Harness-G: Reformulating Search Agent Retrieval as Finite Action Selection on Graphs

_reachsumit · x · 2026-07-31

Harness-G is a graph-structured framework designed for reinforcement learning (RL) search agents. The paper identifies a phenomenon called retrieval-equivalence collapse during the training of models like Search-R1: agents continuously generate distinct query strings for the same question, but their accumulated evidence sets increasingly overlap, leaving trajectories with little effective retrieval contrast.

To address this, Harness-G redesigns the policy-environment interface by reformulating free-form query generation into finite action selection over a graph. The policy selects an evidence sentence, an entity, or chooses to answer, while the environment constructs the action menu, tracks retrieval state, and validates execution. The framework also introduces Structured Non-myopic Credit (SNC), which uses a frozen answer scorer to compare selected actions against their alternatives, optimizing the agent's multi-step retrieval decisions.

Related event: Harness-G Tackles Retrieval Aliasing in Search Agents(2 posts)→

Original post →

More from coding & agent

coding & agent channel →