unguided, the solver's performance can stall or even decline from round to round. Existing approaches split into two camps. Unguided methods steer question generation with signals like difficulty, learnability or diversity. Those keep the questions hard and varied, yet they never say which unresolved reasoning weaknesses the later rounds ought to aim at. Guided methods get their direction from outside: human examples, document corpora, or set difficulty targets. That makes them depend on task information supplied from beyond the self-play loop. One camp has no aim. The other has an aim, but borrows it. The gap between them is a source of direction that is both specific and internal.

DiagEvo takes direction from a third place: the solver's own history of failures. A diagnostician reads that history, pulls out the causes of error that keep recurring, and files them in a hierarchical memory of error causes. The curriculum is then built from information self-play produces anyway, with no external task resources. The claim being tested is that the direction guided methods import from outside can be recovered from inside, simply by paying attention to where the solver has already gone wrong.

The memory groups related causes under skill nodes. Each cause is tracked as either Active or Mastered, depending on how self-consistent the solver is on questions targeted at it. The challenger reads those states, along with how often each cause recurs, and balances generating questions aimed at specific causes against free exploration. Targeting keeps pressure on known weaknesses. Exploration keeps the challenger from asking only about problems it already knows of.

One more filter shapes what gets kept. Double-confidence filtering retains questions of intermediate difficulty, but only when the solver's most common answer has a clear lead in the vote. Two conditions, then, must both hold: the question is neither trivial nor hopeless, and the solver's majority answer is decisive rather than a near tie.

With its default 4B diagnostician, DiagEvo beats every baseline on mean accuracy across all nine benchmarks, and it does so for each of three solvers: Qwen3-4B, Qwen3-8B and OctoThinker-8B.

Taking Qwen3-8B as the example, mean accuracy over five mathematical reasoning benchmarks reaches seventy-two point three percent, four point five percentage points above R-Zero. Over all nine benchmarks it is fifty-seven point four percent, one point one points above DARC. Ablations credit both the hierarchical error-cause memory and double-confidence filtering with part of the gain.

The move is to treat failure as the curriculum. Instead of asking an outside source what to practise next, the loop asks what it keeps getting wrong, and keeps asking until those causes are marked Mastered. Nothing new enters from outside; the solver's mistakes, sorted and remembered, do the steering that a human example or a document corpus would otherwise have done.