Harness-1 addresses a core frustration for anyone building retrieval‑augmented systems: the model is forced to learn both what to search and how to remember everything it sees. When a single policy handles search decisions, evidence tracking, deduplication, and stopping criteria, reinforcement learning gets tangled in low‑level bookkeeping, slowing convergence and limiting generalization. Researchers from UIUC, UC Berkeley, and Chroma show that this split hurts performance, especially on tasks far from the training data.
Their solution is Harness‑1, a 20‑parameter‑billion retrieval subagent built on gpt‑oss‑20b that delegates all state maintenance to an external harness. The policy focuses purely on semantic actions—choosing queries, selecting documents, verifying claims, and deciding when to stop. The harness holds a working memory that stores a compressed candidate pool, an importance‑tagged curated set (capped at 30 documents), an evidence graph built from simple regex extraction, a full‑text store, and verification caches. By offloading these routine structures, the agent can learn richer search strategies without being overwhelmed by memory management.
Training follows the same division: supervised fine‑tuning teaches the model to operate the harness interface, while reinforcement learning improves search decisions over the preserved state. Using on‑policy CISPO with a tool‑diversity bonus, Harness‑1 achieves an average curated recall of 0.730 across eight diverse benchmarks—11.4 points above the next best open subagent and close only to the frontier Opus‑4.6 model. Gains are largest on held‑out tasks (+17.0 points), indicating that the learned search operations transfer well.
For developers, this means a ready‑to‑use, open‑weight retrieval module that can be served with vLLM, SGLang, or Transformers, plugged into any downstream answer generator. It reduces the engineering burden of building custom search state machines while delivering stronger, more generalizable evidence collection. Harness‑1 proves that separating policy from harness is a practical path to better retrieval agents without scaling up model size unnecessarily.
#AI #MachineLearning #Retrieval #ReinforcementLearning #OpenSource #RAG