Agent self-improvement with frozen weights: failure traces add just 4pp, blind search gets 60%
burny_tech · x · 2026-11-12
A new arXiv paper (2611.11342) tests whether an agent's harness — notes around a frozen-weight 9B model, written by a larger teacher — can genuinely learn from failure traces. On generated customer-service decisions, mistake-informed notes beat mistake-blind notes by only 4.0 percentage points, and a matched search that never sees failures already achieves 60% of the total gain. In full customer-service conversations the notes showed no demonstrated improvement; a strict lesson screen admitted only a few unverified lessons, and a neutral note outscored the learned ones. A blind audit also judged 146 of 240 tool-use test items defective — a cautionary result for the 'agents learning from experience' narrative.
More from coding & agent
- Emulating Humanlike Mouse Moves and Key Timing to Make Agents Undetectable — Darpinian · 2026-11-12
- Stanford NLP course's best homework: build your own agent harness — stanfordnlp · 2026-11-12
- Anthropic details agent memory: dreaming organizes Claude's memory, wiki+search hits 61% — RLanceMartin · 2026-11-12
- Talk preview: scaling 1,000 Fable 5.2 agents to formalize Fermat's Last Theorem — RLanceMartin · 2026-11-12
- Why Real Consumer PCs Are the Only Unblockable Way to Run Computer Use Agents — Darpinian · 2026-11-12
- Opinion: AI infra is becoming a coding runtime as intelligence units outgrow tokens — prateekj · 2026-11-12