publications
Selected publications and preprints.
2026
-
Where Did It Go Wrong? Process-Level Evaluation of Web Agents with Semantic State TrackingCOLM, 2026A benchmark that enables process-aware evaluation of web agents through automatic semantic state tracking, revealing exactly where and why each agent fails.Developed a controlled benchmark for diagnosing process-level failures in web agents, enabling fine-grained analysis of exploration, execution, and decision-making behaviors.
-
Towards Preference-Aligned 3D Quality AssessmentWorkshops on Image Processing and Image Understanding (IPIU), 2026A large-scale, MLLM-annotated dataset and lightweight evaluator for human-aligned 3D asset quality — accurate enough to steer iterative asset refinement.Preference-aligned 3D quality assessment for scalable, human-aligned evaluation and curation of 3D assets.