Harness-G: Graph-Structured Retrieval Framework Cuts Retrieval Aliasing, Beats Graph-R1 by 10.7 F1

Yanning Hou · hf · 2026-07-31

Harness-G is a graph-structured retrieval framework for RL search agents. It addresses retrieval aliasing (retrieval-equivalence collapse) by reformulating query generation as finite action selection: policy picks evidence sentence/entity or answer, environment builds menu and validates. This reduces aliasing and enables comparable alternatives. Structured Non-myopic Credit (SNC) uses a frozen scorer to assign downstream gains. On six QA benchmarks, Harness-G achieves highest average F1, outperforming Graph-R1 by 10.74 points at 1.5B and 3.98 at 3B.

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

Original post →

More from coding & agent

coding & agent channel →