Researchers from Meta, Duke University and the University of California, Davis have developed a system of self-improving branches that automatically optimizes the framework surrounding a large language model. Results show a 34.8% relative improvement in Olympiad-level math reasoning, raising accuracy from 46.0% to 62.0% with the Gemini 3 Flash model, without retraining the model itself. The approach also delivered gains of 11.6% on Terminal-Bench 2.0 and 3.8% on SWE-bench Lite. The system, published as a preprint on arXiv on September 29, 2026, succeeds Meta-Harness from March 2026 by splitting the search into multiple specialized branches that learn from their own history.
Source: Read the original article
Disclaimer: this content is for information purposes only and is not financial advice. Cryptocurrencies are highly volatile: you may lose all of your capital. Always do your own research. Legal notice
