Temporal Conflict Resolution in Video Large Language Models

Video large language models (video-LLMs) process dynamic video in which objects change color, shape, count, and identity over time, yet current temporal video benchmarks evaluate them by assuming a single consistent interpretation per clip. The knowledge-conflict literature isolates contradictory evidence, but it confines that scope to text or static images, leaving uncharacterized how a video-LLM arbitrates contradictory visual evidence within a single continuous stream. We introduce TCR-Bench, a framework that casts temporal conflict resolution as an explicit evaluation target. The pipeline extracts target objects from consistent video clips using promptable video segmentation, and then injects a visually plausible contradictory state into a controlled temporal window to produce matched control and conflict sequences. Because the benchmark manipulates the conflict type and its temporal duration independently, it isolates a model's ability to identify specific states from its capacity to detect logical contradictions. The evaluation suite classifies each response as recency-, frequency-, multi-state-, or abstention-consistent, and it quantifies change-localization error while modeling the logistic pull of recency against frequency. Evaluating seven leading video-LLMs exposes a vulnerability to content-level visual prompt injection: rather than flagging the temporal conflict, models default to the injected reality. As the edit dose grows to the full temporal window, the rate at which models adopt the injected state rises from 7.4% to 27.8%, while explicit conflict detection plateaus at 25.0% and roughly 47% of answers remain correct, so the failure is a graded erosion rather than a wholesale collapse. TCR-Bench provides the instrumentation needed to diagnose how visual evidence is integrated across time, a prerequisite for safe deployment in dynamic environments.