Benchmarking Dynamic Affective Reasoning: A Viewer-Centric Video Emotion Dataset

2026-07-11

AI Summary

This work addresses the limitation of traditional video emotion analysis, which typically treats emotions as static categories and overlooks their dynamic evolution across temporally continuous and causally linked events. To bridge this gap, we propose a dynamic emotion reasoning framework and introduce DAR—the first large-scale, viewer-centric video emotion dataset—comprising 15,087 videos and 36,908 event-aligned emotional segments. We define three core tasks: emotion segmentation, fine-grained classification, and causal reasoning. A novel annotation paradigm is introduced that emphasizes temporal density and explicit causal relationships. Furthermore, we design a two-stage model, DAR-R1, which integrates supervised fine-tuning with group-wise relative policy optimization, achieving new state-of-the-art performance in emotion localization and causal reasoning across more than ten multimodal large language models.

Abstract

Video emotion analysis is typically framed as a static classification problem, treating each clip as an independent labeled unit. However, such a formulation overlooks a key psychological fact: emotions change as a result of cumulative reactions to consecutive causal events. To bridge this gap, we introduce Dynamic Affective Reasoning, the first large-scale benchmark for viewer-centric affect transitions and causal reasoning over consecutive video events. DAR contains 15,087 videos and 36,908 event-aligned affective segments annotated with 27 emotion categories. Unlike existing video-based emotion datasets, DAR presents a new viewer-centric perspective on fine-grained emotional expressions and transitions, and provides dense, temporally grounded, and causally explicit reasoning chains. Based on DAR, we formally define three challenging tasks: affective segmentation, fine-grained emotion classification, and affective reasoning. Complementing this benchmark, we propose DAR-R1, a two-stage framework that combines supervised fine-tuning with Group Relative Policy Optimization. Experiments across 10+ MLLMs show that DAR-R1 sets a new state-of-the-art for dynamic affective reasoning, in terms of both emotional localization and affective reasoning. Project page: https://github.com/Zhang-Zhiyan/DAR.

Problem

• dynamic affective reasoning
• video emotion analysis
• affective transitions
• viewer-centric emotion
• causal reasoning

Innovation

• Dynamic Affective Reasoning
• Viewer-Centric Emotion
• Causal Emotion Transitions
• Temporal Emotion Segmentation
• Group Relative Policy Optimization