Google WikiSkill Shows Smaller Models Beating Larger Ones via Skill Evolution
bittingthembits · x · 2026-08-29
Google Research released WikiSkill, splitting agent intelligence into raw experience, a knowledge base, and reusable skills. Experiments show a 9B model with evolved skills scored 47.4%, beating a 27B model without skills at 39.4%. Skills transfer across model families; on ALFWorld, a skill from a 27B model boosted a 9B model to 70.2%, outperforming the 9B model's self-generated skill. This suggests performance gains come from inheriting accumulated knowledge rather than just scaling model size.
More from coding & agent
- Are the fastest GPU kernels now written by AI? Kernels are hyper-verifiable, argues one dev — fleetwood___ · 2026-10-16
- Anthropic lead engineer's 4-question prompt makes agents more ambitious, now a free skill — daniel_mac8 · 2026-10-16
- ChainForge dev shows 'feature map' for orchestrating subagents on large PRs — IanArawjo · 2026-10-16
- aiDotEngineer benchmarks agent orchestration pipelines, staging humans as Foreman and workers — vikvang1 · 2026-10-16
- Developer mass-produces 200 videos in hours using Claude Code + Hyperframes — mcraddock · 2026-10-16
- Penny University hosts software factory discussion; dev shares $40/mo vendor-agnostic AI stack — JnBrymn · 2026-10-16