Sep 14, 2026 · cs.LGJ/K move · Enter open · S save
Mantek Singh, Jeshwanth Challagundla, Siddharth Raina, Jasmin Jarsania
Liverpool John Moores University, Liverpool, England · Carnegie Mellon University, Pittsburgh, USA · Meta, Sunnyvale, USA · University of Texas at Arlington, Texas, USA
We present an efficient method to distill reasoning capabilities into compact video-language models (VLMs) for video question answering (VideoQA). Our approach fine-tunes a 2B-parameter model using only
∼900 uncertainty-selected examples, each augmented with synthetic chain-of-thought (CoT) rationales generated by a 4B teacher. Despite its minimal compute cost - under two hours on a single A100 GPU - our method enables the 2B model to outperform VLMs up to 4
× larger, and generalize across CinePile, ActivityNet-QA, and MLVU, approaching the performance of its own 4B teacher. A key finding is that placing CoT rationales after the answer - contrary to standard prompting - substantially improves reasoning in compact models. This insight challenges prevailing CoT conventions and reveals new alignment strategies under limited model capacity. Our findings offer a practical blueprint for training deployable, reasoning-rich VLMs suited for mobile and edge applications.