SearchAtlas is a framework that converts raw search trajectories produced by LLM agents into structured evidence graphs showing how evidence propagates through multi-step reasoning, using an automated parsing pipeline that reaches an 86.0% mean edge F1 score against human annotation. Applying it across five search agents and three benchmarks, the authors find systematic differences in search strategy and show that process-level failures, such as fragmented answer support and unverified knowledge chains, correlate more strongly with wrong answers than final-answer accuracy alone. The tool is offered as a way to diagnose and debug agentic search/RAG pipelines beyond simple accuracy metrics.
