Which Way Did It Move? Diagnosing and Overcoming Directional Motion Blindness in Video-LLMs

Jongseo Lee, Hyuntak Lee, Sunghun Kim, Sooa Kim, Jihoon Chung, Jinwoo Choi, [0.3em] Kyung Hee University · 2026

Read Which Way Did It Move? Diagnosing and Overcoming Directional Motion Blindness in Video-LLMs by Jongseo Lee, Hyuntak Lee, Sunghun Kim, Sooa Kim, Jihoon Chung, Jinwoo Choi, [0.3em] Kyung Hee University (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-22823.txt.

Introduction

Lee and five colleagues diagnose directional motion blindness in video language models, the failure to tell whether a lone object moved left, right, up, or down. Most models scored near chance, and tracing the signal showed direction stayed readable inside the network but failed to bind to the correct spoken answer. Their DeltaDirect objective, which predicts motion vectors from adjacent-frame feature changes, lifted synthetic accuracy from about 26 percent to about 86 percent.

Catalog record

CREATORJongseo Lee, Hyuntak Lee, Sunghun Kim, Sooa Kim, Jihoon Chung, Jinwoo Choi, [0.3em] Kyung Hee University
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE169245
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.22823

Browse the full library · About the Tower of Babel