March 12, MI Colloquium: Unveiling Reasoning Small Language Models - Debanjan Ghosh

Unveiling Reasoning in Small Language Models: Do They Truly Reason or Just Sound Like It?

What it’s about
How well do small language models actually “reason”? This talk shares findings from two experiments:
Commonsense reasoning: models select items from related lists and generate short passages that follow commonsense rules (physical, social, temporal, and mixed). Outputs are often fluent—but can be generic or illogical.
Physics-based reasoning: a second experiment tests how these models handle physical reasoning beyond everyday commonsense. 

About the speaker   
Debanjan Ghosh works at the intersection of Natural Language Processing and Machine Learning. He is currently a Principal Scientist at Analog Devices, developing small multimodal reasoning models. Previously, he led automated content generation at ETS, supporting learning and assessment products. He publishes at venues such as ACL and EMNLP, serves as an area chair for ACL Rolling Reviews, and is involved with SemEval Workshop 2026. He earned his PhD from School of Communication & Information at Rutgers University and completed postdoctoral research at MIT (Brain and Cognitive Sciences).