Google's WikiSkill: Persistent Wikis Let Agents Turn Experience into Reusable Skills
Google Research released the WikiSkill preprint, arguing that long-running agents should not rely solely on weights or ever-growing context: execution history should be distilled into a maintainable knowledge base and compiled into reusable skill files. The work was independently covered by @montymike, @Crescitaly, @tuvllms, @dairai, @SabooShubham, and @xiaohu; @tuvllms notes a key motivation is that agents editing prompts or skill files typically discard logs after failures, losing lessons.
Confirmed
- Three-layer design: raw execution experience, accumulated knowledge (wiki), and executable skills are managed separately, with a persistent wiki guiding skill updates to combat knowledge forgetting.
- Evaluation: validated on 5 benchmarks and 5 models, with authors reporting consistent superiority over current skill-evolution SOTA.
- Small beats large: per @montymike, a 9B model with evolved skills beats a 27B model; @SabooShubham frames it as a skilled small model defeating a 3x-larger unskilled one; @tuvllms adds an average accuracy gain of 23 on a 27B model (as stated in the post).
- Ablation: removing the wiki eliminates the gains, cited as evidence of its effectiveness.
- Cross-model transfer: evolved skills transfer across model families.
- Scaling gains: @tuvllms notes gains do not diminish—and may increase—with model scale, akin to chain-of-thought.
- Negative finding: @montymike reports skills produced by weak models can degrade Gemini's performance.
Why it matters
- Persistent knowledge accumulation offers a path from context-stuffed experience to maintainable knowledge assets, giving small models low-cost competitiveness, with scaling-compatible gains.
- Cross-model transfer suggests skills as model-independent assets, while the weak-skill backfire finding highlights quality control. All conclusions come from the authors' preprint and await peer review and independent replication.
2026-08-30 ~ 2026-09-01 · 11 related posts
Primary sources
- Google's WikiSkill: a 9B model beats a 27B one, but one bad skill craters Gemini to 18.1% — montymike ·
- Preprint: smaller models with evolved skills beat larger ones; WikiSkill keeps a persistent wiki — Crescitaly ·
- WikiSkill Paper: Compiling Agent Experience into Persistent Knowledge for Skill Evolution — tuvllms ·
- [source] Preprint: smaller models with evolved skills beat larger ones; WikiSkill keeps a persistent wiki — Crescitaly · 2026-08-30
- Google WikiSkill: Co-evolving Agent Skills with a Persistent Knowledge Base — tuvllms · 2026-08-31
- [source] Google's WikiSkill: a 9B model beats a 27B one, but one bad skill craters Gemini to 18.1% — montymike · 2026-08-31
- WikiSkill: Compiling Agent Experience into Persistent Knowledge — tuvllms · 2026-09-01
- WikiSkill Gains Increase with Model Scale, Outperforming SOTA — tuvllms · 2026-09-01
- Google Research's WikiSkill uses logs to teach agents new skills — tuvllms · 2026-09-01
- Google Research: Agents improve by maintaining a wiki — Saboo_Shubham_ · 2026-09-01
- Google WikiSkill: Persistent knowledge bases boost agent evolution — Saboo_Shubham_ · 2026-09-01