When Web Agents Finish but Still Fail: Reproducible Triggers and Trace Diagnostics for Parallel Web Exploration

The paper introduces Parallel WebBench, a benchmark for evaluating long-horizon web agents, comprising 1,679 verified records to analyze failure modes in web exploration. The authors train WebExplorer-style agents using GRPO, achieving a significant improvement in completion rates from 50.7% to 96.0% and an F1 score increase from 0.2489 to 0.4529 at 16k context and 16 interaction rounds. Despite these advancements, the study highlights persistent issues with context-bound search loops, premature termination, and synthesis collapse, indicating a need for enhanced evidence-grounded coverage and diagnostics to address completion-correctness gaps in AI systems.