Vijay Vankadaru

I am a master’s student in data science at UC Berkeley. My research is on mechanistic interpretability of language and vision-language models: when a model has the right information internally, what determines whether its output uses it. I am applying to PhD programs this cycle for a Fall 2027 start.

Two projects make up the interpretability work. With Dr. Vasha DuTell, I study how vision-language models weigh an image against text that contradicts it, using activation patching and logit lens to locate where that decision is made (ECCV 2026; extension under review at ICLR 2027). With Prof. Tanya Roosta, I led a neuron-level study of hallucination in medical LLMs: linear probes on hidden activations detect hallucination reliably, but steering the same neurons does not suppress it (COLM 2026).

The medical work is both empirical and representational. At the UCSF M2AI Lab, I externally validated three published glioma MRI classifiers on 3,380 patients and found age-dependent errors that AUROC does not show, using an LLM pipeline that labeled 9,377 pathology reports to make the subgroup tests possible. At the University of Maryland School of Medicine, I designed Bio MAF v4, an interpretable representation of a tumor’s somatic mutations, and showed in a compute-matched benchmark that it outperforms a mutation transformer trained from scratch for cancer typing. At DASION, a clinical AI startup, I have built and deployed ML systems since 2021, including a multimodal depression detection model that is now a solo-author paper under review.

Selected papers

• ECCV 2026eXCV · first author
  Vision-Language Models Trust What They Can DiscernModels follow the input they can discern more reliably. Activation patching locates where the decision is made.
• COLM 2026DAIH · first author
  Readable but Not ControllableProbes detect hallucination in medical LLMs at AUROC 0.77 to 0.86; steering the same neurons does not suppress it.
• ICLR 2027under review
  Believing Is SeeingThree LLaVA models, four conflict types, and an evaluation on real photographs.
• AI in Medicinein press
  Rethinking Medical LLM HallucinationsA system-level survey of hallucination in medical LLMs: definitions, benchmarks, detection, and mitigation.

All publications  /  Research  /  CV

Research interests

• Mechanistic interpretabilityActivation patching, probing, and steering, and when detection and control diverge.
• Trust and cross-modal conflictHow models weigh input evidence against their priors, and whether that decision can be read from activations before generation.
• Clinical MLEvaluation that matches deployment conditions, subgroup analysis, interpretable representations, and detection methods that do not rely on an LLM judge.

News

• Oct 2026Presented Readable but Not Controllable at the DAIH workshop, COLM 2026, San Francisco.
• Sep 2026Believing Is Seeing, the full-length extension of the ECCV paper, submitted to ICLR 2027.
• Aug 2026Vision-Language Models Trust What They Can Discern accepted to the eXCV workshop at ECCV 2026.
• Jul 2026Readable but Not Controllable accepted at COLM 2026; preprint on arXiv.
• Jun 2026Joined the UCSF M2AI Lab (Dr. Madhumita Sushil) and the Maron group at UMSOM (Dr. Bradley Maron) through The Tensor Lab.
• 2026Rethinking Medical LLM Hallucinations: A System-Level Survey accepted at Artificial Intelligence in Medicine.
• 2026Reviewer for InterpScience (NeurIPS 2026), Actionable Interpretability (COLM 2026), and eXCV (ECCV 2026).
• Apr 2026Began working with Dr. Vasha DuTell on modality arbitration in vision-language models.
• Jan 2026Began working with Prof. Tanya Roosta on hallucination in medical LLMs.