The Big Picture
Public version histories of skills contain useful, complementary update knowledge to runtime traces; distilling and adaptively fusing that history into self‑evolution improves skill updates across tasks and models.
ON THIS PAGE
The Evidence
Public skill version histories and execution traces expose different kinds of fixes: traces concentrate on a few instruction-level fixes seen in current runs, while public histories cover a wider set of changes (scripts, configs, intents). Turning raw diffs into structured, reusable update events and selectively combining them with trajectory-derived proposals leads to consistently better automatic skill updates. Across several benchmarks and models, adding distilled historical experience historical experience improved evolution outcomes and helped evolved skills transfer better to other models.
Not sure where to start?Get personalized recommendations
Data Highlights
1Compared 400 public version-change units against 400 trajectory-derived changes to show complementary coverage between the two sources.
2Evaluated the method on 5 benchmarks using 4 large language models (Qwen3.5-27B, GPT-5.2, DeepSeek-v3.2, claude-sonnet-5).
3Integrated the approach with 3 different self-evolution methods and found consistent improvements; ablating distilled experience or adaptive attention reduced gains across benchmarks.
What This Means
Engineers building AI assistants or agent systems who want more reliable, maintainable skills should care—this shows a practical way to reuse public maintenance history to guide automated fixes. Technical leaders working on agent reliability and teams that maintain skill libraries can adopt distilled version-change experience to reduce blind, overfit edits and improve cross-model robustness LLM-as-Judge Pattern.
Key Figures
Fig 2: Figure 2: Overview of VCE-Skill. The left stage distills the version-change experience from public skill versions. The right stage enhances the skill self-evolution by allocating adaptive attention between external experience and trajectory-derived proposals.
Fig 3: Figure 3: Illustration of experience distillation.
Fig 4: Figure 4: The performance of different variants of VCE-Skill.
Fig 6: Figure 2: Skill counts, substantive version updates, and affected-component shares across 12 task domains in the broader public-history survey. Component shares are non-exclusive because one update may affect several artifacts.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreYes, But...
Public version histories are noisy and repository-specific; raw diffs must be abstracted before reuse, or they may inject irrelevant or brittle changes. The experiments cover five benchmarks and four models, so results may vary for other domains, models, or private skill repositories. Legal, licensing, or compatibility concerns around reusing public code or content were not the paper's focus and need attention in production use. noisy histories
Methodology & More
The work shows that past public edits to skill repositories are a helpful, distinct source of guidance for automated skill repair and improvement. Rather than treating raw diffs as directly usable, the method first distills adjacent version changes into structured update events (what component changed, what operation was applied, and the likely intent). These events are generalized into an experience bank of reusable update patterns. During iterative self‑evolution, a base evolver still generates proposals from recent execution traces; an adaptive selection mechanism then retrieves relevant historical experience and fuses it with the trace-derived proposal using attention weights that are adjusted based on feedback from each optimization step. A pilot comparison used 400 matched public changes and 400 trajectory-derived changes and found complementary coverage: execution traces focused on a narrow set of instruction edits, while public histories included broader fixes (scripts, checks, templates). The full evaluation applied the distilled, adaptive approach adaptive approach across five task benchmarks and four language models, and plugged it into three different self-evolution pipelines. Results show consistent improvements over the base evolvers and better transfer of evolved skills to other models. Practically, the takeaways are: keep a clean distillation step to turn diffs into portable advice, use an adaptive filter to avoid irrelevant history, and treat historical experience as a prior rather than a replacement for execution evidence.
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
All authors have very low or unspecified h-indices and no listed affiliations; arXiv preprint and lack of clear institutional signals indicate emerging/limited credibility.