Meta paper introduces 'agent plasticity': measuring self-improving agent gains per dollar spent on learning
omarsar0 · x · 2026-10-09
Meta Superintelligence Labs released a paper on self-improving agents, tackling the question: does self-improvement actually pay off?
- Proposes agent plasticity: gains on held-out tasks per dollar spent on learning, with frozen weights and a fresh context each run.
- Key finding: the best-performing model is often a different one from the most efficient learner. In chess, Go, and Hex, Claude Fable 5 reaches the highest final score, while GPT-5.6 Sol gains the most per dollar.
- In NetHack, only Claude Opus 5.5 improves significantly, gaining 66 normalized points for about $1,073 of learning.
- The paper explores which variables (notes, skills, tool calls across runs) actually drive self-improvement.
More from coding & agent
- Pine Computer launches an AI-native cloud computer claiming 1/25 model cost and 2-5x speedup — ThePeterMick · 2026-10-10
- Running 5 coding agents in parallel isn't automation: the 4 layers of an agentic software factory — intellectronica · 2026-10-10
- AI Agent That Runs Your Apple Search Ads Workflow Via Public API — jdluk87 · 2026-10-10
- The ultimate Agent Skills resource collection: creators, benchmarks, security scanners — bibryam · 2026-10-10
- Chollet backs Busabase: open-source MIT database lets AI agents turn chat output into reusable team assets — fchollet · 2026-10-10
- Ex-OpenAI Operator Researcher Previews Real-World RL Gains on High-Precision Tasks — ZhaoMandi · 2026-10-10