WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution

TL;DR

Agents that discover their own skills from experience keep the reasoning that produced each skill scattered across optimization histories, so it cannot be reused. WikiSkill separates raw execution experience, accumulated knowledge and executable skills into three layers, continuously consolidating experience into a persistent wiki that later skill updates build on (source).

Authors & Org

Not published in anything read. The HuggingFace Daily Papers snapshot carries the title and abstract but no author list, and arxiv.org answers EGRESS_BLOCKED from this run's sandbox, so no authors or affiliation are recorded rather than guessed. HuggingFace Daily Papers, 2026-08-31, 18 upvotes; arXiv publication date 2026-08-27 (source).

Method

Three levels held separately, which is the paper's structural claim:

Layer	What it holds
Raw execution experience	what actually happened in an interaction
Accumulated knowledge (the wiki)	consolidated insight, persistent across iterations
Executable skills	the reusable resource an agent invokes
Skills and the knowledge base co-evolve: experience is consolidated into the
wiki continuously, and each subsequent skill update reads what the wiki already
holds rather than re-deriving it from an optimization history.

The stated problem it addresses is that in prior skill-discovery work "the insights that guide skill development typically remain scattered across optimization histories", which limits systematic reuse across iterations — not across tasks.

Results

The abstract reports no absolute figure, names no benchmark and names no model, and this page reports none rather than reaching for one. It states four comparative findings:

1. Outperforms state-of-the-art skill-evolution methods across the benchmarks and models tried, and beats a no-skill baseline in most model-benchmark settings — "most", not all, is the paper's own word.
2. Skill evolution complements model scaling. Larger models generally benefit more from evolved skills; and smaller models with skills can outperform substantially larger models without them.
3. Skills transfer across models and across model families.
4. Skills evolved by one model can outperform a model's self-evolved skills — which is the finding least predicted by the framing, since it says the best skill library for a model is not necessarily the one it wrote.

An ablation is reported confirming that persistent accumulation in the wiki is critical; the size of that effect is not stated in the abstract.

Significance

This is the first paper this wiki holds that measures the pattern this wiki is — LLM Knowledge Bases (LLM-curated personal wikis) describes an LLM-maintained persistent knowledge base as an operating discipline, sourced to Karpathy's LLM-wiki note and to this repository's own practice. WikiSkill puts an ablation under the central assumption: that consolidating experience into a durable document beats leaving it in the history it came from. The ablation's direction agrees; its magnitude is unpublished.

Set against Agentic Reinforcement Learning, the contrast is where the improvement is stored. Agentic RL moves the weights; this line leaves the model frozen and grows an artefact beside it, which is the same shape as the results Post-Training Scaling collects — with the difference that a wiki is readable and portable, and the transfer result above is what portability looks like when measured.

The transfer finding also cuts against a common assumption in this cluster: that self-improvement is self-referential. If another model's skills can beat your own, the artefact is a shared asset rather than a private one.

Open Questions

• What was actually measured. No benchmark, model or figure is named in the abstract, so none of the four findings above can be checked, ranked against Agentic ESOpt: Fine-Tuning Long-Horizon LLM Agents with Minimal GPU Requirements (arXiv:2608.17310) or any other result here, or placed on a scale. This page should be revisited if the full paper becomes readable.
• What stops the wiki degrading. A knowledge base that only accumulates is the failure mode this repository runs placeholder-check.py and a weekly lint against. Nothing read describes retraction, contradiction handling, or a size bound.
• Whether "most model-benchmark settings" hides a pattern. Beating a no-skill baseline in most but not all settings is the interesting half; the abstract does not say which settings lost.
• Whether cross-family transfer survives a capability gap, or only holds between models of comparable strength.

Cite

WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution.
arXiv:2608.27454 (2026).

• HuggingFace Daily Papers, 2026-08-31
• arXiv:2608.27454