Harness-G: A Graph-Structured Harness for Search Agents
The paper says search-agent training is losing useful contrast because different generated queries end up retrieving the same evidence.
Harness-G replaces free-form query generation with finite choices over evidence sentences, entities, or answering. The environment builds the action menu, tracks retrieval state, and validates each choice. Its Structured Non-myopic Credit method scores selected actions against alternatives and credits earlier actions that made later gains possible. Across six QA benchmarks, the authors report the best average F1 at both tested model sizes, beating Graph-R1 by 10.74 points at 1.5B and 3.98 points at 3B. HF Daily Papers' note
Harness-G replaces free-form query generation with finite choices over evidence sentences, entities, or answering. The environment builds the action menu, tracks retrieval state, and validates each choice. Its Structured Non-myopic Credit method scores selected actions against alternatives and credits earlier actions that made later gains possible. Across six QA benchmarks, the authors report the best average F1 at both tested model sizes, beating Graph-R1 by 10.74 points at 1.5B and 3.98 points at 3B. HF Daily Papers' note
score 4