This repository contains the weights for the model presented in the paper Learning Spatiotemporal Sensitivity in Video LLMs via Counterfactual Reinforcement Learning.

Video large language models (Video LLMs) often answer video questions through shortcuts such as single-frame cues and language priors rather than by tracking spatiotemporal dynamics. Counterfactual Relational Policy Optimization (CRPO) is a dual-branch RL framework designed to improve spatiotemporal sensitivity.

CRPO constructs counterfactual videos through horizontal flips and temporal reversals, training on both original and counterfactual branches. It introduces a Counterfactual Relation Reward (CRR) between their answers, which encourages answers to change for dynamic questions and remain unchanged for static questions. This makes it difficult for shortcut policies to be consistently rewarded.

The model was evaluated on DyBench, a paired counterfactual video benchmark with 3,014 videos covering reversible dynamics, moving direction, and event sequence. On Qwen3-VL-8B, CRPO significantly improves spatiotemporal sensitivity metrics (such as DyBench P-Acc and TimeBlind I-Acc) over the base model while maintaining competitive general video performance.