Bridge Evidence: Static Retrieval Utility Does Not Predict Causal Utility in Multi-Step Agentic Search
About one-third of the documents an agent used were “bridge” evidence: not useful to a static reader, but causally important to the search path.
The paper tests this by replaying 1,000 HotpotQA agent runs and deleting each read document to see what changed downstream. Across 23,322 document observations, static RAG utility and counterfactual trajectory utility were nearly independent, with Spearman rho at -0.026. The authors argue the useful bridge documents often work by supplying discriminative entities that shape the agent’s next query. ArXiv · AI/CL/LG's note
The paper tests this by replaying 1,000 HotpotQA agent runs and deleting each read document to see what changed downstream. Across 23,322 document observations, static RAG utility and counterfactual trajectory utility were nearly independent, with Spearman rho at -0.026. The authors argue the useful bridge documents often work by supplying discriminative entities that shape the agent’s next query. ArXiv · AI/CL/LG's note
score 5