Researchers have proposed VepAgent, an agentic framework designed to improve Video Event Prediction (VEP) by combining causal-transition reasoning with tool-augmented reinforcement learning. The work addresses a limitation in current Multimodal Large Language Models (MLLMs), which the authors say rely on retrospective summarization and text-centric priors, hindering their ability to bridge unobserved causal transitions.
Rather than passively projecting future trajectories from historical dependencies, VepAgent explicitly models the logical progression from terminal observed states to future events. To support this, the authors constructed futurebench-4K, a chain-of-thought dataset for supervised fine-tuning that structures the deduction of unobserved intermediate states.
The framework also includes a diagnostic tool library for state tracking, frame retrieval, and region magnification, allowing the agent to dynamically augment its reasoning and recover missing spatio-temporal evidence during inference. A composite reward mechanism jointly optimizes prediction accuracy, causal coherence, and reliable priors, pushing the agent toward genuine visual grounding instead of superficial textual similarity.
According to the paper, extensive evaluations on the FutureBench and NEPBench datasets show the method achieves state-of-the-art performance, outperforming larger MLLMs and validating the agentic, future-oriented reasoning paradigm.