Before It Fades: Reinforcing Temporal Representations at Inference Time in VideoLLMs

Open paper

Benchmark	Model	Rank	Results
video-question-answering-on-tempcompass-mcq	Qwen3-VL-8B + TAI	#4	Accuracy: 74.0
video-question-answering-on-mvbench	InternVL2.5-8B + TAI	#23	Avg.: 71.6