Harness-1: Reinforcement Learning for Search Agents with State-Externalizing Harnesses
What this work does
Externalizes candidate pools, document memory, evidence graphs and curation state while training the policy to make semantic search and selection decisions.
Connection to discovery
A concrete division between reliable harness bookkeeping and trainable evidence-seeking decisions.
Evidence
Retrieval benchmarks separate final curated recall from trajectory recall, and same-checkpoint ablations disable individual state-management components.
Scope and limitations
Retrieval relevance and recall do not verify a new scientific claim. Harness ablations at inference measure dependence on the trained interface, not an isolated universal contribution to capability.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.