Vision-Language Models Trust What They Can Discern
When an image and the text prompt disagree, CLIP-LLaVA models follow the input they can discern more reliably. Activation patching and logit lens locate where that decision is made.
Publications
Also on Google Scholar.
Peer-Reviewed Publications
Readable but Not Controllable: Neuron-Level Evidence for Medical LLM Hallucination
Linear probes on hidden activations detect hallucination in medical LLMs at AUROC 0.77 to 0.86. Steering the neurons most associated with hallucination does not reliably change behavior. Detection and control are separate properties.
Rethinking Medical LLM Hallucinations: A System-Level Survey
A survey of hallucination in medical LLMs covering definitions, benchmarks, detection, and mitigation. It argues that hallucination is a structural property of probabilistic generation and should be treated as a system-level risk management problem.
Under Review and In Preparation
Believing Is Seeing: How LLaVA Models Weigh a Contradicting Image
Full-length extension of the ECCV paper: three CLIP-LLaVA models, four sources of image-text conflict, a real-photograph evaluation on NaturalBench, and multiple-testing correction across all reported effects.
Potential for Age-Based Bias in AI-Informed Neurosurgical Decision-Making: A Sociodemographic Subgroup Assessment Approach for Open-Source Glioma Molecular Subtype Classifiers
External validation of three published glioma subtype classifiers on 3,380 UCSF patients, with subgroup tests powered by an LLM labeling pipeline over 9,377 pathology reports. Errors concentrate by age: specificity 54% for IDH-wildtype patients under 40 versus 80% at 55 and older, a difference AUROC does not show.
Benchmarking Somatic Mutation Representations for Cancer Typing and HRD Prediction
We introduce Bio MAF v4, a 948-feature interpretable representation of a tumor’s somatic mutations, and benchmark it against mutational signatures and a MuAt transformer trained from scratch on 8,800 TCGA tumors across 20 cancer types. The feature table outperforms the transformer (0.738 vs 0.528 balanced accuracy) and transfers to 913 PCAWG tumors.
A Recipe for Multimodal Multi-Instance Depression Detection: Frozen Teachers, Schedule-Aware Training, and Audio Contribution Diagnostics
A frozen-teacher multimodal multi-instance model for depression detection on DAIC-WOZ (DepRoBERTa text, emotion-finetuned wav2vec2 audio, gated-attention pooling), plus the audio occlusion delta, a diagnostic for whether the audio branch is contributing at all.