MetaSkill paper: Builder AI learns to build harnesses, boosting another AI by 12+ points
青稞AI · wechat · 2026-10-10
A writeup of the MetaSkill (AI4AI) paper frames AI-helping-AI as an Advisor–PhD relationship: a Target AI executes tasks while a Builder AI builds and improves its harness — the environment of tools, memory, and execution flows.
The Builder observes the Target's feedback, accumulates meta-skills, and turns them into usable tools and workflows rather than text-only advice. Experiments across two benchmarks and three target models show the Builder-built harness beats handing the Target the same advice by 12.02 points on average; a single model playing both roles gains 18.71 points with meta-skills. The takeaway: when discussing RSI, an agent system's ability to keep improving its own working environment matters as much as raw model capability. Paper: arxiv.org/abs/2609.38143, code open-sourced.
More from Research
- EMNLP paper: diffusion LMs internally encode latent denoising timestep representations — gsarti_ · 2026-10-23
- KU Leuven and VIB open two Group Leader professorships in AI for Biology, start 2027 — jpillowtime · 2026-10-23
- Subspace Tuning curbs LLM catastrophic forgetting by freezing principal singular directions — scottniekum · 2026-10-23
- Perceptic launches PercepticOS, an AI workspace for 'scientific superintelligence' — nathanbenaich · 2026-10-23
- EPFL's Bosselut to talk long context as test-time learning at ELLIS seminar — andre_t_martins · 2026-10-23
- DRATS: Adaptive Task Sampling Boosts Worst-Task Performance in Multi-Task RL — Cohere · 2026-10-23