aiqbench analysis
This research addresses the critical challenge of detecting harm in multimodal models where benign images and captions combine to create unsafe content, framing it as a cross-modal reasoning problem. The authors introduce MuPHI, a diagnostic benchmark revealing that current Vision-Language Models fail to bind evidence into harm mechanisms, leading to poor out-of-distribution transfer. To solve this, they propose CHIL, a learning framework that prioritizes the reasoning chain and counterfactual robustness over simple label prediction, significantly improving generalization capabilities.
Learning to Reason About Implicit Multimodal Harm: From Diagnosis to Generalization
arXiv:2605.29951v2 Announce Type: replace-cross Abstract: A benign image paired with a benign caption can become harmful only in combination, making detection fundamentally a cross-modal reasoning problem. Existing multimodal harm benchmarks evaluate only classification accuracy, leaving unclear whether models identify the cross-modal mechanism that produces harm. We introduce MuPHI (Multimodal Pragmatic Harm Interpretation), a controlled diagnostic benchmark in which harm arises purely from image-text composition, with benign counterfactuals and annotated harm rationales. Using MuPHI, we systematically evaluate where compositional harm reasoning breaks in vision-language models (VLMs). MuPHI reveals that VLMs reliably ground each modality yet fail to bind the evidence into the harm mechanism and their predictions respond weakly to label-determining counterfactual interventions. Moreover, existing harm detection approaches trained on labels substantially improves in-domain (ID) detection but transfers poorly across harm distributions. We therefore propose Compositional Harm Interpretation Learning (CHIL), which rewards evidence identification in each modality, cross-modal binding and verdict consistency in the reasoning chain rather than just predicted label. CHIL improves 81% OOD transfer evaluations by an average of +2.5 macro-F1 and correct counterfactual flip rate by up to +3.8 and +5.4 % points, while also improving overall reasoning quality over label-only optimization. These results suggest that optimizing how a model reasons, not just what it concludes, offers a promising direction for multimodal safety systems that generalize beyond dataset-specific correlations.
Abstract:A benign image paired with a benign caption can become harmful only in combination, making detection fundamentally a cross-modal reasoning problem. Existing multimodal harm benchmarks evaluate only classification accuracy, leaving unclear whether models identify the cross-modal mechanism that produces harm. We introduce MuPHI (Multimodal Pragmatic Harm Interpretation), a controlled diagnostic benchmark in which harm arises purely from image-text composition, with benign counterfactuals and annotated harm rationales. Using MuPHI, we systematically evaluate where compositional harm reasoning breaks in vision-language models (VLMs). MuPHI reveals that VLMs reliably ground each modality yet fail to bind the evidence into the harm mechanism and their predictions respond weakly to label-determining counterfactual interventions. Moreover, existing harm detection approaches trained on labels substantially improves in-domain (ID) detection but transfers poorly across harm distributions. We therefore propose Compositional Harm Interpretation Learning (CHIL), which rewards evidence identification in each modality, cross-modal binding and verdict consistency in the reasoning chain rather than just predicted label. CHIL improves 81% OOD transfer evaluations by an average of +2.5 macro-F1 and correct counterfactual flip rate by up to +3.8 and +5.4 % points, while also improving overall reasoning quality over label-only optimization. These results suggest that optimizing how a model reasons, not just what it concludes, offers a promising direction for multimodal safety systems that generalize beyond dataset-specific correlations.
Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Machine Learning (cs.LG); Multimedia (cs.MM)
arXiv:2605.29951 [cs.AI]
(or arXiv:2605.29951v2 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2605.29951 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 317 words
Related coverage
Evidence-Grounded Trustworthy Multimodal Reasoning and Evaluation Benchmark in Complex Urban Scenes
arXiv · Sep 1, 2026
From Recognition to Reasoning: Advancing Multimodal Harmful Meme Detection via Chain-of-Thought Alignment
arXiv · Aug 25, 2026
Evidence-Grounded Trustworthy Multimodal Reasoning and Evaluation Benchmark in Complex Urban Scenes
arXiv · Aug 28, 2026
SyRHM: Symbolic-Language-Enhanced Reasoning with Associative Retrieval for Zero-shot Harmful Meme Detection
arXiv · Sep 15, 2026