Investigation of temporal divergence vectors shows that VideoLLMs often fail to distinguish between forward and reversed video frames. The research identifies that temporal reasoning capacity peaks at intermediate layers before diminishing at the output.
HOW THIS AFFECTS YOU
●
researcherThis provides a framework for diagnosing where temporal reasoning breaks down in transformer architectures.