Beyond Confidence: Test-Time Scaling for Multi-Turn Search Agents via Retrieval Grounding

The paper introduces Retrieval-Grounded Voting (RGV), a new test-time scaling method for multi-turn search agents that retrieves and conditions on external documents. It identifies that confidence-based voting fails in this setting due to copy inflation, where tokens copied from retrieved documents inflate log probabilities and flatten confidence scores. RGV scores each rollout by lexical overlap between the final answer and retrieved documents, avoiding contaminated context and additional LLM calls, and achieves consistent accuracy gains across multiple benchmarks and models.

Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv AI.

• llms
• agents
• benchmarks