Retrieval Grounding Outperforms Confidence Voting for Search Agents

_reachsumit · x · 2026-08-26

Research reveals that confidence-based voting performs poorly in multi-turn search agent settings due to "copy inflation": tokens copied from retrieved documents receive systematically inflated log probabilities, flattening confidence scores. The paper proposes Retrieval-Grounded Voting (RGV), which scores rollouts by lexical overlap between the final answer and retrieved documents, bypassing the contaminated context. Experiments show RGV consistently outperforms confidence-based voting across four benchmarks.

Original post →

More from Research

Research channel →