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.
Diagnosing VLM failures
I decompose visual questions into probes that test perception, grounding, and reasoning separately, each with exact ground truth, so a wrong answer can be traced to the skill that failed. I apply this to fine-grained domains such as charts and faces, where models often answer correctly for the wrong reasons.
Video understanding
I study temporal reasoning in video LLMs through repetition counting. In RepProbe, a linear probe on a model's final layer recovers the count better than the model's own answer, locating the failure at readout rather than perception.
Video generation and action
I am extending the same diagnostic lens to models that generate and act: steering video diffusion models to produce a requested number of repetitions, and benchmarking vision-language-action policies against world action models on long-horizon manipulation.
