Microsoft has developed Agent Lightning v1.0, a system that connects harnesses to reinforcement learning for agent training. This new work utilizes an endpoint proxy to integrate any harness, enabling the trainer to interact with the environment loop. By employing this method with modest compute and 6,000 training examples, Microsoft successfully improved the performance of Qwen3.5-9B on the SWE-bench Verified benchmark from 41.8% to 56.4%. AI
IMPACT This development could enhance agent training efficiency and performance on coding benchmarks, potentially influencing future AI agent capabilities.
RANK_REASON The item describes a new research work from Microsoft detailing a method for agent training and its performance improvement on a benchmark. [lever_c_demoted from research: ic=1 ai=1.0]
Read on X — Omar Sanseviero (HF research) →
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
