Research map / Vision-language and multimodal / Multimodal LLMs
MASP: Multi-Aspect Guided Emotion Reasoning with Soft Prompt Tuning In Vision-Language Models
SangEun Lee, Yubeen Lee, Eunil Park, Wonseok Chae
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MASP, a multi-aspect guided emotion reasoning method for vision-language models that uses soft prompt tuning instead of costly instruction tuning to capture multi-faceted emotional cues in images.
Problem
Emotion understanding from images in vision-language models; existing methods rely on global visual representations that miss nuanced multi-faceted emotional cues and typically use costly instruction-tuning datasets.
Approach
Guides emotion reasoning with multiple emotion-related aspects (local visual cues) and trains soft prompts rather than fine-tuning on instruction data, targeting a cheaper alternative to instruction tuning.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add trained aspect modules to zero-shot VLM without soft prompt (ablation) | Qwen2-VL-7B · FI | accuracy (%): 58.16 -> 44.54 | Hurt |
| add trained soft prompt to zero-shot VLM (ablation) | Qwen2-VL-7B · FI | accuracy (%): 58.16 -> 76.63 | Improved |
| add trained soft prompt to zero-shot VLM (ablation) | Qwen2-VL-7B · Emotion6 | accuracy (%): 51.68 -> 70.88 | Improved |
| add aspect modules on top of the soft prompt (ablation) | MASP · FI | accuracy (%): 76.63 -> 80.06 | Improved |
| add aspect modules on top of the soft prompt (ablation) | MASP · Emotion6 | accuracy (%): 70.88 -> 74.75 | Improved |
| use random untrained soft prompt with aspect modules instead of trained prompt (ablation) | MASP · FI | accuracy (%): 80.06 -> 41.47 | Hurt |
| use all six aspect modules instead of one (ablation) | MASP · FI | accuracy (%): 78.72 -> 80.06 | Improved |
| use both brightness and colorfulness modules instead of one (ablation) | MASP · Abstract | accuracy (%): 39.13 -> 45.65 | Improved |
| increase soft prompt length from 32 to 128 (ablation) | MASP · ArtPhoto | accuracy (%): 53.09 -> 42.59 | Hurt |
| reduce soft prompt length from 32 to 8 (ablation) | MASP · Emotion6 | accuracy (%): 74.75 -> 70.20 | Hurt |
| use soft prompt tuning instead of QLoRA instruction tuning | Qwen2-VL-7B · trainable parameters | trainable parameters: 2,523,136 -> 114,688 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- EmoVIT — EmoVIT: Revolutionizing Emotion Insights with Visual Instruction Tuning (CVPR 2024) · better on 4 of 5 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- EmoVIT: Revolutionizing Emotion Insights with Visual Instruction Tuning (CVPR 2024 · similarity 0.89)
- Catch Your Emotion: Sharpening Emotion Perception in Multimodal Large Language Models (ICML 2025 · similarity 0.88)
- Customizing Visual Emotion Evaluation for MLLMs: An Open-vocabulary, Multifaceted, and Scalable Approach (ICLR 2026 · similarity 0.86)
- Interpreting and Enhancing Emotional Circuits in Large Vision-Language Models via Cross-Modal Information Flow (ICML 2026 · similarity 0.86)
- AffectGPT: A New Dataset, Model, and Benchmark for Emotion Understanding with Multimodal Large Language Models (ICML 2025 · similarity 0.86)
- Facial-R1: Aligning Reasoning and Recognition for Facial Emotion Analysis (AAAI 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
