Meta Researchers Unveil Branching Technique for AI Harness Optimization

Researchers from Meta and two top US universities have developed a new way to improve AI agents without modifying their underlying models. The breakthrough, published in a preprint titled 'Mixture of Self-Improving Branches for Agent Harness Optimization', involves creating multiple self-improving teams that work together to optimize the 'harness' - the code framework wrapped around large language models.

The team's approach, which builds on an earlier system called Meta-Harness, uses a technique called branching. This involves splitting the search process into several specialized branches, each of which evolves independently and refines its strategy based on past performance. A router then selects the best branch head for each incoming input.

The researchers claim that their method has delivered significant gains in accuracy on various benchmarks. On Olympiad-level mathematical reasoning, for example, they achieved a 34.8% relative improvement, moving from 46.0% to 62.0%. This was accomplished without modifying the model itself or accessing the test set.

The team behind the work includes Haoyu Dong and Zihao Lin, affiliated with Meta and UC Davis respectively. The paper has been posted on arXiv but has not yet undergone formal peer review.