arXiv subjects

Mengzhu Xu

Publications and source records attributed to Mengzhu Xu.

3 recordsLinked to original sources

arXiv2026

Right Diagnoses, Decorative Reasoning:A Perturbation Audit of Medical Chain-of-Thought

Clinicians read chain-of-thought (CoT) rationales as evidence of medical reasoning, but whether the visible chain plays that role is rarely tested. General-domain CoT-faithfulness probes ignore clinical cost, and medical LLM evaluations treat the chain as a black box. We close this gap with a medical perturbation audit: a 30-operator battery edits both the chain and the question with clinically motivated operators (severity reversal, negation flip, demographic swap, evidence ablation), paired with a chain-update times answer-flip joint analysis that classifies each model by its failure mode. Applied to 14 LLMs on four medical QA benchmarks, three independent tests converge: the Chain-Decoupling Rate (CDR; chain does not register the edit and the answer does not flip) is 72.9% panel-wide on clinically meaningful destructive edits, chain corruption leaves accuracy unchanged, and removing CoT prompting does not reduce accuracy. Two board-certified clinicians re-annotate N=197 perturbed questions; 98.5% leave the gold defensible. The pattern holds across medical and reasoning fine-tuning and scale; on the closed-source tier, where the chain text is unavailable, the answer-side signals are consistent with the same decoupling. Our framework and CDR provide a reusable yardstick for auditing whether medical CoT is faithful or merely documentation.

cs.AI↗

arXiv2026

Toward Real-Time Circadian Phase Estimation with Low Latency from Wearable Sensing Data

Accurate estimation of the human circadian phase plays an important role in personalized health monitoring, but most existing wearable-based approaches operate retrospectively and require full circadian cycle recordings, leading to high estimation latency and substantial data and computational burden for real-time deployment on edge devices. In this study, we investigated whether circadian phase can be estimated in real time using only short historical windows of wearable data. We propose a low latency framework that estimates instantaneous circadian phase from past observations, with a cosinor-fitted core body temperature rhythm serving as the reference. Data from a free-living field study involving 14 participants were used to systematically evaluate the effects of sensor modality selection, historical window length, and model class under participant-based cross-validation. The results showed that estimation accuracy improves with increasing window length but saturates at approximately 8 hours of history. Tree-based models reached a performance plateau beyond 480 minutes, whereas sequence-based models continued to benefit from longer temporal contexts. When relying solely on light exposure and physical activity, the proposed approach achieved a mean circular mean absolute error (CMAE) of 1.19 h. These findings provide practical guidance for efficient and deployable real-time circadian phase monitoring using wearables.

eess.SP↗

arXiv2025

Affordance-First Decomposition for Continual Learning in Video-Language Understanding

Continual learning for video--language understanding is increasingly important as models face non-stationary data, domains, and query styles, yet prevailing solutions blur what should stay stable versus what should adapt, rely on static routing/capacity, or require replaying past videos. We aim to explicitly specify where stability lives and where plasticity should be focused under realistic memory and privacy constraints. We introduce Affordance-First Decomposition (AFD): videos are mapped to slowly varying affordance tokens that form a shared, time-aligned substrate, while a lightweight, query-routed, conflict-aware scheduler concentrates adaptation and grows capacity only when needed. The substrate is stabilized via weak alignment and teacher consistency, and training uses question-only replay. AFD achieves state-of-the-art across protocols: 51.6% average accuracy with -1.8% forgetting on domain-incremental VideoQA, ViLCo R@1@0.5 of 29.6% (MQ) and 20.7% (NLQ) with 18.4% stAP@0.25 (VQ), and 39.5% accuracy with -1.6% forgetting on time-incremental iVQA. Overall, AFD offers an explicit, interpretable split between a stable interaction-centered substrate and targeted adaptation.

cs.CV↗