extract evidence links

Existing Generative Engine Optimization (GEO) approaches are confined to single-page optimization and overlook the dynamic process by which LLM agents gather and integrate evidence across multi-step browsing trajectories in web environments. This work proposes EcoGEO, a novel framework that models GEO from an ecosystem perspective, introducing TRACE—a trajectory-aware coordinated evidence system. TRACE jointly designs entry and supporting pages through shared terminology, internal linking, and consistent product attributes to guide agent search paths. By moving beyond traditional page-level optimization, EcoGEO significantly improves target product recommendation rates on OPR-Bench and enhances key trajectory-level metrics, including target-relevant crawling, follow-up searches, and internal link clicks.

Current AI web agents rely on passive inspection of execution logs for multi-step task verification, which often leads to information overload and low interpretability. This work proposes an interactive verification approach that identifies critical evidence pages from agent trajectories and reconstructs lightweight snapshots preserving contextual states—such as filtering criteria, search queries, and scroll positions—to enable users to actively trace decision-making through a navigable visual interface, with explicit alerts when provenance is unavailable. The method achieves the first trajectory summarization technique that retains contextual fidelity while supporting verifiable traceability, attaining 83.7% precision and 88.8% recall on AssistantBench and Online-Mind2Web, along with a 61.6% reduction in trajectory size. User studies demonstrate significant reductions in verification time and cognitive load, alongside improved usability and error detection capability.

Existing search agents are constrained by transient context or single-page browsing mechanisms in multi-turn complex tasks, hindering effective reuse of previously acquired information and leading to redundant retrieval and rendering. This work proposes the Fetch-then-Explore framework, which decouples page selection from evidence extraction for the first time and introduces a file system–based persistent workspace, enabling agents to revisit and extract content from previously visited pages on demand throughout the task. Built upon a unified ReAct architecture, the approach integrates three backbone models and is evaluated on the BrowseComp and WideSearch benchmarks. Experimental results demonstrate state-of-the-art accuracy on BrowseComp and consistently competitive or superior performance on WideSearch, with gains primarily attributed to efficient revisitation and secondary utilization of already accessed pages.

Existing datasets of scientific ideation trajectories struggle to comprehensively capture the full research process—from literature exploration and tool utilization to the evolution of intermediate artifacts and final proposals. This work proposes a reverse-to-forward synthesis mechanism that emulates the uncertainty, evidence integration, and phased convergence characteristic of real scientific inquiry through a Generator–Advisor architecture. By leveraging action–observation–editing sequence modeling, context-aware verification, and process-level supervision, the approach generates multi-turn trajectories aligned with authentic research practices, starting from high-quality papers and proposals. The study yields the first trajectory dataset spanning the complete scientific workflow and establishes a generalizable paradigm for synthesizing process-supervised data for scientific agents.