ESC: Emotional Self-Correction for Reliable Vision-Language Models

Vision-language models (VLMs) have achieved strong performance across diverse multimodal tasks, yet they remain vulnerable to unreliable reasoning. Existing self-correction methods mitigate these issues but typically rely on post-training or carefully engineered feedback, incurring high computational cost. In this work, we revisit this challenge through the lens of emotional cues, asking whether they can activate latent self-correction behaviors in VLMs without additional training. $\mathbf{\text{We}}\mathbf{\text{find}}\mathbf{\text{that}}\mathbf{\text{emotional}}\mathbf{\text{signals}}\mathbf{\text{serve}}\mathbf{\text{as}}\mathbf{\text{an}}\mathbf{\text{effective}}\mathbf{\text{trigger}}\mathbf{\text{for}}\mathbf{\text{self-correction,}}\mathbf{\text{encouraging}}\mathbf{\text{more}}\mathbf{\text{cautious}}\mathbf{\text{and}}\mathbf{\text{reflective}}\mathbf{\text{reasoning}}$. Motivated by this finding, we propose \escabstract ($\underset{‾}{\mathbf{\text{E}}}$motional $\underset{‾}{\mathbf{\text{S}}}$elf-$\underset{‾}{\mathbf{\text{C}}}$orrection), a training-free self-correction framework. ESC introduces an external verifier that detects potentially incorrect initial responses and injects emotional feedback to encourage model to reflect, and produce a better revised response without additional training. Extensive experiments across safety, hallucination, vision-centric perception, and multimodal reasoning benchmarks show that ESC consistently improves reliability while preserving overall model utility. These results suggest that emotion can function not only as an ability to be recognized, but also as a practical control signal for scalable self-correction in VLMs. $\mathbf{\text{We}}\mathbf{\text{therefore}}\mathbf{\text{believe}}\mathbf{\text{that}}\mathbf{\text{ESC}}\mathbf{\text{provides}}\mathbf{\text{a}}\mathbf{\text{strong}}\mathbf{\text{foundation}}\mathbf{\text{for}}\mathbf{\text{a}}\mathbf{\text{new}}\mathbf{\text{reliable}}\mathbf{\text{human-like,}}\mathbf{\text{emotion-integrated}}\mathbf{\text{research}}\mathbf{\text{direction.}}$ Our project is publicly available at \textcolor{red}{https://genai4e.github.io/ESC/}.