Which Way Did It Move? Diagnosing and Overcoming Directional Motion Blindness in Video-LLMs
The research diagnosed directional motion blindness in Video-LLMs, where motion direction information is accessible but fails to bind correctly to answer options, especially in new scenarios. The solution introduced DeltaDirect, a training objective that strengthens signed displacement cues at the projector level. This intervention significantly improved performance on complex out-of-domain tasks by boosting the signal strength of motion direction.