Problem
Research questions and friction points this paper is trying to address.
multimodal harm
implicit reasoning
vision-language models
context-dependent semantics
harm detection
MuPHI: Learning Implicit Multimodal Harm Reasoning via Semantically Grounded Reward Optimization
🤖 AI Summary
Current vision-language models struggle to detect context-dependent implicit multimodal harms in image-text pairs due to insufficient intent awareness and reasoning capabilities. To address this limitation, this work introduces MuPHI, the first multimodal dataset of implicit harms annotated with explicit reasoning rationales, and proposes the MuPHIRM training framework. MuPHIRM leverages a multi-perspective reward mechanism to enhance semantic alignment and foster coherent cross-modal reasoning chains, thereby moving beyond reliance on superficial features. Experimental results demonstrate that MuPHIRM significantly improves both implicit harm detection accuracy and reasoning quality, while exhibiting superior generalization and robustness in out-of-distribution scenarios.
📝 Abstract
Understanding how harm emerges from interaction between otherwise benign image-text pairs requires intent-aware cross-modal reasoning beyond surface-level features. Existing vision-language models (VLMs) excel at literal reasoning over perceptual cues but often fail to derive harmful semantics that rely on implicit, context-dependent reasoning. To evaluate VLMs on compositional harm detection and reasoning, we introduce Multimodal Pragmatic Harm Interpretation (MuPHI), a dataset containing image-text pairs where harm is encoded in subtle multimodal cues. MuPHI spans diverse harm categories and includes annotated harm rationales for assessing VLM reasoning chains. To improve both detection and reasoning in VLMs, we propose MuPHIRM, a reasoning-augmented training framework which learns joint semantics by optimizing multi-perspective rewards. MuPHIRM improves both harm detection and reasoning quality of VLMs while demonstrating superior out-of-distribution robustness compared to both trained and inference-time baselines. Our findings suggest that reasoning-oriented reward optimization offers a promising direction towards building multimodal systems that generalize beyond benchmark-specific shortcuts.
Innovation
Methods, ideas, or system contributions that make the work stand out.
multimodal harm reasoning
reward optimization
vision-language models
implicit semantics
out-of-distribution robustness
🔎 Similar Papers
No similar papers found.
💼 Related Jobs
