Video event prediction asks models to infer what happens next from partial footage. Most multimodal LLMs translate what they see into text before reasoning, but the authors of a new paper argue that this verbalization discards fine-grained motion, geometry, and interaction cues. Their proposed framework, Future-L1, instead lets the model alternate between ordinary language tokens and continuous latent visual spans during decoding, effectively "imagining" intermediate visual states before committing to a prediction.

The team built a training set called Future-L1-50K by selecting examples where future visual hints help prediction and aligning latent states to embeddings of the actual future frames. They then optimized sampled latent trajectories with LA-DAPO, a reinforcement-learning objective that rewards outcome contrast and temporal diversity. This two-stage approach teaches the model not only to generate useful latent representations but also to explore varied plausible futures.

On FutureBench, Future-L1 improves Qwen3-VL-8B from 61.0 to 85.4, surpassing the previous best model, Video-CoE, by 10.4 points. On TwiFF-Bench, the average score rises from 2.44 to 3.04. The results suggest that preserving intermediate visual semantics in latent space—rather than translating every reasoning step into text—can reduce visually ungrounded hallucinations and improve future-oriented video understanding. The paper has been accepted to EMNLP 2026.