ResearchPod Summary
In open-domain multi-hop question answering (QA), models often generate multiple search trajectories to gather information. However, these trajectories are frequently redundant, conflicting, or noisy, making it difficult to select the correct final answer. The authors investigate how to improve this final selection stage by moving away from comparing raw, unstructured search logs toward a more systematic, evidence-based approach.
The authors propose STEC (Evidence Compression for Deep Search), a framework that processes search outputs in three distinct stages:
Experimental results across four benchmarks (HotpotQA, 2WikiMultiHopQA, MuSiQue, and Bamboogle) demonstrate that STEC consistently outperforms representative baselines. By organizing evidence at the candidate level, the system effectively filters out noise and redundant information that typically plagues multi-trajectory search. Ablation studies confirm that the compression mechanism is a critical contributor to the performance gains, as it provides the verifier with a cleaner, more comparable signal for each candidate.
This work addresses a common bottleneck in agentic search systems: the "final selection" problem. As LLMs become more capable of performing complex, multi-step research, they often produce a deluge of information. STEC provides a scalable way to synthesize this information, ensuring that the final answer is grounded in the most relevant evidence rather than being swayed by the sheer volume of noisy or conflicting search results.
AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.