Harness-G: A Graph-Structured Harness for Search Agents

Harness-G proposes a graph-structured retrieval interface that replaces free-form query writing with finite action selection.

The agent chooses among evidence sentences, entities, and answer actions, which helps reduce retrieval aliasing.

It also adds Structured Non-myopic Credit to better reward early retrieval decisions that help later outcomes.

Across six question-answering benchmarks, the approach outperforms prior baselines with stronger average F1.