AI

Don't Forget What You Just Learned: Self-Directed Repair Turns Agent Mistakes Into Lasting Lessons

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 26/10/2026, 04:03:24 UTC.

Briefings

A new mechanism called Self-Direction enables AI agents to learn from mistakes by converting temporary in-context reflections into permanent, lightweight low-rank adapters. This approach combines the efficiency of short-term retries with the stability of full-parameter fine-tuning, freezing the original model weights while training new parameters equivalent to 0.02% of the base model's size. Experiments using Qwen family models across open-ended research, tool-use workflows, and terminal tasks demonstrate that this method improves success rates by 8.9 to 17.0 percentage points and reduces repeated mistakes by nearly doubling reliability.

Key points

• Self-Direction converts temporary in-context reflections into permanent low-rank adapters by freezing original model weights.
• Agents save learned lessons in a persistent library and automatically route them to solve related problems later.

Context

Intelligent agents should learn from their mistakes so that experience leads to continuous improvement. Existing methods struggle with temporary in-context reflections being easily forgotten across sessions, while full-parameter fine-tuning is expensive and can interfere with existing capabilities. This finding addresses the gap in how AI agents can effectively utilize past errors for future improvement.

Limitations

Results are based on experiments with open-weight models from the Qwen family across three distinct research domains: open-ended research, tool-use workflows, and terminal tasks. The study does not specify the exact hardware or software environment used, nor does it report results from closed-weight models or specific commercial platforms.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 26/10/2026, 06:36:38 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeiglngp7rkk2g3qcptbkiqhy65e42vanzxm7aoyo64z7csap3ncshe