Benchmarking Dynamic Affective Reasoning Viewer-Centric Video Emotion Dataset DAR

Jul 16, 2026

Video emotion analysis is commonly treated as static clip-level classification, ignoring how emotions evolve through cumulative reactions to consecutive causal events. DAR (Dynamic Affective Reasoning) introduces a large-scale viewer-centric benchmark for affect transitions and causal reasoning across temporal video events. DAR includes 15,087 videos with 36,908 event-aligned affective segments annotated across 27 emotion categories, offering dense, temporally grounded, causally explicit reasoning chains. Using DAR, three tasks are defined: affective segmentation, fine-grained emotion classification, and affective reasoning, and a two-stage DAR-R1 model is proposed and evaluated on 10+ MLLMs.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages25

Reading Time1 h 3 min

Abstract

Introduction

Dynamic Affective Reasoning Tasks

Benchmark Contributions and Annotation Design

Example Tasks and Outputs

Affective Segmentation

Fine-grained Emotion Classification

Affective Reasoning

arXiv :2607 . 10238v1 [ cs .CV] 11 Jul 2026 Benchmarking Dynamic Affective Reasoning: A Viewer-Centric Video Emotion Dataset Zhiyan Zhang 1, Peipei Song 1⋆ , Jinpeng Hu2, Jingyang Jia 1, Xun Yang 1, and Xiaojun Chang 1 1 University of Science and Technology of China, Hefei, China [zzyhang02@gmail.com](zzyhang02@gmail.com) , [beta.songpp@gmail.com](beta.songpp@gmail.com) , [jjygood@mail.ustc.edu.cn](jjygood@mail.ustc.edu.cn) , [xyang21@ustc.edu.cn](xyang21@ustc.edu.cn) , [xjchang@ustc.edu.cn](xjchang@ustc.edu.cn) 2 Hefei University of Technology, Hefei, China [135858hjp@gmail.com](135858hjp@gmail.com) 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 DAR (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 anew 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 DARR1 , 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](https://github.com/Zhang-Zhiyan/DAR). Keywords: Dynamic Affective Reasoning · Video Emotion Analysis · Large-scale Dataset · Reinforcement Learning 1 Introduction Affective computing [31] aims to equip machines with the ability to perceive, interpret, and reason about human affect, enabling applications from human– computer interaction [13] to psychological counseling and psychotherapy [27,43, 44,48] . As AI systems become embedded in everyday products [45], robust emo tion understanding is increasingly essential for natural, adaptive, and socially aware interaction [37] . Recent Multimodal Large Language Models (MLLMs) [1,⋆ Corresponding author. 2 Z. Zhang et al. 0.0s 5.5s 8.5s 16. 1s Task 1: Affective Segmentation Split 1 Split 2 Temporal Localization 0.0s-5 .5s 5.5s-8 .5s 8.5-16. 1s Task 2: Fine-grained Emotion Classification From 27 Categories Anxiety Fear Empathic Pain 0.0s 5.5s 8.5s 16. 1s [{Start: 0.0s, End: 5.5s. }, {Start: 5.5s, End: 8.5s. }, {Start: 8.5s, End: 16.1s. }] [{Start: 0.0s, End: 5.5s, Emotion: Anxiety. },{Start: 5.5s, End: 8.5s, Emotion: Fear. }, {Start: 8.5s, End: 16.1s, Emotion: Empathic Pain. }] Task 3: Affective Reasoning [{0 .0s-5.5s, Anxiety, The viewer feels anxiety due to the stark …}, Causal Logic Explanation 0.0s ue tAnxioet…y 5.B5s auseFear 8. .5s EThempamthain ic Psa 16. 1s{5{8 ..5s5s--8.165s1,searEm,ThepathivciewPaerin,’sTehemvoitionewershnoiftswffromeelsaEnmxpietyathtoicPfeaainr}, Fig. 1: Overview of the DAR Benchmark Tasks. We formulate three hierarchical tasks: (1) Affective Segmentation for locating temporal boundaries of emotion shifts; (2) Fine-grained Emotion Classification utilizing a 27-category viewer-centric taxonomy; and (3) Affective Reasoning for generating causal explanations grounded in visual evidence. 9, 46, 47] have further advanced video-based emotion understanding and recognition [10, 16] . Foundational benchmarks such as DFEW [19] and VCE [29] enabled systematic evaluation of video emotion understanding,

FAQs

What problem does DAR address in existing video emotion analysis benchmarks?

Existing benchmarks often treat video emotion as a static label for an entire clip. DAR targets the overlooked dynamic nature of emotions as they change due to consecutive causal events.

What does the DAR dataset contain and how is it annotated?

DAR provides 15,087 videos and 36,908 event-aligned affective segments. Each segment is annotated with 27 emotion categories and aligned to event-level temporal segments.

What are the three benchmark tasks introduced with DAR?

DAR defines affective segmentation to locate emotion-shift boundaries, fine-grained emotion classification using a 27-category taxonomy, and affective reasoning to generate causal logic grounded in visual evidence.

Jul 16, 2026

Upload by Seraphina