My work asks whether a model's apparent competence reflects real understanding of the structure in its data, or a correlation that happened to hold where it was tested. I study this through robustness, multimodal reasoning benchmarks, and physics-grounded generative models.
When inputs shift in several ways at once, what breaks, and can a model learn to withstand all of it together?
Are vision-language models reading the image, or leaning on priors and guesses?
How do we make generative models respect the physics of the instrument, instead of hallucinating detail?
Microscopes see far less detail along the depth axis. Can a generative model fill it in without making things up? MicroFM trains on data simulated with realistic optics and starts its flow from the actual observation rather than noise, which keeps reconstructions faithful and makes sampling fast.
Can vision-language models solve the puzzles, series and analogies used in aptitude exams? NTSEBench collects such problems from India's NTSE exam and finds models well behind people, often guessing rather than reasoning through the problem.
Robustness is usually tested one perturbation at a time, but real inputs are messy in several ways at once. We study how language models behave under multiple simultaneous perturbations, and show a mixed-training strategy can recover robustness to all of them without hurting clean accuracy.
Do models actually follow a flowchart? FlowVQA asks questions that require tracing logic through diagrams. Flipping a flowchart upside down noticeably hurts accuracy, suggesting models lean on a learned reading direction rather than the diagram itself.
Can an LLM turn a natural-language hardware spec into correct SystemVerilog assertions? ChIRAAG structures the spec, generates assertions, and repairs them using simulation feedback.
* and † denote equal contribution. For a complete, up-to-date list see Google Scholar.
Vatsal Gupta