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

Hugging Face Daily Papers Papers
self-evolution error-memory language-model self-play diagnosis-guided double-confidence-filtering hierarchical-memory

Summary

DiagEvo improves language-model self-evolution by deriving training direction from internal failure history via hierarchical error-cause memory and double-confidence filtering, outperforming baselines that rely on external resources.

Self-play is an effective paradigm for language-model self-evolution, but without guidance, solver performance can plateau or decline across rounds. Unguided methods steer question generation with signals such as difficulty, learnability, or diversity. These signals keep questions challenging and varied but do not specify which unresolved reasoning weaknesses later rounds should target. Guided methods obtain direction from external task resources, including human examples, document corpora, or specified difficulty targets, and therefore rely on task information supplied outside the self-play loop. We show that the needed direction can instead be derived from the solver's own failure history. We introduce DiagEvo, whose diagnostician extracts recurring error causes from this history and stores them in a hierarchical error-cause memory. The memory groups related causes under skill nodes and tracks 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. Double-confidence filtering retains intermediate-difficulty questions only when the most common solver answer has a clear vote lead. DiagEvo derives its curriculum from information produced during self-play, without external task resources. With the default 4B diagnostician, DiagEvo outperforms every baseline in mean accuracy across all nine benchmarks for each of the three solvers: Qwen3-4B, Qwen3-8B, and OctoThinker-8B. On Qwen3-8B, it reaches 72.3% mean accuracy across five mathematical reasoning benchmarks, 4.5 percentage points above R-Zero. Its mean accuracy across all nine benchmarks is 57.4%, 1.1 percentage points above DARC. Ablations show that the hierarchical error-cause memory and double-confidence filtering both contribute to these gains.
Original Article
View Cached Full Text

Cached at: 09/02/26, 03:43 AM

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

Source: https://huggingface.co/papers/2609.00768 Published on Sep 1

·

Submitted byhttps://huggingface.co/dingyii

dingyion Sep 2

Abstract

DiagEvo improves language-model self-evolution by deriving training direction from internal failure history via hierarchical error-cause memory and double-confidence filtering, outperforming external-resource baselines.

Self-playis an effective paradigm for language-model self-evolution, but without guidance, solver performance can plateau or decline across rounds. Unguided methods steer question generation with signals such as difficulty, learnability, or diversity. These signals keep questions challenging and varied but do not specify which unresolved reasoning weaknesses later rounds should target. Guided methods obtain direction from external task resources, including human examples, document corpora, or specified difficulty targets, and therefore rely on task information supplied outside theself-playloop. We show that the needed direction can instead be derived from the solver’s own failure history. We introduce DiagEvo, whosediagnosticianextracts recurring error causes from this history and stores them in a hierarchicalerror-cause memory. The memory groups related causes under skill nodes and tracks each as Active or Mastered according toself-consistencyon targeted questions. Thechallengeruses these states and recurrence counts to balance cause-targeted generation with free exploration.Double-confidence filteringretains intermediate-difficulty questions only when the most common solver answer has a clear vote lead. DiagEvo derives its curriculum from information produced duringself-play, without external task resources. With the default 4Bdiagnostician, DiagEvo outperforms every baseline in mean accuracy across all nine benchmarks for each of the three solvers: Qwen3-4B, Qwen3-8B, and OctoThinker-8B. On Qwen3-8B, it reaches 72.3% mean accuracy across five mathematical reasoning benchmarks, 4.5 percentage points above R-Zero. Its mean accuracy across all nine benchmarks is 57.4%, 1.1 percentage points above DARC. Ablations show that the hierarchicalerror-cause memoryanddouble-confidence filteringboth contribute to these gains.

View arXiv pageView PDFProject pageAdd to collection

Get this paper in your agent:

hf papers read 2609\.00768

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2609.00768 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2609.00768 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2609.00768 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

Selection Is Not Learning: Choosing Model Futures

Hugging Face Daily Papers

The paper identifies a selection-to-update gap in self-evolving language model loops and proposes HeirCommit and EvoHeir, which separate parameter and experience inheritance with fresh evaluation to confirm updates, reducing harmful inheritance and improving math and general reasoning performance.

CurateEvo: Data-Curation Evolving for Agentic Post-Training

arXiv cs.CL

CurateEvo is a failure-driven dynamic evolution framework for agentic post-training data curation. It iteratively rewrites curation strategies using failed trajectories, improving effectiveness and efficiency on benchmarks like ACEBench-Agent and BFCL-V4.