AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

World-Model Inspired Emotion-aware Token Refinement for Training-Free Multimodal Emotion Recognition

Kejun Liu, Zhe Chen, Yuanyuan Liu, ke Wang, Yibing Zhan, Wei Xiang, Hongyan Zhang

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Training-free token reduction for MLLMs

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 a training-free, world-model-inspired emotion-aware token refinement method to improve multimodal emotion recognition with MLLMs.

Problem

MLLMs for multimodal emotion recognition are unreliable because sparse emotional cues are overwhelmed by redundant context; fine-tuning is costly and chain-of-thought prompting relies on heuristic prompts.

Approach

Refines multimodal tokens in a training-free manner, guided by a world-model-inspired view of emotion, instead of relying on heuristic prompting or fine-tuning (details beyond the abstract are not available).

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add NTS (noise-suppressed token selection) (ablation)frozen InternVL3.5-8B · MER23 (InternVL3.5-8B)F1 (%): 63.26 -> 67.84Improved
add STR (state-strengthened token reweighting) (ablation)frozen InternVL3.5-8B · MER23 (InternVL3.5-8B)F1 (%): 63.26 -> 66.22Improved
add STR on top of NTS (ablation)WETR · MER23 (InternVL3.5-8B)F1 (%): 67.84 -> 68.48Improved
use contrastive query scoring instead of random token selection (ablation)NTS · MER23 (InternVL3.5-8B)UAR (%): 57.07 -> 63.24Improved
use contrastive scoring instead of Qpos alone (ablation)NTS · MER23 (InternVL3.5-8B)UAR (%): 61.48 -> 63.24Improved
use contrastive Qpos-Qneg instead of Qneg only (ablation)NTS · MER23 (InternVL3.5-8B)F1 mean (%): 65.99 -> 68.17Improved
add anchor protection to STR (ablation)WETR · MER23 (InternVL3.5-8B)F1 (%): 68.07 -> 68.48Improved
add STR to NTS under 30% post-NTS noise (ablation)WETR · MER23 (InternVL3.5-8B)F1 drop: -2.70 -> -1.76Improved
apply WETR at highest noise level (eps=7)InternVL3.5-8B · MER23 (InternVL3.5-8B)F1: +9.18%Improved
apply WETR to InternVL3.5-8BInternVL3.5-8B · MAFWWAR (%): 55.08 -> 59.64Improved
apply WETR to InternVL3.5-8BInternVL3.5-8B · MER23WAR (%): 60.34 -> 68.13Improved
apply WETR to Qwen3-VL-8BQwen3-VL-8B · MAFWWAR (%): 52.78 -> 56.79Improved

Experimental setup

Baselines: chain-of-thought prompting

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy