AI BriefingPickharness2: Fix Model Weights, Recursively Optimize Only the Harness for 11.5-Point Performance Gain
google-research/harness2
About the project
A framework that recursively optimizes only the surrounding system, the harness, instead of fine-tuning the model itself to improve LLM agent performance. It modifies eight components, including prompts, memory, and tool configurations, and compares execution results to find the most efficient combination.
Designed to operate in open-world environments without ground-truth data. It iterates through Propose-Probe-Compose stages, refining the harness based on execution evidence, and features a continuous improvement loop that applies lessons learned from previous tasks to subsequent ones.
Demonstrated performance improvements of up to 11.5 points on WorkBuddy-Bench and 4.5 points on JobBench using GLM-5 and Gemini 3.5 Flash models. Supports parallel and sequential recursive modes, enabling evaluation of legal, office, and coding-related professional benchmarks via OpenCode and Codex runners.
google-research/harness2
The original page has no description.
Python
This introduction was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report errors, attribution issues, or removal requests via Contact.