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.
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.
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.
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.
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.
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.
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.
How does a vision-language model decide whether to trust an image or its own language prior when the two disagree? Using a semantic image pair protocol on CLIP-LLaVA models, activation patching, and logit lens, we showed that models follow whichever input they can discern more reliably, and located where in the network that decision forms. First paper at ECCV 2026 (eXCV); full-length extension under review at ICLR 2027.
Linear probes on hidden activations detect hallucination in medical LLMs at AUROC 0.77 to 0.86, yet steering the same neurons does not reliably change behavior. The project established this detection-control gap across four open-weight models and 36K free-generation answers. Presented at COLM 2026.
External validation of three published glioma subtype classifiers (3D CNN, ViT with ClinicalBERT, CNN with radiomics) on an IRB-approved cohort of 3,380 UCSF patients. An LLM pipeline labeled 9,377 pathology reports at F1 0.96 to 0.97, giving every subgroup enough samples to test. Errors concentrate by age: specificity for IDH-wildtype patients under 40 is 54%, against 80% at 55 and older.
A compute-matched comparison of mutation representations for cancer typing. We designed Bio MAF v4, 948 interpretable per-tumor features of mutation consequence, driver genes, hotspots, and allele fraction, rebuilt the MuAt mutation transformer from scratch, and benchmarked both on 8,800 TCGA tumors. The feature table outperformed the transformer for cancer typing (0.738 vs 0.528 balanced accuracy) and HRD prediction (AUROC 0.91 vs 0.75), and transferred to 913 PCAWG tumors.
Clinical ML at an NSF-funded startup since 2021: a multimodal depression detection model on interview audio and transcripts (F1 0.87 on DAIC-WOZ), imaging classifiers for lung and cardiac scans, and BERT-based clinical documentation tools, deployed on AWS to partner institutions.