acceptodds
Under review as a conference paper at ICLR 2027

Visual Distraction Undermines Moral Reasoning in Vision-Language Models

Abstract

Moral reasoning is fundamental to safe Artificial Intelligence (AI), yet ensuring its consistency across modalities becomes critical as AI systems evolve from text-based assistants to embodied agents. Current safety techniques demonstrate success in textual contexts, but concerns remain about generalization to visual inputs. Existing moral evaluation benchmarks rely on text-only formats and lack systematic control over variables that influence moral decision-making. Here we show that visual inputs substantially alter moral decision-making in state-of-the-art (SOTA) Vision-Language Models (VLMs), revealing limitations in text-based safety alignment. We introduce Moral Dilemma Simulation (MDS), a multimodal benchmark grounded in Moral Foundation Theory (MFT) that enables mechanistic analysis through orthogonal manipulation of visual and contextual variables. The evaluation reveals that the vision modality induces more intuition-like decision patterns and weaker deliberative sensitivity than text-only contexts. These findings expose critical fragilities where language-tuned safety filters fail to constrain visual processing, demonstrating the urgent need for multimodal safety alignment.

Primary area:
datasets and benchmarks
Keywords:
large vision-language models, morality, safety

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_RDhLOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Moral Dilemma Simulation (MDS), a generative benchmark for evaluating moral reasoning in VLMs. MDS generates sandbox-style pixel-art scenes with independent control over conceptual variables (personal force, intention of harm, self-benefit) and character attributes (species, profession, age, wealth, etc.). The authors generate 84k samples and evaluate six VLMs under three input settings: Text (structured descriptions), Caption (model-generated captions with OCR), and Image (raw images). The experiments examine how input modality affects moral decisions, including sensitivity to utilitarian considerations, deontological constraints, and preferences across demographic groups. The results show substantial differences between text and visual inputs.

Critical strengths
  1. MDS extends moral reasoning evaluation beyond text-only benchmarks. Its controlled generation process allows individual factors to be varied independently, making it possible to study how different scenario attributes affect model decisions at scale.

  2. The three-setting evaluation separates the effects of information loss and input modality. Comparing Text with Caption and Caption with Image provides a way to analyze which parts of the performance gap come from the visual representation itself.

  3. The paper includes extensive validation, including human identifiability experiments, a full-dataset grounding audit, rewrite fidelity metrics, per-model OCR checks, and a VQA control. The VQA results show that differences in moral decisions can remain even when models correctly identify the relevant visual information.

Critical weaknesses
  1. Caption mode mixes captioning errors with moral reasoning errors. The authors’ own ablations show that this distinction matters: replacing LLaVA’s generated transcript with ground truth reduces its quantity MAE from 0.2886 to 0.0570. For this model, much of the reported modality gap comes from transcription errors rather than moral reasoning. The main text does not adequately distinguish between these two sources of error when presenting the overall findings.

  2. The model coverage is insufficient for the paper’s SOTA framing. The evaluation excludes frontier models such as GPT-5-class models and Claude. The open-source models use 4-bit quantization without a corresponding control experiment. In addition, Gemini is evaluated with safety thresholds manually disabled, which affects the interpretation of the reported refusal rates.

  3. The dual-process explanation (System 1 vs. System 2) lacks direct evidence. The argument mainly relies on an analogy and a single Grad-CAM example. The appendix also acknowledges that the ablations do not identify the underlying mechanism.

  4. The paper lacks a human moral judgment baseline. Text mode is implicitly treated as the reference for better moral reasoning, but some of its value preferences, such as favoring younger people over older people or civilians over criminals, reflect normative assumptions. Smaller differences in Image mode do not necessarily indicate worse moral reasoning and could instead reflect different value preferences.

Questions for authors
  1. How much of the Text vs. Caption gap comes from caption quality rather than missing information? How would the results change with fixed, high-quality captions?

  2. Given the LLaVA ablation results, can the authors distinguish models where transcription errors explain the modality gap from those where they do not?

  3. Does the reduction in Gemini’s refusal rate still hold under its default safety settings?

  4. Are human moral judgments available for these dilemmas? Without such a baseline, similarity to Text mode measures consistency across input settings rather than alignment with human judgments.

9 comments
AuthorsOpenReview ↗

OCR, proposition presentation, and downstream decisions

We thank the reviewer for recognizing MDS's independent factor control, three-setting evaluation, and extensive validation. We address the suggestions on caption quality, model coverage, experimental settings, and the interpretation of our findings below.

W1

Caption mode mixes captioning errors with moral reasoning errors. The authors’ own ablations show that this distinction matters: replacing LLaVA’s generated transcript with ground truth reduces its quantity MAE from 0.2886 to 0.0570. For this model, much of the reported modality gap comes from transcription errors rather than moral reasoning. The main text does not adequately distinguish between these two sources of error when presenting the overall findings.

Q2

Given the LLaVA ablation results, can the authors distinguish models where transcription errors explain the modality gap from those where they do not?

The cited MAEs concern the Text–Caption context gap, rather than the Caption–Image modality gap, as distinguished in Sec. 3.3 and Figure 4. Text provides a concise, structured dilemma description, whereas Caption decisions use the scene description and extracted proposition. Both conditions reason from language. Their difference concerns the information and contextual presentation supplied to the decision model.

The transcript ablation keeps the visual description fixed and replaces only the extracted proposition with its ground-truth text. A reduction in MAE therefore indicates that transcription contributed to the context gap: correcting the proposition brings Caption decisions closer to the structured Text reference. It does not measure a reduction in the Caption–Image modality gap.

Appendix E.3, Table A3 of the revised PDF reports a LLaVA Caption→SelfCap+GT MAE change of 0.2886→0.0570. Across Quantity records, the comparison is:

ModelCaption MAE to TextSelfCap+GT MAE to Text
LLaVA-v1.6-34B0.28860.0570
Qwen3-VL-8B0.08880.0628
Qwen3-VL-32B0.03690.0333
LLaMA-3.2-90B0.02260.0243

For LLaVA, the MAE drops by 0.2316 (80.3%), and fitted Smax⁡S_{\max} shifts from −0.2442 to +0.1773. Qwen3-VL-8B and Qwen3-VL-32B show smaller MAE decreases (0.0260 and 0.0036), whereas LLaMA increases slightly (0.0017). Transcript errors can thus strongly affect LLaVA's reported context gap, but not all models. Crucially, an 80.3% reduction in distance to Text is not an estimate of the proportion of moral reasoning errors caused by OCR. The description source also affects decisions after transcription is corrected: with the ground-truth transcript fixed, switching Qwen3-VL-8B from SelfCap+GT to Config+GT or GemCap+GT changes its MAE from 0.0628 to 0.0850 or 0.1344 (Appendix E.3). Thus, transcription is a major contributor to LLaVA's reported Text–Caption discrepancy, a smaller contributor for Qwen3-VL-8B, and not a general explanation across models.

Caption–Image modality comparison. Our main analysis compares linguistic and direct visual processing of the same rendered scene, following Sec. 4.1. Qwen3-VL-8B's maximum quantity sensitivity decreases from 0.7502 in Caption to 0.2945 in Image, and LLaMA's decreases from 0.5441 to 0.3747. These results show reduced sensitivity to numerical trade-offs under direct visual input. The transcription controls above explain part of the separate context comparison, while the Caption–Image analysis evaluates the modality-dependent decision patterns central to the paper.

Presenting the proposition as text. The separate Image+Text control removes embedded text from the scene image and supplies the exact proposition as text tokens. Following the Quantity analysis in Sec. 4.1, the results are:

ModelImage MAE to TextImage+Text MAE to Text
LLaVA-v1.6-34B0.82190.1015
Qwen3-VL-8B0.21110.1956
Qwen3-VL-32B0.03850.0112
LLaMA-3.2-90B0.17660.2914

Supplying the proposition as text brings three models closer to their Text profiles, with a particularly large change for LLaVA. LLaMA instead moves farther away. This supplementary control shows that proposition presentation also affects decisions in a model-dependent way.

Together, these controls show that transcript correction can account for part of the Text–Caption gap, but not the full pattern of cross-modal differences. Proposition presentation and the integration of scene pixels with text also matter, in model-dependent ways.

Please see Appendix E.3, Figure A5 and Table A3 for transcript and description-source comparisons, and Figure A4 for Image+Text. Appendices E.4–E.5 provide complementary Single Feature and Interaction controls.

AuthorsOpenReview ↗

Model coverage and 4-bit quantization

W2

The model coverage is insufficient for the paper’s SOTA framing. The evaluation excludes frontier models such as GPT-5-class models and Claude. The open-source models use 4-bit quantization without a corresponding control experiment.

Model coverage. We added GPT-5.6 Luna and Sol, evaluating the same 2105 Quantity and 4941 representative Single Feature configurations in Text, Caption, and Image modes. Both use reasoning.effort=medium and an 8192-token output cap. The new models continue to show modality-dependent moral decisions.

Quantity response curves (Sec. 4.1). The fitted maximum sensitivity Smax⁡S_{\max} is:

ModelTextCaptionImage
GPT-5.6 Luna0.68880.85120.7766
GPT-5.6 Sol1.35410.77331.1926

The response profiles differ across input modes in both models. For Sol, the fitted sensitivity is lower in Image than Text, while Caption produces a further change. The complete curves are included in Appendix E.11, “Evaluation of GPT-5.6 models,” of the revised PDF.

Moral constraints and preferences (Sec. 4.2). The conceptual-factor coefficients are:

Model / factorText log-oddsCaption log-oddsImage log-odds
Luna / Intention of Harm−0.5864−0.4802−0.4933
Luna / Self Benefit+0.0217−0.0227+0.3797
Sol / Intention of Harm−0.5186−0.4143−0.6058
Sol / Self Benefit−0.2164−0.1885−0.0141

Self Benefit shifts toward greater permissibility in Image Mode for both models. Luna's coefficient becomes substantially more positive, while Sol's negative coefficient moves close to zero. Foundation preferences also shift. Care preference increases from Text to Image, from 0.7709 to 0.8431 in Luna and from 0.7504 to 0.8779 in Sol. Fairness decreases from 0.5970 to 0.5080 and from 0.7943 to 0.5769, respectively.

These results extend our core finding to the newer models: changing the input modality changes moral decisions, including responses to self-benefit and priorities across moral foundations. Please see Appendix E.11, “Evaluation of GPT-5.6 models,” for the model descriptions, result tables and accompanying curves in the revised PDF.

Matched precision control. Qwen3-VL-32B and LLaVA-v1.6-34B are the two models evaluated with 4-bit quantization in the original experiments. We therefore compare NF4 and BF16 for these two models, using 2105 Quantity and 1000 stratified Single Feature configurations per model. Each condition is analyzed using its own valid outputs within these predefined sets. Checkpoints, processors, prompts and greedy decoding are held fixed. Caption comparisons use identical captions and OCR. Following the Quantity analysis in Sec. 4.1, the fitted maximum sensitivity is:

ModelPrecisionText Smax⁡S_{\max}Caption Smax⁡S_{\max}Image Smax⁡S_{\max}
Qwen32NF40.44000.76090.3552
Qwen32BF160.51210.76990.3624
LLaVA-34BNF40.26430.08403.1229*
LLaVA-34BBF160.24830.09440.1767

Qwen32 retains a lower Image slope than Text and Caption under both precisions. LLaVA's Image curve remains close to always taking the action, while its Text and Caption curves occupy a much lower range. *For NF4 LLaVA, the Image curve is near the ceiling and the fitted inflection lies outside the tested ratio range, so the plotted curve is more informative than the extrapolated maximum slope. Under BF16, LLaVA's Image curve likewise remains at high action probabilities. Thus, the cross-modal response differences are also present under BF16.

The Single Feature analysis in Sec. 4.2 gives a consistent example in Qwen32: its Self Benefit Image−Text log-odds shift is +0.488 under NF4 and +0.394 under BF16. Please see Appendix E.10, “Matched precision control,” for the protocol and results, and Appendix E.11 for the newer-model extension.

AuthorsOpenReview ↗

Gemini safety settings and refusal interpretation, dual-process interpretation and mechanistic claims

W2

In addition, Gemini is evaluated with safety thresholds manually disabled, which affects the interpretation of the reported refusal rates.

Q3

Does the reduction in Gemini’s refusal rate still hold under its default safety settings?

We confirm that safety settings were disabled in the original experiments to obtain decisions on moral dilemmas involving harm. The original Gemini 2.5 Flash refusal rates were reported in Appendix E.2, Figure A2. To address the reviewer's question, we repeated the comparison under default safety settings using the same-family gemini-3.5-flash-lite, since the original Gemini 2.5 Flash endpoint was unavailable (Google's official notice). The follow-up covers 200 Single Feature and 100 Interaction configurations with MINIMAL thinking.

The refusal rates are compared below. Follow-up rates are design-weighted.

Model / safety settingSubsetText (%)Caption (%)Image (%)
Gemini 2.5 Flash / disabled (original)Single Feature1.061.030.51
Gemini 3.5 Flash-Lite / defaultSingle Feature36.4738.688.24
Gemini 2.5 Flash / disabled (original)Interaction0.550.910.06
Gemini 3.5 Flash-Lite / defaultInteraction82.1691.9934.22

Although refusal rates are higher in the follow-up model under default safety settings, Image still has substantially lower refusal rates than both Text and Caption in the two subsets. The reduction therefore persists under default safety settings in this same-family follow-up. Please see Appendix E.12, “Gemini thinking and safety controls,” for the configuration comparison and Appendix E.2, Figure A2 for the original results.

W3

The dual-process explanation (System 1 vs. System 2) lacks direct evidence. The argument mainly relies on an analogy and a single Grad-CAM example. The appendix also acknowledges that the ablations do not identify the underlying mechanism.

We would like to clarify that dual-process theory provides a phenomenological framework for interpreting the observed behavioral shifts. The Introduction frames the analogy conditionally: “If VLMs exhibit a similar pattern.” Section 4.4 introduces the explanations with “We hypothesize,” and Appendix A states that the study “primarily remains at the phenomenological level of behavioral analysis.” Our contribution is the controlled benchmark and the systematic characterization of modality-dependent moral decisions. Appendix E.6 provides preliminary attribution evidence to inform the discussion. These existing qualifications distinguish the behavioral findings from the further mechanistic investigation identified as future work.

Reviewer_AJhAOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

At present, the moral alignment of the visual-language model is mainly trained on text data, but visual input will significantly change the moral judgment made by the model. The MDS moral dilemma simulation benchmark built by this work can orthogonally control the two variables of concepts and roles in the moral scenario.It can be seen from the results of the three groups of comparative experiments: text, description and images that visual input will weaken the utilitarian trade-off ability of the model, and will also invalidate the judgment of the social value hierarchy, exposing a significant modal gap. This also shows that it is obviously insufficient to do only the alignment at the text level, and it is necessary to further promote the relevant research on security alignment under multi-modal conditions.

Critical strengths
  1. The generator is the main contribution of this article. Ortho-controlled manipulation, well-found true values and public code make it possible to match and compare the same dilemma between different input formats, which is not available in existing image-based ethical benchmarks.
  2. The appendix is detailed and frank: rendering review of the full data set, rewriting fidelity check, OCR/VQA check, and Image+Text dissipation experiment with appropriate qualification instructions. These allow the evidence to be verified.
Critical weaknesses
  1. The conclusion that "visual input amplifies demographic bias" (Section 4.3) is a comparison between Image and Caption. What if it is compared with Text?
  2. Gemini's thinking setting is not reported. This confusing factor weakens the paper's explanation of Gemini, and also weakens the conclusion that "modal gap is a solvable alignment problem".
Questions for authors
  1. Regarding W1: Please report the comparison between Text and Image of Character and Quantity in Figure 8, as well as the unnatified SHAP values.
  2. Regarding W2: Does Gemini-2.5-flash runtime open thinking? If it is turned on, will there be a modal gap after lowering (or turning off when feasible) thinking?
3 comments
AuthorsOpenReview ↗

Text–Image SHAP comparison and absolute feature contributions

We appreciate the reviewer's recognition of our controllable generator and the detailed, verifiable evidence in the appendix. We address the requests for direct Text–Image SHAP comparisons and Gemini thinking controls below.

W1

The conclusion that "visual input amplifies demographic bias" (Section 4.3) is a comparison between Image and Caption. What if it is compared with Text?

Q1

Regarding W1: Please report the comparison between Text and Image of Character and Quantity in Figure 8, as well as the unnatified SHAP values.

Our modality comparison is Caption–Image (Sec. 3.3 and Figure 4). Both start from the same rendered scene. Caption Mode converts the scene into a linguistic description and extracts its embedded text, whereas Image Mode processes the image directly. This comparison addresses our research question of how decisions change when the same rendered dilemma is processed through language or vision. Text Mode provides an additional structured-description reference. We therefore use Caption–Image for the modality analysis and also report the direct Text–Image comparison requested here.

Figure 8 already displays all three modes, with the corresponding composition and interaction-intensity tables in Appendix E.5. We use normalized contributions because our analysis in Sec. 4.3 concerns how decisions distribute their relative reliance among numerical trade-offs, character attributes and action bias. Expressing the three contributions as percentages of their total directly captures this balance across input modes.

The surrogates are fitted separately for each mode, so cross-modal comparisons of raw SHAP magnitudes require consideration of surrogate fidelity and the evaluated samples. Normalized contributions are more appropriate for our comparison of relative reliance. As in the analysis in Sec. 4.3, Image assigns a larger proportion of decision contributions to Character than Caption in five of the six models. This indicates that decisions place relatively greater weight on character attributes under direct visual input. The requested Text–Image comparison is:

ModelQuantity (%), Text→ImageCharacter (%), Text→Image
LLaVA-v1.6-34B4.08→14.9394.49→80.62
Qwen3-VL-8B11.30→3.6678.68→94.52
Qwen3-VL-32B5.17→17.0093.33→80.61
LLaMA-3.2-90B5.63→7.0087.79→90.23
GPT-4o-mini7.81→10.9687.69→83.44
Gemini-2.5-Flash36.17→39.2118.71→16.16

Relative to Text, Qwen8 and LLaMA allocate a larger proportion of decision contributions to Character in Image, whereas the other four models allocate a smaller proportion. Thus, the Text–Image comparison also reveals changes in the balance between numerical trade-offs and character-based preferences, with the direction varying across models. The Caption–Image result concerns the matched comparison of linguistic and visual representations of the same rendered scene.

We have also added the unnormalized SHAP values for all six models and three modes in the unnormalized-contributions table in Appendix E.5 of the revised PDF.

AuthorsOpenReview ↗

Gemini thinking configuration

W2

Gemini's thinking setting is not reported. This confusing factor weakens the paper's explanation of Gemini, and also weakens the conclusion that "modal gap is a solvable alignment problem".

Q2

Regarding W2: Does Gemini-2.5-flash runtime open thinking? If it is turned on, will there be a modal gap after lowering (or turning off when feasible) thinking?

We thank the reviewer for highlighting thinking as a potentially important factor. We did not explicitly configure thinking in the original Gemini 2.5 Flash experiments and used its default thinking configuration. Google's thinking guide lists thinking as enabled by default for Gemini 2.5 Flash. We have added the original thinking configuration to Appendix D.3.5.

To address whether a modality gap remains at low thinking, we evaluated the same-family Gemini 3.5 Flash-Lite at MINIMAL, its lowest supported thinking level. Google lists the original Gemini 2.5 Flash as deprecated with access limited to prior active users (official notice), and it was unavailable to our follow-up project. The new evaluation covers all 2105 Quantity configurations in Text, Caption and Image modes. For this model, Default also corresponds to MINIMAL according to Google's thinking guide.

Following the Quantity sensitivity analysis in Sec. 4.1, the fitted maximum sensitivity is:

Model / thinkingText Smax⁡S_{\max}Caption Smax⁡S_{\max}Image Smax⁡S_{\max}
Gemini 3.5 Flash-Lite / MINIMAL0.64701.07650.6373

Image has lower sensitivity to numerical trade-offs than Caption even at the lowest thinking level. Thus, modality-dependent moral decisions persist under minimal thinking in this same-family model. Please see Appendix E.12, “Gemini thinking and safety controls,” of the revised PDF for the configuration and full Quantity results.

We also compared Default (MINIMAL) and HIGH on the same 36 configurations spanning Quantity, Single Feature and Interaction. Among configurations with binary decisions under both thinking settings and both compared modes, Caption–Image agreement increases from 83.33% to 100% in Quantity and from 60% to 80% in Single Feature. Text–Image agreement instead decreases from 100% to 83.33% and from 83.33% to 75%, respectively. Thus, higher thinking makes Caption and Image decisions more aligned in this limited sample, while the effect varies across modality pairs.

These paired results support the reviewer's suggestion that greater thinking can improve Caption–Image alignment. The relatively aligned profile of Gemini 2.5 under default adaptive thinking and the remaining Caption–Image difference in Gemini 3.5 at MINIMAL are qualitatively consistent with this interpretation. Thinking is therefore a plausible contributor to the original Gemini 2.5 profile, with the same-model MINIMAL–HIGH comparison providing the more direct evidence. These preliminary results also complement our discussion in Sec. 4.4 that modality gaps can be mitigated, and suggest inference-time thinking as a possible way to improve Caption–Image consistency in the tested configurations.

Due to the limited time available during rebuttal, we could not extend the HIGH condition to the full experimental grid. A larger paired evaluation is a direction for future work. The full MINIMAL results above establish the observed low-thinking modality difference, while the smaller comparison provides preliminary evidence that thinking can affect alignment.

AuthorsOpenReview ↗

Seeking your feedback on our rebuttal

We sincerely thank you for your constructive feedback and encouraging assessment of our work. We have carefully addressed both of your main concerns:

  • SHAP analysis: Added direct Text–Image comparisons and unnormalized SHAP values, clarifying model-dependent feature contributions (Appendix E.5).
  • Gemini thinking: Clarified the original configuration and conducted additional thinking-level controls, demonstrating its influence on cross-modal consistency (Appendix E.12).

We hope these additions have resolved your concerns and further strengthened your assessment of the paper. If you find our responses satisfactory, we would be very grateful if you could consider updating your rating accordingly.

We would also welcome any further questions or suggestions. Thank you again for your time and consideration.

Reviewer_cGaDOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

In this paper, the authors investigate whether visual inputs affect moral decision-making in state-of-the-art VLMs and whether text-based safety alignment generalizes to the visual modality. Specifically, they introduce the MDS benchmark, which is grounded in Moral Foundation Theory (MFT) and allows controlled manipulation of conceptual and character variables. The benchmark evaluates multiple VLMs under three conditions: text, caption, and image. The results show differences in models' moral decision-making across these conditions, particularly when visual inputs are used. Based on these findings, the authors argue that language-centered safety alignment does not reliably transfer to visual processing.

Critical strengths
  1. The benchmark design allows easy manipulation of conceptual and character variables. This controlled design is useful for systematically analyzing how different factors influence moral decision-making across modalities.

  2. The evaluation is extensive, covering multiple input conditions, ablations, and supporting analyses in the appendix. The additional experiments help better understand the factors contributing to the observed differences across modalities.

Critical weaknesses
  1. Although the paper motivates the benchmark through embodied agents and safety-critical real-world deployment, the main evaluation relies on synthetic pixel-art scenes with embedded text. Without broader validation on real-world images or realistic embodied scenarios, it is difficult to determine whether the observed "visual distraction" reflects a meaningful and generalizable multimodal safety failure. The limited style-transfer analysis is too small and inconsistent to close this gap. Furthermore, the 84K samples are isolated dilemmas rather than consecutive observations from evolving agent trajectories, further limiting the applicability of the findings to embodied-agent settings.

  2. The paper does not provide a human moral-judgment baseline. Specifically, the human study only verifies whether people can identify content from the images, but does not compare model decisions against human moral judgments on the same dilemmas. The additional experiments strengthen the analysis of cross-modal consistency; however, it remains unclear whether the observed differences represent degraded moral reasoning or simply changes in model behavior across input modalities. This distinction is important because the paper draws broader conclusions about moral reasoning and multimodal safety alignment that are not fully supported by the current evaluation.

Questions for authors
  1. How well do the findings in the paper generalize beyond synthetic pixel-art scenes to real-world images or realistic embodied-agent settings? The paper provides a style-transfer analysis, but it is limited to five representative dilemmas; what evidence supports the broader implications for embodied agents?

  2. The paper does not provide a human moral-judgment baseline. Without comparing model decisions against human moral judgments on the same dilemmas, how can the authors establish that the observed cross-modal differences represent degraded moral reasoning or alignment instead of just behavioral inconsistencies across input modalities?

  3. In Section D.1.1, the caption-generation prompt explicitly tells the model that the image contains a moral dilemma. What happens if this instruction is removed? Could this explicit guidance influence the comparison between Caption and Image Modes and, consequently, the observed modality gap?

  4. I'm wondering whether the paper's findings still hold for more recent frontier VLMs. Several models evaluated in the paper are relatively old. Would the same modality gap persist in newer models with stronger multimodal reasoning and safety alignment?

10 comments
AuthorsOpenReview ↗

Realistic imagery, rendering style, and external validity

We thank the reviewer for recognizing the value of MDS's controlled design and extensive evaluation. We address the constructive suggestions on realism, human baselines, caption framing, and newer models below.

W1

Although the paper motivates the benchmark through embodied agents and safety-critical real-world deployment, the main evaluation relies on synthetic pixel-art scenes with embedded text. Without broader validation on real-world images or realistic embodied scenarios, it is difficult to determine whether the observed "visual distraction" reflects a meaningful and generalizable multimodal safety failure. The limited style-transfer analysis is too small and inconsistent to close this gap.

Q1

How well do the findings in the paper generalize beyond synthetic pixel-art scenes to real-world images or realistic embodied-agent settings? The paper provides a style-transfer analysis, but it is limited to five representative dilemmas; what evidence supports the broader implications for embodied agents?

MDS is designed for matched modality comparisons under controlled dilemma configurations. Real photographs combine scene content, character appearance, lighting, viewpoint, and background, making it difficult to vary one moral variable while holding the others fixed. MDS makes quantities, character attributes, and causal constraints independently adjustable, so the same configuration can be compared across different input forms. This controlled design complements real-image moral evaluation and perturbation studies such as MORALISE [4] and Liu et al. [5], which we have added to the related-work discussion in Sec. 2.2. It also follows established experimental practice: The Moral Machine experiment (Nature, 2018) uses pictorial moral dilemmas [1], CLEVR uses synthetic scenes for visual reasoning diagnostics [2], and virtual moral-action studies use simulated dilemmas to study behavior [3].

The existing style experiment evaluates five representative dilemmas across nine visual styles, including realistic-photo rendering, oil painting, and sketches (Appendix F.5, Figure A13 and Table A15). Across these examples, model decisions vary with style, while realistic rendering does not consistently restore the Text-Mode response pattern. The observed differences therefore extend beyond the original pixel-art presentation in these tested scenes. The 2047-pair perception-oriented VQA evaluation (Appendix E.7) and the human visual-identification study (14 participants, 95.71% accuracy, Appendix F.9) additionally support the recognizability of the rendered content. The full-dataset renderer audit reports 80,297/84,240 configurations jointly passing the structural checks (Appendix F.10).

The controlled construction therefore serves our research objective of comparing moral decisions across input forms while independently varying specific dilemma variables. The validation results show that the differences in the tested scenes are not attributable solely to the original pixel-art style, and support the use of recognizable, structurally consistent stimuli for this controlled analysis.

[1] Awad, Edmond, et al. "The moral machine experiment." Nature 563.7729 (2018): 59-64.

[2] Johnson, Justin, et al. "Clevr: A diagnostic dataset for compositional language and elementary visual reasoning." Proceedings of the IEEE conference on computer vision and pattern recognition. 2017.

[3] Francis, Kathryn B., et al. "Simulating moral actions: An investigation of personal force in virtual moral dilemmas." Scientific Reports 7.1 (2017): 13954.

[4] Lin, Xiao, et al. "Moralise: A structured benchmark for moral alignment in visual language models." In Proceedings of International Conference on Machine Learning (ICML), 2026.

[5] Liu, Zhining, et al. "Do VLMs Have a Moral Backbone? A Study on the Fragile Morality of Vision-Language Models." Findings of the Association for Computational Linguistics: ACL 2026. 2026.

AuthorsOpenReview ↗

Embodied trajectories

W1

Furthermore, the 84K samples are isolated dilemmas rather than consecutive observations from evolving agent trajectories, further limiting the applicability of the findings to embodied-agent settings.

The MDS configurations are independent, single-step dilemmas. This design isolates changes in quantities, character attributes, and input representation from additional variation introduced by observation history, action feedback, and planning. A controllable benchmark together with the systematic characterization of this single-step phenomenon constitutes a substantive contribution in its own right.

Multi-step agents must also make individual decisions from visual observations, so the representation sensitivity identified here is already a reliability issue they need to address. A longer trajectory does not by itself resolve this decision-level vulnerability. It introduces additional interactions with observation history, planning, and action feedback, making the behavior of this vulnerability over time a natural next research question. Appendix A already discusses dynamic visual settings and temporal evaluation as future work. The programmatic generation design described in Appendix F.2 provides a basis for extending MDS to such settings.

AuthorsOpenReview ↗

Moral framing in the Caption prompt

Q3

In Section D.1.1, the caption-generation prompt explicitly tells the model that the image contains a moral dilemma. What happens if this instruction is removed? Could this explicit guidance influence the comparison between Caption and Image Modes and, consequently, the observed modality gap?

Thank you for raising this question. We removed the explicit moral framing from the Caption prompt:

Original wordingNeutral wording
“moral dilemma”“scenario”
“dilemma mechanisms, consequences and related moral conflict”“action mechanisms, consequences and related conflict”

The remaining caption instructions, OCR extraction and downstream Yes/No prompt are unchanged. We compare Neutral Caption and Image following the analyses in Secs. 4.1–4.2, using each mode's valid outputs on the same prespecified configurations. For Qwen8 and Qwen32, we reuse the original Image decisions for all 2105 Quantity and the 4941 prespecified Single Feature configurations. The Image Quantity values below reproduce those reported in the submitted paper. For LLaVA, both modes use the newly evaluated subset of 236 Quantity and 264 Single Feature configurations with matched inference settings.

Quantity sensitivity (Sec. 4.1). The maximum fitted slope is:

ModelNeutral Smax⁡S_{\max}Image Smax⁡S_{\max}
Qwen3-VL-8B0.65250.2945
Qwen3-VL-32B0.71400.2628
LLaVA-v1.6-34B (subset)0.4553≈0*

Removing moral framing retains reduced sensitivity to utilitarian trade-offs in Image. For Qwen8 and Qwen32, changes in the numbers of affected individuals have less influence on Image decisions than on Neutral Caption decisions. *LLaVA Image remains near the ceiling across quantity ratios, with little empirical response to the numerical trade-off and a near-zero fitted slope (R2≈0R^2\approx0).

Single Feature results (Sec. 4.2). The conceptual-factor analysis gives the following log-odds:

ModelHarm: NeutralHarm: ImageSelf Benefit: NeutralSelf Benefit: Image
Qwen3-VL-8B−0.3047−0.4358+0.2208+0.1910
Qwen3-VL-32B−0.2367−0.3398+0.0774+0.1171
LLaVA-v1.6-34B (subset)−0.3130−1.1981+0.7044−0.7775

The more negative Image coefficients indicate stronger inhibition against intentionally causing harm. Qwen32 becomes more permissive toward self-benefit in Image, while LLaVA changes from favoring to avoiding it. These estimates show changes in the strength and direction of moral constraints after explicit moral framing is removed.

Moral-foundation and character preferences also change. In the LLaVA subset, Care preference increases from 0.6000 in Neutral Caption to 0.9730 in Image, while Fairness decreases from 0.7400 to 0.1800. Image thus places greater weight on Care and less on Fairness. Qwen32's Species category gap changes from +0.5605 to −0.1140 under the signed log-odds analysis in Figure 7(b), reversing its species preference.

The Neutral Caption–Image comparison therefore retains reduced utilitarian sensitivity and modality-dependent moral preferences. We have added this control and its results to Appendix E.9, “Caption framing control,” of the revised PDF.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Overall response

We thank all three reviewers for recognizing MDS's controlled generation, cross-modal evaluation, and validation. We have replied individually and updated the manuscript. Our follow-up asks whether the observed differences extend beyond experimental artifacts, and what the evidence can establish.

1. An independent human reference (cGaD, RDhL). Text Mode measures within-model consistency, not moral truth. We therefore collected 4800 judgments from 400 participants (200 per Text/Image condition) on 41 matched stimuli from 12 dilemmas (Appendix E.13). Human numerical sensitivity was similar across conditions (Smax⁡=0.6555S_{\max}=0.6555 in Text and 0.66080.6608 in Image), with the same moral-foundation ordering; the task-averaged Image–Text Yes-rate difference was −1.10 percentage points (95% CI [−3.44, 1.08]), although Image responses took 28.2% longer. On those same stimuli, all six original VLMs showed lower agreement with human responses in Image than in Text, by 2.57–18.21 percentage points. In one high-consensus dilemma, 95% of participants in both modes rejected saving one neighbor at the cost of ten babies; all six models rejected this action in Text, but four endorsed it in Image.

2. Perception and representation controls (RDhL, cGaD). Our analyses distinguish the Text–Caption context gap from the Caption–Image comparison of linguistic and direct visual processing (Sec. 3.3). Replacing only the predicted OCR transcript with the exact proposition changes LLaVA-v1.6-34B's Quantity-curve MAE to Text from 0.2886 to 0.0570 (Appendix E.3, Table A3). This affects the context gap, not the Caption–Image gap or an isolated moral-reasoning mechanism. Effects are smaller for Qwen and slightly adverse for LLaMA. With propositions fixed, model-generated, configuration-grounded, and external captions change decisions non-monotonically; external captions do not uniformly recover Text behavior. In the separate Image+Text control (scene pixels plus the exact proposition as text), Quantity MAE moves closer to Text in three of four open-weight models (LLaVA 0.8219→0.1015) but farther in LLaMA (0.1766→0.2914), with residual differences in Single Feature and Interaction (Appendices E.3–E.5). No single factor—OCR, scene description, proposition presentation, or visual–text integration—explains all models.

3. Robustness to prompts, precision, and newer models (cGaD, AJhA, RDhL). A neutral Caption prompt preserves modality-dependent quantity sensitivity and preference shifts (Appendix E.9). Matched NF4/BF16 experiments for Qwen3-VL-32B and LLaVA-v1.6-34B retain cross-modal differences at BF16 (Appendix E.10). GPT-5.6 Luna and Sol, tested on 2105 Quantity and 4941 Single Feature configurations in all three modes, also exhibit decision shifts, including greater Self-Benefit permissibility in Image (Appendix E.11). We documented Gemini's original thinking configuration; same-family Gemini 3.5 Flash-Lite follow-ups probe thinking and default safety settings, with remaining modality differences (Appendix E.12). Direct Text–Image and unnormalized SHAP comparisons also show model-dependent character contributions (Appendix E.5).

4. Validation and scope (cGaD, RDhL). We structurally audited all 84,240 samples: 80,297 (95.32%) passed all checks; 3922 trolley samples had partially clipped but visible sprites, and 21 shark-attack samples had negative diver counts (Appendix F.10). These structural checks complement the 2047-question VQA and 14-person legibility study. The nine-style test on five dilemmas suggests the pixel-art aesthetic is not the sole source (Appendix F.5). MDS is a controlled diagnostic benchmark, not a deployment simulator.

What the paper establishes. MDS provides a controllable benchmark and reveals a key vulnerability: the same moral dilemmas can elicit substantially different decision patterns across modalities, affecting numerical sensitivity, moral constraints, and character preferences. These differences persist under targeted controls and can reduce agreement with human judgments.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.