STEC: Evidence Compression for Deep Search in Open-domain Multi-Hop QA

STEC is an evidence-compression framework for selecting final answers in open-domain multi-hop QA when search agents produce multiple, noisy trajectories. It first normalizes and groups trajectories by answer identity, then converts each group into a candidate-specific evidence representation. An evidence-guided verifier compares these representations and selects among existing candidates, shifting verification from raw trajectory comparison to aligned evidence comparison. According to the abstract, STEC performs best overall among compared methods on four open-domain multi-hop QA benchmarks, while ablations indicate that answer-level evidence compression contributes to the gains. The provided abstract does not report benchmark names or numerical results.