Ablation study on transferring sparse attention weights to dense attention across tasks.

Model	LongVideoBench	MLVU	TimeScope	LongTimeScope	Tomato	VSIBench
Qwen2.5-VL-7B	58.7	51.2	81.0	40.7	22.6	29.7
Dense-SFT	57.8 (-1.5%)	51.2 (+0.0%)	76.8 (-5.2%)	40.2 (-1.2%)	21.7 (-4.0%)	30.6 (+2.1%)
Dense-NSA	56.1 (-4.4%)	51.6 (+0.8%)	83.0 (+2.5%)	40.9 (+0.5%)	23.4 (+3.5%)	33.1 (+10.7%)
VideoNSA	59.4 (+1.1%)	51.8 (+1.2%)	82.7 (+2.1%)	44.4 (+9.1%)	26.2 (+15.9%)	36.1 (+20.3%)