Research map / Vision-language and multimodal / Multimodal LLMs

EMO-R3: Reflective Reinforcement Learning for Emotional Reasoning in Multimodal Large Language Models

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 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 EMO-R3, a GRPO-based RL framework for visual emotion reasoning that adds a structured three-step emotional thinking prompt and a reflective reward (image-text consistency plus emotional coherence via re-prompting the model), evaluated on Qwen2.5-VL-3B.

Problem

MLLMs struggle with subjective visual emotion understanding; SFT generalizes poorly and vanilla GRPO reasoning traces are not emotion-oriented and often disconnect from the final emotion answer.

Approach

Structured Emotional Thinking (trigger identification, human emotional reflection, emotional conclusion) combined with a Reflective Emotional Reward that feeds reasoning steps back into the model to check visual grounding (Yes/No) and whether the inferred emotion matches the label, added to accuracy and format rewards under GRPO.

Key results

• On Qwen2.5-VL-3B with rollout-8, EMO-R3 reaches 60.42 overall accuracy vs 58.28-59.16 for GRPO/DAPO baselines (reported best at rollout-4: 60.50 vs GRPO 59.97).
• Cold-Start-Emo variant (EMO-R3#) achieves 62.13 overall accuracy vs 60.92 for GRPO in the ablation table.
• Ablation: adding SET raises overall accuracy from 60.92 to 61.21; adding RER further raises it to 61.79.

Surprising findings

• DAPO failed to complete training on the emotion task, attributed to a mismatch between its filtering and the binary emotion reward.
• SFT achieves strong in-domain EmoSet accuracy (77.15) but drops sharply out-of-domain (Emotion6 34.51, WebEmo 17.75).

Limitations

• Experiments limited to a 3B model and two-to-three emotion datasets with random 2,000-sample subsets.
• Reflective reward relies on the same model for consistency and coherence judgments, with no independent verification shown.
• Reported results are median of three runs; variance and significance not discussed in detail.

Experimental setup

Datasets: EmoSet Emotion6 WebEmo

Benchmarks: EmoSet Emotion6 WebEmo

Metrics: accuracy average in-domain accuracy average out-of-domain accuracy

Baselines: Vanilla Qwen2.5-VL-3B-Instruct SFT GRPO DAPO SEPM LLaVA1.5-7B Vanilla SEPM on LLaVA1.5-7B

Builds on: GRPO DAPO EasyR1 Qwen2.5-VL SEPM

Closest papers by meaning

• Unveiling the Cognitive Compass: Theory-of-Mind–Guided Multimodal Emotion Reasoning (ICLR 2026 · similarity 0.91)
• RLVER: Reinforcement Learning with Verifiable Emotion Rewards for Empathetic Agents (ICLR 2026 · similarity 0.90)
• MME-Emotion: A Holistic Evaluation Benchmark for Emotional Intelligence in Multimodal Large Language Models (ICLR 2026 · similarity 0.89)
• EmotionThinker: Prosody-Aware Reinforcement Learning for Explainable Speech Emotion Reasoning (ICLR 2026 · similarity 0.89)
• Catch Your Emotion: Sharpening Emotion Perception in Multimodal Large Language Models (ICML 2025 · similarity 0.89)
• Emotion-Coherent Reasoning for Multimodal LLMs via Emotional Rationale Verifier (AAAI 2026 · similarity 0.89)