When Can Models Trust Their Training Data?
Diagnosing when and why multimodal models generalize, and turning the diagnosis into fixes.
Problem. Models learn whatever correlations the training distribution offers, including shortcuts that fail under distribution shift. Knowing when a learned correlation can be trusted is the difference between a demo and a deployable system.
Findings across three works:
- LCA-on-the-Line (ICML 2024, Oral): class-taxonomy distance (LCA) predicts out-of-distribution generalization across model families, turning label hierarchies into a practical model-selection signal. Project page
- Natural experiments in real-world data (arXiv 2026): real-world datasets contain naturally occurring interventions; treating them as such, via Markov-blanket-based causal feature selection, improves downstream performance over purely observational treatment.
- The question-first paradox (arXiv 2026): logit-lens and attention probes show that placing a question before the image guides perception but starves answer generation of the question. Prompt echoing, restating the question on both sides of the image, fixes it training-free, up to +19 accuracy points across NaturalBench, POPE, Winoground, and VQAv2.
Why it matters. Mechanistic diagnosis turns “the model fails sometimes” into a specific, fixable cause, and each diagnosis above shipped with its fix.
See the publications page for the papers.