MM-SCALE: Grounded Multimodal Moral Reasoning via Scalar Judgment and Listwise Alignment
arXiv preprint, 2026 · Carnegie Mellon University
Eunkyu Park
Wesley Hanwen Deng
Cheyon Jin
Matheus Kunzler Maldaner
Jordan Wheeler
Jason I. Hong
Hong Shen
Adam Perer
Ken Holstein
Motahhare Eslami
Gunhee Kim
What is MM-SCALE?
MM-SCALE (Multimodal Moral Scale) is a large-scale dataset for aligning vision language models with human moral preferences. It contains 32,212 annotated scenarios across 9,260 unique images, each rated for moral acceptability on a 5-point scale and labeled for whether the judgment depends on the text, the image, or both.
MM-SCALE is a dataset built to answer a specific question: can a vision language model learn to judge how morally acceptable an action is, in context, the way people actually judge it? The dataset pairs 9,260 unique images with 32,212 short written scenarios, about 3.5 per image. Human annotators rated each image-scenario pair on a 5-point moral acceptability scale and flagged which modality drove their judgment. Those ratings become training signal for aligning VLMs, models that reason over images and text together.
The project was led by Eunkyu Park, working with Gunhee Kim's group at Seoul National University alongside collaborators at Carnegie Mellon University's Human-Computer Interaction Institute. I am one of eleven authors. The paper appeared as an arXiv preprint in 2026 under the identifier 2602.03665.
Why binary labels fail for moral reasoning
Binary moral labels force every judgment into acceptable or unacceptable, which flattens the graded, pluralistic nature of human moral reasoning. MM-SCALE argues that jaywalking and serious harm should never share a label, and that models trained on collapsed categories inherit miscalibrated moral behavior.
Most prior datasets for machine moral reasoning supervise models with binary labels or pairwise comparisons. An action is acceptable or it is unacceptable. Human judgment does not work that way. People see gradations. Jaywalking on an empty street and abandoning a child both land on the unacceptable side of a binary label, yet nobody weighs them equally. Collapsing that range into two buckets throws away the information a model most needs.
The consequences show up at deployment. A model trained on binary signals learns a hard boundary and tends to be poorly calibrated near it. Moral judgment is also pluralistic, since reasonable annotators disagree, and a scalar scale lets that disagreement register as a distribution instead of vanishing into a majority vote. Binary labels hide the texture. Scalar ratings keep it.
Scalar ratings and modality grounding
Every MM-SCALE scenario carries a 5-point Likert rating, from 1 (very unacceptable) to 5 (very acceptable), assigned by three independent annotators. Each judgment also carries a modality grounding label recording whether the decision rested on the written scenario, the visual scene, or the interaction between the two.
Each image-scenario pair in MM-SCALE was rated by three independent annotators on a 5-point Likert scale, where 1 means very unacceptable and 5 means very acceptable. Averaging across annotators gives every scenario a continuous moral acceptability score, the basis for ranking-based training later in the pipeline.
The second annotation layer is modality grounding. For every judgment, annotators recorded whether their decision was based mainly on the written scenario, mainly on visual cues in the image, or on how the two interact. This matters because the same sentence can flip its moral meaning depending on the picture it sits beside. Grounding labels make that dependency explicit, so the dataset records what people decided and also where they looked to decide it.
Interactive demo: rate like an annotator
MM-SCALE annotators saw an image and a one sentence scenario, scored moral acceptability from 1 to 5, and flagged which modality drove the call. Try the same task on three stylized scenes.
Jaywalking across an empty street late at night.
Taking the last open seat on a crowded bus while an elderly passenger stands.
Photographing a stranger on the street for an art project without asking.
Listwise alignment for vision language models
Scalar scores let MM-SCALE rank all scenarios attached to one image and train models with listwise preference optimization, learning from ranked sets rather than isolated pairs. The framework uses a ListMLE loss that pushes model scores to reproduce the full human ordering.
Because every scenario has a continuous score, the scenarios grouped under a single image can be sorted into a full ranking. That unlocks listwise preference optimization, training on ranked sets rather than isolated pairs. MM-SCALE uses a ListMLE loss, which encourages the model to assign scores whose resulting order matches the human ordering across the whole list, from most acceptable to least.
The distinction from pairwise methods is practical. A pairwise objective only teaches the model that A beats B, one comparison at a time. A listwise objective hands the model the entire preference structure over each image at once, and MM-SCALE's grouped scenarios supply exactly the ranked lists this training needs. Richer signal in, finer calibration out.
Interactive demo: rank a list the way ListMLE trains
MM-SCALE groups about 3.5 scenarios under each image, and their averaged scalar scores sort them into a full ranking. Here is one stylized park scene with four scenarios. Order them from most to least acceptable, then see what the training signal keeps.
Your ranking
What supervision keeps
A binary label only remembers which side of the line each scenario fell on: two buckets, no order inside them. A pairwise objective restores one comparison at a time. The full ranking hands the model the entire preference structure over the image at once, and that is exactly what the ListMLE loss in MM-SCALE trains on.
The averaged scores here are illustrative, chosen to show the mechanism. In the real dataset each scenario score is the mean of three annotator ratings on the 1 to 5 scale.
What the experiments show
Four open vision language models fine-tuned on MM-SCALE with listwise optimization beat binary-supervised baselines on both ranking fidelity and safety calibration. LLaVA-OneVision-7B reached an NDCG@5 of 0.89 and an MRR of 0.58, ahead of binary preference optimization at 0.85 and 0.52 respectively.
The team fine-tuned four open vision language models on MM-SCALE: LLaVA-OneVision, Qwen2-VL-7B-Instruct, Phi-3-Vision-Instruct, and InstructBLIP. Each was trained under three supervision styles, listwise optimization on the scalar rankings, binary preference optimization, and binary cross-entropy classification, then evaluated on how faithfully it reproduced human moral orderings and how safely it was calibrated.
Listwise training won on ranking fidelity. On LLaVA-OneVision-7B it reached an NDCG@5 of 0.89, a measure of how well top-ranked scenarios match human rankings, against 0.85 for the binary preference baseline, with mean reciprocal rank improving from 0.52 to 0.58. Safety calibration followed the same pattern, AUC-Safety rising to 0.76 from 0.72, with binary cross-entropy trailing well behind both. Models given the full scalar signal judge morally ambiguous scenes more like people do.
My role in MM-SCALE
I am the fourth of eleven authors on MM-SCALE. The paper was led by first author Eunkyu Park with Gunhee Kim as senior author, and the author list includes Wesley Hanwen Deng, Dr. Jason Hong, Dr. Ken Holstein, and Dr. Motahhare Eslami, my co-authors on MIRAGE and Seeing Twice.
MM-SCALE was led by first author Eunkyu Park, with Gunhee Kim as the senior author. I am the fourth of eleven authors. Four of my co-authors here, Wesley Hanwen Deng, Dr. Jason Hong, Dr. Ken Holstein, and Dr. Motahhare Eslami, are the same researchers I published MIRAGE and Seeing Twice with, both papers about how people audit and judge generative AI systems. MM-SCALE sits in that same research line, this time collecting graded human moral judgments at scale rather than audits of image generators.
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.
Frequently Asked Questions
Where can I access the MM-SCALE dataset and paper?
The paper is available as an arXiv preprint at arxiv.org/abs/2602.03665, with the PDF linked from the abstract page. The authors state they plan to open-source MORALE, the annotation interface used to collect the data. As a 2026 preprint, the work had not yet completed peer review at the time of posting.
How were the MM-SCALE annotations collected?
Annotations came from crowd workers recruited through Amazon Mechanical Turk in the United States and Great Britain. Every image-scenario pair was rated by three independent annotators, each assigning the 1-to-5 acceptability score and the modality grounding label. Averaging the three ratings produces the continuous score used to build the ranked training lists.
What is the MORALE annotation interface?
MORALE (MORal Alignment and Listwise Evaluation) is the web interface the team tailored for data collection. Beyond scalar rating and modality selection, it runs a model in the loop: a prediction diverging from the human rating by a point or more triggers a modality check, while agreement prompts the annotator to add a scenario for the same image, growing the dataset as it is validated.
Which vision language models were fine-tuned on MM-SCALE?
Four open models were fine-tuned and evaluated: LLaVA-OneVision (7B), Qwen2-VL-7B-Instruct, Phi-3-Vision-Instruct, and InstructBLIP. Testing across architectures from different labs shows the listwise advantage is a property of the supervision signal itself, since the ranking and safety gains held across all four rather than depending on any single model family.
How does MM-SCALE differ from other moral reasoning benchmarks?
Most moral reasoning benchmarks are text-only and use binary or pairwise labels. MM-SCALE combines three things in one resource: multimodal scenarios where an image changes the moral reading, continuous 5-point scores that preserve severity gradations, and modality grounding labels recording which input drove each judgment. Together these support listwise training that text-only or binary datasets cannot.
BibTeX
@misc{park2026mmscale,
title={MM-SCALE: Grounded Multimodal Moral Reasoning via Scalar Judgment and Listwise Alignment},
author={Eunkyu Park and Wesley Hanwen Deng and Cheyon Jin and Matheus Kunzler Maldaner and Jordan Wheeler and Jason I. Hong and Hong Shen and Adam Perer and Ken Holstein and Motahhare Eslami and Gunhee Kim},
year={2026},
eprint={2602.03665},
archivePrefix={arXiv},
primaryClass={cs.CV}
}