Yue Chen陈越

Structural Robustness Auditing for Document Parsing

• Led the development of ProSA, an auditing framework requiring no access to parser internals. To address Footprint Bias in measuring parsing damage by perturbation area, designed the block-level structural loss rate B-SLR to jointly assess block locations and text preservation.
• Led the research direction, experimental design and core validation; evaluated MinerU and PP-StructureV3 and analyzed how structural errors propagate to downstream question answering and retrieval.
• Across 1,000 document pages and 29 fixed perturbation configurations, B-SLR achieved fitted R² values of 0.727 / 0.916 against changes in OCR character error rate, exceeding 0.384 / 0.110 for the area-based metric.

Sprint Video Analysis and Evidence-Augmented Reports

• Contributed to the SEE–ASSESS–EXPLAIN framework, combining pose estimation from single-view video, phase-specific rule assessment, literature retrieval and LLM explanations to generate technique analysis reports with cited evidence.
• Contributed to expert agreement, retrieval and evidence comparison evaluations. Across 1,008 fixed phase-specific rule assessments from 112 sprint trials, rule judgments agreed with the majority judgment of three experts in 93.2% of cases.

Multimodal Physical State Estimation for Robots

• Fuse RGB-derived poses, robot states and interaction history using an MLP + GRU to estimate object mass, friction and center of mass, building a physical state representation continuously updated through interaction.
• Conduct training under multiple conditions, temporal ablations and generalization evaluations. In completed in-domain tests, full history reduced friction estimation MAE by approximately 65% relative to a short-history baseline.