arxiv:2603.14935

Video-CoE: Reinforcing Video Event Prediction via Chain of Events

Published on Mar 16

· Submitted by

xiaochonglinghu on Mar 19

AMAP-ML

Upvote

Authors:

Rui Chen ,

Abstract

A new Chain of Events paradigm is introduced for video event prediction that improves temporal modeling and logical reasoning in multimodal language models through structured event chains and enhanced training protocols.

AI-generated summary

Despite advances in the application of MLLMs for various video tasks, video event prediction (VEP) remains relatively underexplored. VEP requires the model to perform fine-grained temporal modeling of videos and establish logical relationships between videos and future events, which current MLLMs still struggle with. In this work, we first present a comprehensive evaluation of current leading MLLMs on the VEP task, revealing the reasons behind their inaccurate predictions, including lack of logical reasoning ability for future events prediction and insufficient utilization of visual information. To address these challenges, we propose Chain of Events (CoE) paradigm, which constructs temporal event chains to implicitly enforce MLLM focusing on the visual content and the logical connections between videos and future events, incentivizing model's reasoning capability with multiple training protocols. Experimental results on public benchmarks demonstrate that our method outperforms both leading open-source and commercial MLLMs, establishing a new state-of-the-art on the VEP task. Codes and models will be released soon.

View arXiv page View PDF Add to collection

Community

xiaochonglinghu

Paper submitter 1 day ago

An interesting work Video-CoE: Reinforcing Video Event Prediction via Chain of Events (CVPR26)

avahal

1 day ago

the big move here is replacing end-to-end option guessing with a constructed chain of events that ties observed video content to plausible futures. the two-stage setup—coe sft for reasoning and coe grpo for grounding—is neat in spirit, but i worry about how much the sft-generated chain biases the final prediction. an ablation where you drop coe sft or swap in a smaller, less biased chain generator would help quantify how much the chain quality actually drives vep gains. arxivlens breakdown helped me parse the method details; the walkthrough on arxivlens at https://arxivlens.com/PaperView/Details/video-coe-reinforcing-video-event-prediction-via-chain-of-events-843-ddd5f9c8 covers this nicely. overall it's a clean move toward grounded temporal reasoning, and i’m curious how it holds up when futures are truly open-ended and the chains must generalize to unseen event types.