VideoJev scores video decisions directly and reuses context

When a video system is asked several structured questions about one clip, it may need to locate an event or compare answer choices each time. It remains unclear whether those decisions can be made without generating text for every answer, while also avoiding repeated work on the same visual context.

VideoJev scores candidate time intervals or answer choices with one shared model component and reuses video context across related questions.

Researchers compared VideoJev-4B and VideoJev-35B-A3B with their corresponding Qwen models evaluated in instruction mode.

Across three event-location tests, mean interval overlap scores were higher by 6.7 and 10.8 points; across four multiple-choice video-question tests, mean scores were higher by 8.0 and 5.7 points.

VideoJev had been trained to score intervals or choices directly, while the Qwen models generated outputs. The results show higher task-family averages for the direct-scoring models, though the comparison also differs in training phase and readout.

The backbone comparisons do not show that the readout alone caused the gains: the paper does not specify whether each model and its corresponding Qwen comparator used identical checkpoints. On the evaluated video-question tests, the models’ probabilities also tended to overstate correctness.