SkillGLoW: Storing shared solving procedures beats per-task skill pools, +17.2 points with 3.6x smaller library
rohanpaul_ai · x · 2026-09-06
A new arXiv paper, SkillGLoW (Global-Local Weave), tackles how self-improving LLM agents should remember what worked. Existing approaches fail on long-horizon, heterogeneous workloads in opposite ways: a global skills document collapses into generic discipline, while a per-task skill pool inflates with entries bound to the instance that wrote them.
The authors argue the right unit of reuse is the solving procedure shared by a cluster of related tasks. SkillGLoW aggregates local skills from task executions into procedural families, compresses them into de-instantiated global priors, regenerates instance details per task instead of storing them, and uses a commit gate that only admits a prior when real execution shows it doesn't degrade the deployed library.
Key results:
- Tested across four benchmarks (math reasoning, terminal automation, software repair, embodied control) and three models
- +17.2 points on hard tasks over the no-skill baseline on average, with positive gains in all 12 continual-improvement runs; 18.0 with local regeneration
- One prior per procedural family, making the library 3.6x more compact than a per-task pool
Related event: SkillGLoW: Agent Memory Shrinks 3.6x While Scores Rise 17.2 Points(2 posts)→
More from coding & agent
- NInfer fork runs 555k-token context on a single RTX 5090 with custom NVFP4 KV cache and YaRN — Lumpy-Comedian-1027 · 2026-09-06
- lume agent reads its own errors to debug itself and propose tasks — BLUECOW009 · 2026-09-06
- Agent Astra ports three.js demos to Godot in minutes, surprising one skeptical dev — teortaxesTex · 2026-09-06
- Dev builds surprisingly fun isometric puzzle-combat game entirely with GPT-6Astra High — IanArawjo · 2026-09-06
- Armin Ronacher: reasoning levels live in the system prompt, and switching them busts your KV cache — mitsuhiko · 2026-09-06
- mitsuhiko: switching reasoning levels busts KV caches; Fable 5.1 is the exception — mitsuhiko · 2026-09-06