Harness-1: a 20B retrieval sub-agent trained with reinforcement learning for stateful search
Original title:Meet Harness-1: A 20B Retrieval Subagent Trained With Reinforcement Learning Inside a Stateful Search Harness on gpt-oss-20b
AI summary
UIUC and Chroma have jointly launched Harness-1, a 20B-parameter retrieval sub-agent. It is trained via reinforcement learning in a stateful search framework that maintains a candidate pool, importance annotation sets, an evidence graph, and verification records, with a policy deciding when to search, filter, verify, and stop. Harness-1 achieves 0.730 average curated recall across 8 benchmarks, 11.4 percentage points above the next-best open-source sub-agent and trailing only Opus-4.6. Model weights and framework code have both been made public.
Key Takeaways
- UIUC and Chroma have jointly launched Harness-1,a 20B-parameter retrieval sub-agent.
- It is trained via reinforcement learning in a stateful search framework that maintains a candidate…
- Harness-1 achieves 0.
- 730 average curated recall across 8 benchmarks,11.
- 4 percentage points above the next-best open-source sub-agent and trailing only Opus-4.
Quick Facts
- Published
- Updated
- Source
- MarkTechPost(RSS)
- Category
- Model Releases
- AI score
- 73
UIUC和Chroma放出的这个20B检索子代理,用RL训练出了0.73的平均召回,把开源竞品甩开11.4分,只比Opus-4.6低一点。权重、Harness全开源,搞RAG的可以真刀真枪试试了。
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: Harness-1: a 20B retrieval sub-agent trained with reinforcement learning for stateful search. June 7, 2026. https://new.nat6.net/en/news/cmq3eppdj06j7sl97koo7u9xf/