Apollo: Scaling Consistency and FPS Sampling Unlock Efficient Video Understanding in Large Multimodal Models
The Apollo paper presents a systematic study of what actually drives video understanding in Large Multimodal Models (LMMs), an area previously dominated by unjustified design choices due to high computational costs. A key discovery — "Scaling Consistency" — shows that design and training decisions made on smaller models and datasets reliably transfer to larger ones, dramatically reducing research overhead. Leveraging these insights, the authors identify FPS-based sampling and optimal vision encoder selection as critical factors, producing the Apollo model family that achieves state-of-the-art results across multiple video benchmarks at both 3B and 7B parameter scales.