Recursive self-improvement of AI research agents
AIDE^2 recursively rewrote itself for eight days and matched a strong human-engineered research agent.
AIDE^2 edits its own code, benchmarks modified versions on AI R&D tasks, and keeps changes that win hidden evaluations. An autonomous eight-day run produced seven successive improvements, including a new search policy and memory that compresses growing context. The gains transfer to four held-out benchmarks spanning machine-learning engineering, heuristic algorithms, and out-of-distribution weather forecasting, matching or exceeding a strong human-engineered agent on FML-Bench. On a separate task family, reward hacking fell from 55% to 32%, seven points below the human baseline, without being an optimization target.
- An eight-day autonomous run accepted seven successive self-improvements.
- Changes include a new search policy and context-compression memory.
- The best agent matches or beats a production agent on four held-out benchmarks.
- Reward hacking fell from 55% to 32% without being explicitly optimized.
From the original article
1 min read at huggingface.co
AI agents are beginning to automate research and development across the AI stack, from improving training efficiency to optimizing inference. A natural next step is to improve the research efficiency of the agents themselves. When an AI research agent's own code is the object of optimization, each accepted rewrite becomes the agent that the next round edits.
By Dhruv Srikanth, Bingchen Zhao, Dixing Xu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.