MM-SCALE: Grounded Multimodal Moral Reasoning via Scalar Judgment and Listwise Alignment

TL;DR

MM-SCALE addresses the challenge of aligning vision-language models with human moral preferences in multimodal, context-rich scenarios by collecting 32,212 scalar, grounded moral judgments and training with listwise optimization via ListMLE. The dataset provides explicit modality grounding (text, image, or both) enabling analysis of how different modalities influence moral judgments and calibration. Through experiments across multiple VLM backbones, scalar listwise supervision achieves superior ranking fidelity (e.g., high $NDCG@5$ and $MRR$) and more stable safety calibration ($AUC-Safety$) than binary supervision. The work demonstrates that multimodal grounding and scalar supervision yield finer-grained, more robust moral alignment, with an interactive annotation loop (MORALE) to expand coverage and reflect human disagreement. This advances practical safety and ethics in multimodal AI by enabling context-aware, calibrated moral judgments in real-world settings.

Abstract

Vision-Language Models (VLMs) continue to struggle to make morally salient judgments in multimodal and socially ambiguous contexts. Prior works typically rely on binary or pairwise supervision, which often fail to capture the continuous and pluralistic nature of human moral reasoning. We present MM-SCALE (Multimodal Moral Scale), a large-scale dataset for aligning VLMs with human moral preferences through 5-point scalar ratings and explicit modality grounding. Each image-scenario pair is annotated with moral acceptability scores and grounded reasoning labels by humans using an interface we tailored for data collection, enabling listwise preference optimization over ranked scenario sets. By moving from discrete to scalar supervision, our framework provides richer alignment signals and finer calibration of multimodal moral reasoning. Experiments show that VLMs fine-tuned on MM-SCALE achieve higher ranking fidelity and more stable safety calibration than those trained with binary signals.

Figures (15)

• Figure 1: Comparison between existing benchmarks and MM--Scale. (a) SPA-VL zhang2024spavl provides binary preference labels between two model responses, and (b) VLGuard zong2024safetyfinetuningalmostcost classifies a model output as safe or unsafe. (c) MM--Scale presents human--authored scenarios grounded in an image, each labeled with a scalar moral judgment and an attribute indicating on what modality the judgment is grounded. In (c), “Drinking liquid that you...” is text--grounded, since the moral meaning is conveyed entirely by the wording, while “Getting someone water...” is image+text--grounded, as both the situation of helping gesture and textual intent jointly inform the judgment. “Drink something to cool off...” is image--grounded, since the arbitrary something is specified with the visual context.
• Figure 2: Overview of our data annotation pipeline. (a) Situations Sourcing: We source daily norm scenarios that can add details to an action from the Commonsene Normbank jiang2022machineslearnmoralitydelphi dataset. (b) Multimodal Moral Context Generation: A commonsense-based target setting (e.g., “Child crossing a street”) is selected and rendered into a visual scene using a text-to-image (T2I) model. (c) Moral Judgment Annotations: Annotators evaluate multiple moral scenarios grounded in the image using scalar judgments (1-5) and indicate the grounding modality (text, image, or both).
• Figure 3: Comparison of alignment metrics between synthetic images and caption-matched real images from Visual Genome. Differences ($\Delta \leq 0.02$) are trivial across NDCG@5, Unsafe Rate, and Kendall’s $\tau$ metrics.
• Figure 4: The interactive annotation loop in MORALE. For each image--scenario pair, the system compares the annotator’s score with the model’s prediction. Disagreement triggers a modality--grounding check, while agreement prompts the annotator to add a new, image--grounded scenario. See §\ref{['sec:annotation-interface']} for details.
• Figure 5: Grounding modality distribution of human moral ratings in MM--Scale. The left bar chart shows the distribution of scalar ratings (1–5) grounded in text, image, or both modalities. The pie chart shows the proportion of modality reliance.