LMs Accept False Premises in Real-World Communication
ACL 2026.
* Equal contribution.
About
About Sraavya Sambara
Hi! I'm an incoming PhD student at MIT CSAIL and LIDS, and I'm fortunate to be supported by the NSF Graduate Research Fellowship. My research centers on building efficient and reliable multimodal machine learning systems. I have previously worked on making vision-language and language models safer with conformal prediction, evaluating their multimodal reasoning capabilities, and investigating LM alignment failures in real user interactions.
Previously, I studied Computer Science at Harvard University, where I graduated Phi Beta Kappa and was named a Harvard Student Leader in AI. I have worked with Professor Pranav Rajpurkar at Harvard, Professor Irene Chen at BAIR, and Professor Monica Agrawal at Duke NLP.
Beyond my academic work, I have deployed production AI systems to F50 companies at Distyl AI, helped develop the first commercially available multi-finding abdomen-pelvis CT triage AI in the US at a2z Radiology AI, and contributed to multimodal AI research with Eric Schmidt at Schmidt Sciences.
Outside research, I enjoy tennis, reading 20th century British fiction, and Carnatic music. Please reach out at sraavyas@mit.edu if you'd like to chat — I'd love to hear from you!
Selected Research
Publications and preprints-
-
Mitigating Hallucinations in Vision–Language Models via Conformal Prediction
ML4H @ NeurIPS 2024.
* Equal contribution.
-
Improving Multimodal Reasoning in 3D Vision–Language Models
Pacific Symposium on Biocomputing, 2026. Oral presentation. Invited Talk at Microsoft Research.
* Equal contribution.
-
Reasoning Models Don't Necessarily Help: Benchmarking LLMs for Emergency Department Revisit Prediction
Pacific Symposium on Biocomputing, 2026. Oral presentation.