MuPHI: Learning Implicit Multimodal Harm Reasoning via Semantically Grounded Reward Optimization

Anisha Saha; Sophia Wiedmann; Teodora Kamova; Timothy Hospedales; Varsha Suresh; Vera Demberg

REVIEW 2 major objections 2 minor 14 references

MuPHI: Learning Implicit Multimodal Harm Reasoning via Semantically Grounded Reward Optimization

T0 review · 2 major / 2 minor · reviewed 2026-06-29 · grok-4.3

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Optimizing multi-perspective rewards on the MuPHI dataset lets vision-language models detect and reason about implicit multimodal harm more accurately and robustly.

desk verdict MuPHI supplies a useful new dataset for implicit multimodal harm with rationales, and MuPHIRM shows some gains from multi-perspective reward training, but the OOD robustness claim needs tighter verification on whether the test split actually forces novel transfer. read the letter →

arxiv 2605.29951 v1 pith:O76FD3Q3 submitted 2026-05-28 cs.AI cs.CLcs.LGcs.MM

Anisha Saha , Varsha Suresh , Teodora Kamova , Sophia Wiedmann , Timothy Hospedales , Vera Demberg This is my paper · ORCID

classification cs.AIcs.CLcs.LGcs.MM

keywords multimodalharmdetectionvision-languagemodelsrewardoptimizationimplicitreasoningMuPHIdatasetout-of-distributionrobustnesscompositionalsemantics

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Vision-language models typically handle literal image-text cues but miss harmful meanings that depend on subtle context and intent. The paper introduces the MuPHI dataset of such pairs along with annotated rationales and proposes MuPHIRM, a training method that optimizes rewards from multiple perspectives to learn joint semantics. A sympathetic reader would expect this to produce models that generalize better to unseen harm distributions instead of latching onto surface shortcuts. The reported results show gains in both detection accuracy and reasoning quality over trained and inference-time baselines.

What carries the argument

MuPHIRM, a reasoning-augmented training framework that optimizes multi-perspective rewards to learn joint semantics across image-text pairs.

What would settle it

A controlled test set of image-text pairs with new harm categories outside MuPHI training distributions where a MuPHIRM-trained model shows no improvement in detection accuracy or reasoning quality over baselines would falsify the central claim.

Watch

Extended reading notes

Core claim

MuPHIRM is a reasoning-augmented training framework that learns joint semantics by optimizing multi-perspective rewards on the MuPHI dataset of image-text pairs where harm arises from subtle multimodal cues; this produces vision-language models with improved harm detection, higher-quality reasoning chains, and superior out-of-distribution robustness compared with standard trained and inference-time baselines.

Load-bearing premise

Multi-perspective reward optimization on the MuPHI dataset produces joint semantics that transfer beyond the specific harm categories and distributions used in training.

Editorial extensions

If this is right

Vision-language models trained via MuPHIRM achieve higher harm detection accuracy on compositional multimodal inputs.
The same models generate higher-quality reasoning chains supported by the annotated rationales in MuPHI.
MuPHIRM-trained models exhibit stronger performance on out-of-distribution harm examples than both fine-tuned and prompt-based baselines.
Reasoning-oriented reward optimization provides a route to multimodal systems that avoid benchmark-specific shortcuts.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The same multi-perspective reward structure could be applied to other implicit multimodal tasks such as detecting sarcasm or figurative language.
If the learned joint semantics prove general, the method might reduce dependence on exhaustive category-specific annotations for new harm types.
Deployment on live social-media streams would test whether the robustness gains survive noisy, real-world distributions not represented in MuPHI.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

read the letter

The paper's main contribution is a dataset of image-text pairs where harm sits in subtle cross-modal cues rather than obvious signals, plus a training recipe that optimizes several reward views at once to push VLMs toward better detection and chain-of-thought explanations. That setup is new enough in the safety literature to be worth noting.

It does a few things cleanly. The dataset covers multiple harm categories and ships annotated rationales, which lets them measure reasoning quality directly instead of just accuracy. The method frames the problem as learning joint semantics through reward optimization, and the abstract reports better performance than both fine-tuned and inference-only baselines.

The soft spot is the out-of-distribution story. The strongest claim is superior robustness that shows the model is not just memorizing benchmark shortcuts. If the OOD split reuses the same harm categories or rationale styles as the training data, even with different surface cues, the measured gains do not yet demonstrate that the multi-perspective rewards produced transferable joint semantics. The abstract leaves the exact OOD construction and any category-level breakdowns unspecified, so that part of the argument is the least secure.

This work is aimed at researchers doing VLM alignment and multimodal safety. A reader already working on harm detection or reward modeling for vision-language models will get concrete artifacts and a clear experimental direction. The central idea is coherent on its own terms and engages the right prior literature, so it clears the bar for a serious referee even if the OOD results require more scrutiny in revision.

Referee Report

2 major / 2 minor

Summary. The paper introduces the MuPHI dataset of image-text pairs encoding harm via subtle multimodal cues across diverse categories, with annotated rationales. It proposes MuPHIRM, a reasoning-augmented training framework that optimizes multi-perspective rewards to learn joint semantics in VLMs, claiming improvements in harm detection accuracy, reasoning quality, and superior out-of-distribution robustness relative to trained and inference-time baselines.

Significance. If the central empirical claims hold after verification of the OOD construction and ablations, the work would provide a concrete dataset and reward-optimization approach for addressing implicit compositional harm in multimodal models, moving evaluation beyond literal perceptual cues.

major comments (2)

[§5.3] §5.3 (OOD Evaluation): The construction of the out-of-distribution test instances is not specified in sufficient detail to establish that they introduce novel implicit harm combinations outside the training harm categories and rationale styles; without this, the reported robustness gains cannot be confirmed to reflect transfer of joint semantics rather than benchmark-specific patterns.
[§4.2] §4.2 (MuPHIRM Training): The multi-perspective reward formulation is described at a high level but lacks an explicit statement of whether the reward components are derived from external signals independent of the MuPHI training split or involve any fitted parameters on the same data; this directly affects the claim that the method avoids benchmark-specific shortcuts.

minor comments (2)

[Abstract / §3] The abstract states that MuPHI 'spans diverse harm categories' but does not quantify the number of categories or the distribution of implicit vs. explicit cues; adding these statistics in §3 would strengthen the dataset description.
[Table 2] Table 2 (main results) reports aggregate metrics but does not break down performance by harm category; including per-category scores would clarify whether gains are uniform or driven by a subset of categories.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive feedback on our work. We address each major comment below with clarifications and commit to revisions that strengthen the presentation of the OOD construction and reward formulation without altering the core claims.

read point-by-point responses

Referee: [§5.3] §5.3 (OOD Evaluation): The construction of the out-of-distribution test instances is not specified in sufficient detail to establish that they introduce novel implicit harm combinations outside the training harm categories and rationale styles; without this, the reported robustness gains cannot be confirmed to reflect transfer of joint semantics rather than benchmark-specific patterns.

Authors: We agree that §5.3 would benefit from greater explicitness. The OOD test set was constructed by partitioning the MuPHI data such that entire harm categories (e.g., certain implicit threat and manipulation subtypes) and distinct rationale annotation styles are held out from the training split, with no overlap in the underlying semantic combinations. In the revised manuscript we will expand §5.3 (and the corresponding appendix) to include the exact category-disjoint splits, the rationale-style partitioning criteria, and concrete examples of novel multimodal cue combinations. This addition will make clear that the reported robustness reflects generalization of joint semantics rather than memorization of training patterns. revision: yes
Referee: [§4.2] §4.2 (MuPHIRM Training): The multi-perspective reward formulation is described at a high level but lacks an explicit statement of whether the reward components are derived from external signals independent of the MuPHI training split or involve any fitted parameters on the same data; this directly affects the claim that the method avoids benchmark-specific shortcuts.

Authors: We acknowledge the need for an explicit statement. The multi-perspective rewards in MuPHIRM are computed from external, pre-trained models (for semantic grounding and harm classification) together with human-annotated rationales drawn exclusively from a held-out validation portion of MuPHI that is never used in the training split; no parameters of the reward models are fitted on the MuPHI training data. In the revision we will insert a dedicated paragraph in §4.2 (and a short note in the method overview) that states this independence explicitly, thereby reinforcing that the optimization does not rely on benchmark-specific fitted shortcuts. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: method framed as learning from external rewards and new dataset without reduction to fitted inputs

full rationale

The provided abstract and description introduce MuPHI as an external dataset with annotated rationales and MuPHIRM as a framework optimizing multi-perspective rewards to learn joint semantics. Claims of improved detection, reasoning, and OOD robustness are presented as outcomes of this optimization rather than quantities defined by construction from the same fitted parameters. No equations, self-citations, or derivations are shown that would make predictions equivalent to inputs. The approach is described as using reward signals external to the model, aligning with a self-contained derivation chain. No load-bearing steps reduce by the paper's own text to self-definition or renaming of known results.

Assumptions & free parameters 0 free parameters · 0 assumptions · 0 invented entities

Only the abstract is available; no free parameters, axioms, or invented entities are described in sufficient detail to populate the ledger.

how reviews work

0 comments

Cite this review

Pith. "Pith review of MuPHI: Learning Implicit Multimodal Harm Reasoning via Semantically Grounded Reward Optimization." pith.science (2026). https://pith.science/paper/O76FD3Q3

@misc{pith2026260529951,
  author       = {Pith},
  title        = {Pith review of: MuPHI: Learning Implicit Multimodal Harm Reasoning via Semantically Grounded Reward Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/O76FD3Q3}},
  note         = {Machine review of arXiv:2605.29951}
}

read the original 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.

Figures

Figures reproduced from arXiv: 2605.29951 by the authors.

**Figure 2.** Examples of Harmful (a)-(b) and Benign (c)- [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Overview of MuPHIRM training pipeline. We first train a base VLM via SFT on labels followed by GRPO-based reward optimization on the warm-start model to improve harm reasoning. dataset-agnostic and can be used to obtain rationales for other datasets. To assess the quality of the automatically generated annotations, we choose a balanced subset of 330 rationales to be reviewed by two human annotators. The annotator… view at source ↗

Figures from the paper (8 more)

**Figure 4.** Figure 4: Attention attribution comparison between SFT [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

**Figure 5.** Figure 5: Class distribution of harmful memes in MuPHI [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]

**Figure 7.** Figure 7: Reasoning word count distribution [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]

**Figure 8.** Figure 8: Lexical diversity (MTLD) per category. Rationale Type Count Gold (Human-Annotated) 330 Silver 1,264 Total 1,594 [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]

**Figure 6.** Figure 6: Subclass distribution of harmful memes across fine-grained harm categories. [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]

**Figure 9.** Figure 9: Class-wise distribution of errors for MuPHIRM on MuPHI test set. This evaluation tests whether models update their predictions based on the changed image-text interaction rather than relying only on visual cues or textual cues alone [PITH_FULL_IMAGE:figures/full_fig_…

**Figure 10.** Figure 10: Qualitative examples of model-generated reasoning on MuPHI and FHM counterfactual samples of both [PITH_FULL_IMAGE:figures/full_fig_p020_10.png]

**Figure 11.** Figure 11: Attribution map of tokens indicating harm in the model-generated harm reasoning. Letters in [PITH_FULL_IMAGE:figures/full_fig_p022_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references · 2 canonical work pages

[1]

LLMs-as-Judges: A Comprehensive Survey on LLM-based Evaluation Methods

Llms-as-judges: a comprehensive survey on llm-based evaluation methods.arXiv preprint arXiv:2412.05579. Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harri- son Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe

work page Pith review arXiv
[2]

InInternational Conference on Learning Representations, volume 2024, pages 39578–39601

Let’s verify step by step. InInternational Conference on Learning Representations, volume 2024, pages 39578–39601. Hongzhan Lin, Ziyang Luo, Wei Gao, Jing Ma, Bo Wang, and Ruichao Yang. 2024. Towards explain- able harmful meme detection through multimodal de- bate between large language models. InProceedings of the ACM web conference 2024, pages 2359–2370...

2024
[3]

InFindings of the Associ- ation for Computational Linguistics: EMNLP 2023, pages 9114–9128, Singapore

Beneath the surface: Unveiling harmful memes with multimodal reasoning distilled from large language models. InFindings of the Associ- ation for Computational Linguistics: EMNLP 2023, pages 9114–9128, Singapore. Association for Com- putational Linguistics. Zhihang Lin, Mingbao Lin, Yuan Xie, and Rongrong Ji. 2026. Cppo: Accelerating the training of group ...

2023
[4]

InProceedings of the 60th Annual Meet- ing of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8253–8280, Dublin, Ireland

V ALSE: A task-independent benchmark for vision and language models centered on linguistic phenomena. InProceedings of the 60th Annual Meet- ing of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8253–8280, Dublin, Ireland. Association for Computational Linguistics. Shraman Pramanick, Shivam Sharma, Dimitar Dim- itrov, Md. Sha...

2021
[5]

Naquee Rizwan, Paramananda Bhaskar, Mithun Das, Swadhin Satyaprakash Majhi, Punyajoy Saha, and Animesh Mukherjee

Direct preference optimization: Your language model is secretly a reward model.Advances in neural information processing systems, 36:53728–53741. Naquee Rizwan, Paramananda Bhaskar, Mithun Das, Swadhin Satyaprakash Majhi, Punyajoy Saha, and Animesh Mukherjee. 2025. Exploring the limits of zero shot vision language models for hate meme de- tection: The vul...

2025
[6]

DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

Finding and exploring memes in social me- dia. InProceedings of the 23rd ACM conference on Hypertext and social media, pages 295–304. Anisha Saha, Varsha Suresh, Timothy Hospedales, and Vera Demberg. 2026. Mustreason: A benchmark for diagnosing pragmatic reasoning in videolms for multimodal sarcasm detection. InProceedings of the Fifteenth Language Resour...

work page Pith review arXiv 2026
[7]

Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross

PMLR. Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. 2022. Winoground: Probing vision and lan- guage models for visio-linguistic compositionality. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 5238– 5248. Zhuo Xu, Xiang Xiang, and Yifan Liang. 2025. Ov...

2022
[8]

View a series of image-text examples

Show all 14 references

[9]

Review whether these examples belong to the harmful or non-harmful category. 3https://pypi.org/project/bert-score/ 4https://huggingface.co/sentence-transformers/all-mpnet- base-v2 Figure 6: Subclass distribution of harmful memes across fine-grained harm categories. Label Word ...
[10]

Potential Risks and Discomfort: Contents you will see may include:

Read model generated explanations about whether the content is harmful or non-harmful and modify it based on whether the reasoning captures the implicit harm. Potential Risks and Discomfort: Contents you will see may include:
[11]

References to sensitive topics (e.g., race, gender, religion, politics, homophobia, sex- ual content, pornographic imagery, violent graphic depictions, hateful slurs, depictions of physical harm, fraud, hatespeech)
[12]

Implicit or explicit stereotypes
[13]

Notes: If at any time you feel uncomfortable, you may,

Content that discusses harmful or offensive themes. Notes: If at any time you feel uncomfortable, you may,
[14]

"create chaos

Stop annotating immediately. Compensation: The annotators are research as- sistants (RAs) employed in the lab and payed ac- cording to the standard university pay-scale for RAs. B Methodology B.1 Problem Formulation Given an image I, with embedded text T , and the gold harm la...

2024

Pith tools

Reviewed June 29, 2026 · model on record in the stance chip above.

[1] [1]

LLMs-as-Judges: A Comprehensive Survey on LLM-based Evaluation Methods

Llms-as-judges: a comprehensive survey on llm-based evaluation methods.arXiv preprint arXiv:2412.05579. Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harri- son Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe

work page Pith review arXiv

[2] [2]

InInternational Conference on Learning Representations, volume 2024, pages 39578–39601

Let’s verify step by step. InInternational Conference on Learning Representations, volume 2024, pages 39578–39601. Hongzhan Lin, Ziyang Luo, Wei Gao, Jing Ma, Bo Wang, and Ruichao Yang. 2024. Towards explain- able harmful meme detection through multimodal de- bate between large language models. InProceedings of the ACM web conference 2024, pages 2359–2370...

2024

[3] [3]

InFindings of the Associ- ation for Computational Linguistics: EMNLP 2023, pages 9114–9128, Singapore

Beneath the surface: Unveiling harmful memes with multimodal reasoning distilled from large language models. InFindings of the Associ- ation for Computational Linguistics: EMNLP 2023, pages 9114–9128, Singapore. Association for Com- putational Linguistics. Zhihang Lin, Mingbao Lin, Yuan Xie, and Rongrong Ji. 2026. Cppo: Accelerating the training of group ...

2023

[4] [4]

InProceedings of the 60th Annual Meet- ing of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8253–8280, Dublin, Ireland

V ALSE: A task-independent benchmark for vision and language models centered on linguistic phenomena. InProceedings of the 60th Annual Meet- ing of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8253–8280, Dublin, Ireland. Association for Computational Linguistics. Shraman Pramanick, Shivam Sharma, Dimitar Dim- itrov, Md. Sha...

2021

[5] [5]

Naquee Rizwan, Paramananda Bhaskar, Mithun Das, Swadhin Satyaprakash Majhi, Punyajoy Saha, and Animesh Mukherjee

Direct preference optimization: Your language model is secretly a reward model.Advances in neural information processing systems, 36:53728–53741. Naquee Rizwan, Paramananda Bhaskar, Mithun Das, Swadhin Satyaprakash Majhi, Punyajoy Saha, and Animesh Mukherjee. 2025. Exploring the limits of zero shot vision language models for hate meme de- tection: The vul...

2025

[6] [6]

DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

Finding and exploring memes in social me- dia. InProceedings of the 23rd ACM conference on Hypertext and social media, pages 295–304. Anisha Saha, Varsha Suresh, Timothy Hospedales, and Vera Demberg. 2026. Mustreason: A benchmark for diagnosing pragmatic reasoning in videolms for multimodal sarcasm detection. InProceedings of the Fifteenth Language Resour...

work page Pith review arXiv 2026

[7] [7]

Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross

PMLR. Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. 2022. Winoground: Probing vision and lan- guage models for visio-linguistic compositionality. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 5238– 5248. Zhuo Xu, Xiang Xiang, and Yifan Liang. 2025. Ov...

2022

[8] [8]

View a series of image-text examples