Researchers from UIUC and Chroma have introduced Harness-1, a 20 billion parameter retrieval subagent. This system utilizes reinforcement learning within a stateful search harness, maintaining internal bookkeeping such as candidate pools and evidence graphs. Harness-1 has demonstrated strong performance, achieving an average curated recall of 0.730 across eight benchmarks, and its weights and code have been made publicly available. AI
IMPACT Provides a new open-source retrieval subagent for AI research and development.
RANK_REASON Release of an open-source model with performance benchmarks. [lever_c_demoted from research: ic=1 ai=1.0]
Read on Mastodon — sigmoid.social →
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →