Mahsa Khoshnoodi
I work on multimodal AI, studying how models see and reason about the world, and building the tools that reveal when that understanding is real.
I study where and why vision-language and video models fail. I build diagnostic benchmarks and interpretability probes that locate each failure at one of three levels: perception (did the model see it?), grounding (did it link what it saw to the question?), or reasoning (did it combine the evidence correctly?). For example, video LLMs internally encode repetition counts they then fail to report.
I study where multimodal models fail and why, across models that see, models that watch, and models that generate and act.
Diagnosing Vision-Language Models
I split each chart question into probes for perception, grounding, and reasoning, each with exact ground truth. Training on perception alone raised reasoning accuracy in an unseen domain from 13.6 to 80.1.
Temporal Reasoning in Video
Video LLMs encode how many times an action repeats but fail to report it. Across 42K clips, a linear probe on the final layer counts better than the model's own answer.
Video Generation and Embodied Agents
Ongoing
Steering video generation models to produce a requested number of repetitions, and comparing vision-language-action policies with world action models on long-horizon manipulation.
