Self-Direction lets agents learn from mistakes, boosting success rates
Self‑Direction is a new mechanism that lets agents learn from their own mistakes without expensive fine‑tuning. When an agent fails, it reflects on the trajectory, extracts a lesson, and distills it into a lightweight low‑rank adapter. The adapter adds only about 0.02% of the base model’s parameters and keeps the original weights frozen.
Key points
- Self‑Direction uses low‑rank adapters that add only ~0.02% of base model parameters.
- Experiments show 8.9–17.0 percentage‑point improvement in success rates.
- Adapters generalise to dozens of new tasks without extra training.
The authors tested Self‑Direction on Qwen family models across three benchmarks: GAIA for open‑ended research, Claw‑Eval for tool‑use workflows, and Terminal‑Bench 2.1 for terminal tasks. The method raised overall success rates by 8.9 to 17.0 percentage points and nearly doubled the reliability of an agent’s corrections. Learned adapters were stored in a persistent library and could generalise to dozens of new tasks without further training.
This approach offers a persistent, low‑cost way for agents to remember lessons across sessions, reducing the need for full‑parameter fine‑tuning and making open‑weight models more reliable in real‑world applications.
The story so far
6 episodes →- Self-Direction lets agents learn from mistakes, boosting success ratesthis story
Coverage and discussion
2sources- RSI-Opt: Recursive Self-Improvement Optimizer Agents For Model TrainingPrimary source · arXiv cs.AI ·
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- Researchers release LAMB, a Latin language model trained on 74B tokens · 1 src
- Researchers explore how aligned LLMs generalize to new tasks · 1 src
- Researchers propose agentic framework for emotion reasoning with evidence checks · 1 src
- Researchers propose 24 quality dimensions for knowledge graphs used by LLMs · 1 src
- Researchers introduce Schema Coherence Inventory to test belief consistency in language models · 1 src
Comments
via GitHub Discussions