acceptodds
Under review as a conference paper at ICLR 2027

Synth-Ethics: Synthetic Training Data for LLM Constitutions Grounded in Ethical Traditions

Abstract

When a language model stands in for a person, as a simulated user, an annotator, or a judge, it enacts a moral outlook that nobody chose: whatever its training distribution happened to contain, or whatever a prompt can induce. Even when the intended outlook is written down, as with Claude's constitution, what the trained model enacts on a given dilemma stays unknown. We cannot measure it either: there is no dataset of what a utilitarian, a Kantian, a virtue ethicist, and so on would conclude on the same dilemmas. SYNTH-ETHICS provides, to our knowledge, the first such dataset. We operationalise 16 frameworks from six ethical traditions as framework cards and resolve 4,397 dilemma scenarios under all 16, giving 68,520 quality-filtered reasoning traces with structured situations and answer metadata, in the style of an instruction-tuning corpus. Fine-tuned on one framework's traces, models agree with that framework on held-out scenarios at a mean Cohen's K of 0.42, against 0.29 for instruction-tuned models given the framework as a system prompt. Asked without a named theory, they score highest on their own framework's MoReBench rubric, with answers a fraction of the instruct models' length. Scored against the same references, 27 API models from 12 labs, among them GPT, Claude and Gemini, agree most with virtue ethics, and no framework separates US, Chinese and European labs. We release the 16 models, which serve as simulated users, annotators, or judges with a documented outlook, and the references, which report the moral outlook of any other model. Since the references share the format of instruction-tuning data, the same records can enter a model at any training stage and be scored the same way afterwards; where an outlook enters, and how much of it survives later training, become measurable.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
synthetic data generation, pluralistic alignment, moral reasoning, language models, supervised finetuning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_VNVeOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces Synth-Ethics, a synthetic dataset of 68,520 reasoning traces covering 4,397 moral dilemmas and 16 ethical frameworks. Each framework is operationalized through an expert-authored framework card that guides the generation of framework specific reasoning and verdicts. The authors fine-tune nine base models from three families and use MoReBench to evaluate framework agreement, ood behavior, and reasoning adherence. They find that fine-tuning aligns models with the target framework more closely than prompting does, and that fine-tuned models exhibit framework specific reasoning without explicit instructions. The authors also use the dataset to profile the moral judgments of 27 API models and find a general tendency toward virtue ethics.

Critical strengths
  1. The paper moves beyond crowd sourced moral judgments by constructing parallel reasoning traces for 16 explicit ethical frameworks. The expert authored framework cards and controlled causal structure provide a systematic foundation for studying differences in moral judgments. This is a meaningful contribution to pluralistic alignment and moral evaluation.

  2. The evaluation spans nine base models, 144 single framework fine-tunes, and independent benchmarks. The MoReBench results are particularly interesting: fine-tuned models consistently score highest on their target framework's rubric even without explicit framework instructions. This provides evidence that framework-specific behavior can be learned through training rather than induced solely through prompting.

Critical weaknesses
  1. The same LLM-generated framework resolutions serve as both training targets and evaluation references. Although the authors conduct human evaluation on 800 traces, this primarily assesses whether the prescribed reasoning procedures are followed, rather than independently validating the moral verdicts. The ratings are also heavily concentrated near the ceiling (Appendix C). Consequently, agreement with these references may reflect successful imitation of the generator's interpretation of each framework rather than genuine adherence to the underlying ethical theory. This limitation also affects the validity of the conclusions drawn about existing API models.

  2. Unclear responses are excluded from agreement calculations, despite substantial differences in answer coverage across models (Table 6). This makes the fine-tuning versus prompting comparison potentially sensitive to selective nonresponse.

  3. The out-of-corpus evaluation reports overall action selection rates rather than independently validated framework specific judgments. Differences in these rates could reflect general response biases rather than ethical reasoning.

Questions for authors
  1. Can the authors provide independent expert validation of the reference verdicts, particularly on scenarios where the ethical frameworks disagree? How frequently do experts agree with the generated judgments, and how sensitive are the main results to disagreements?

  2. What proportion of held-out scenarios produces judgeable verdicts for each fine-tuned and prompted model? Does the fine-tuning advantage persist when Cohen's κ is computed on the same subset of scenarios or when abstentions are explicitly accounted for?

  3. Can the authors provide stronger ood evidence of framework adherence, such as testing whether models respond appropriately to controlled changes in morally relevant causal factors? The current actionnselection rates alone do not establish that the learned behavior follows the intended ethical principles.

2 comments

We thank the reviewer for the detailed review.

Presentation. In the revision, we will make the paper easier to follow. Section 4.2 will open by listing the four axes on which we evaluate:

  • Framework agreement: On our 250 held-out scenarios, how often a model's verdict agrees with each of the 16 reference corpora (Cohen's (\kappa));
  • Out-of-corpus behaviour (renamed from "Out-of-corpus agreement"): On 400 DailyDilemmas dilemmas, how often a model chooses the proposed action;
  • Reasoning adherence: On MoReBench dilemmas, how well a model's reasoning meets the criteria that experts wrote for each of four ethical frameworks;
  • API model profiles: We ask 27 API models to judge 250 held-out dilemmas without naming any framework, and compare each model's verdicts with those of the 16 frameworks to see which frameworks it agrees with most.

We will also correct several counts, the typos and the citation format where applicable.

W1 and Q1. An independent validation needs several annotators with training in moral philosophy, enough to measure agreement between experts as well as with our verdicts. The differences between the analysis procedures set out in the framework cards are sometimes subtle, and some scenarios require considerable time for reflection (15 minutes or more), especially those that involve multiple parties and the calculation of a numerical approximation of the possible costs and benefits of the outcomes, so that it could not be done within the discussion period. The revision will state this limitation in Section 7. We can still address the concern in two ways: by measuring how sensitive our results are to a different reading of the cards, and with evidence that does not depend on our references.

On sensitivity, we can measure one source of disagreement now: a second generator's (DeepSeek-V4-Pro) reading of the same cards, which reaches the same verdict as ours on 77% of scenarios. Of the 23% of scenarios where DeepSeek-V4-Pro's verdict differs, a third are cases where one model answers "it depends" (conditional). Scored against that model's verdicts instead of ours, fine-tuned models still agree with their own framework far more than prompted ones: (\kappa = 0.404), against (0.258) for prompted instruction-tuned and (0.201) for prompted base models ((0.419), (0.286) and (0.227) against our references). The references are also based on expert input: a researcher with a doctorate in moral philosophy wrote each framework card and refined it over several rounds. Each card sets out the framework's key thinkers, its standard of evaluation, how it should respond to the causal structure of a dilemma, and its known limitations (Appendix B). MoReBench, in addition, does not use our references at all, and there each of the 36 fine-tunes scores highest on its own framework's rubric (W3 below).

For the concern that "this limitation also affects the validity of the conclusions drawn about existing API models", we will phrase the API results as agreement with the card-based references. The references do not favour the model that wrote them: GLM-5.2, our generator, queried without a card, ranks 9th of 27 by mean (\kappa) against both its own references (a rerun) and DeepSeek-V4-Pro's, and DeepSeek-V4-Pro ranks 3rd against its own and 4th against GLM-5.2's.

W2 and Q2. Fine-tuned models give a clear verdict far more often than prompted ones, and fine-tuning stays ahead in both recomputations you suggest. A clear verdict is an answer the judge can label yes, no or "it depends". How often prompted models give one depends strongly on the model: from 12 to 232 of the 250 scenarios (median per model, Table 6). Fine-tuned models give a clear verdict on 183 to 247 of the 250 scenarios, and half of them on 234 or more. The revision will add these numbers to Table 6. Fine-tuned models output an unclear verdict no more often than the reference verdicts they were trained on (7.6% of scenarios, against 8.7% for the references). Both recomputations compare each fine-tuned model with the same base model prompted with the framework, in its instruction-tuned and its base version, for 9 base models (18 comparisons), and a new appendix table will report them:

  1. Same scenarios: On the scenarios where the fine-tuned model, the prompted model and the reference all give a clear verdict, the fine-tuned model agrees more with its framework in all 18 comparisons ((\kappa = 0.435) against (0.284) for instruction-tuned models, and (0.414) against (0.215) for base models).

  2. Unclear verdicts counted: When an unclear verdict counts as a disagreement, the fine-tuned model agrees more with its framework for every framework in every comparison, and it also stays ahead in all 18 comparisons when "unclear" is treated as a verdict of its own.

W3 and Q3. External datasets that judge dilemmas within named ethical frameworks are scarce: in our survey of moral-reasoning benchmarks, MoReBench is the only public dataset that scores a model's reasoning against expert criteria written for each theory. There, with no theory named, each of the 36 fine-tunes for the four frameworks it covers scores highest on its own framework's rubric (Table 11), the result you found "particularly interesting". To go further, we run the test you suggest on the causal structure of our scenarios, which we vary by design. Each framework card states in advance which causal factors should change its verdict, so a fine-tuned model that follows its framework should shift its verdicts in the direction its card states when one of these factors changes.

We use the 150 held-out constructed scenarios, which were never seen in training. Each scenario was built with known values of three causal factors. For each framework and each factor, we compare how often the fine-tuned model approves the action under one value and under the other, and we do the same for its framework's references, which a model that follows its framework should match. We will add the full results in a new appendix:

  • Harm as a means rather than a side effect: Across all 16 frameworks, the fine-tunes change in the same direction as their references in 14. For example, Kantian fine-tunes, whose card prohibits using a person merely as a means, approve the action in 53% of the scenarios where the harm is a side effect but in only 20% where it is a means, close to the Kantian references (43% and 14%). Classical-utilitarian fine-tunes, whose card counts only outcomes, approve it about equally often either way (80% and 83%). In the other two, preference utilitarianism and classical pragmatism, whose cards do not treat this distinction as morally relevant, neither the references nor the fine-tunes change by more than about 3%. By contrast, a model that is only prompted with the Kantian description shows no such drop, so the effect comes from fine-tuning.

  • Harm that could have been avoided: The fine-tunes respond to this factor as their frameworks do. When the harm could have been avoided rather than happening anyway, the references of all 16 frameworks approve the action less often, and so do all 16 fine-tunes, by about the same amount (on average from 68% to 40% for the fine-tunes, and from 62% to 38% for the references).

  • Creating a benefit rather than preventing a harm: This makes little difference even to the references (46% against 52%). This matches Fränken et al. (2024), from whom our three factors are adapted, where a related factor (acting versus failing to prevent) had no clear effect on the judgments of humans, GPT-4 or Claude-2.

To summarize, on the causal factors that matter to the frameworks, the fine-tunes respond as their framework does.

We hope that these answers and the new results address the reviewer's concerns, and we kindly ask the reviewer to reconsider the score.

Reviewer_vnZLOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces a synthetic dataset of moral dilemmas resolved under 16 ethical frameworks, grounded in expert-written framework descriptions. It shows that fine-tuning on these data produces models with distinguishable framework-specific judgments and reasoning, including on the external MoReBench evaluation.

Critical strengths
  • This is a timely and important problem, with relevance to AI safety and the alignment of the values of LLMs and AI agents.
  • The data collection and framework design seem carefully conducted, with substantive involvement from moral philosophy expertise.
  • The external validation on MoReBench is particularly valuable, providing evidence that the induced framework-specific reasoning generalizes beyond the authors’ own dataset.
  • The experiments are substantial, and the paper is well written and clear.
Critical weaknesses

No major weaknesses

Questions for authors
  • Why was GLM-5.2 chosen to generate the framework-conditioned reasoning traces? Were quality, cost, or reproducibility considerations important in this choice?
  • Do the authors think this dataset could support research on moral parliaments, where a model is aligned to a mixture of different theories of normative ethics? How might they envision using the dataset for this purpose?
1 comment

We thank the reviewer for the positive assessment. We are glad that the reviewer finds the problem timely and important, the data collection and framework design "carefully conducted", and the external validation on MoReBench "particularly valuable".

Q1. Mainly cost and reproducibility. GLM-5.2 is a strong open-weight model released under the MIT licence (GLM-5 Team, 2026). Since the weights are public, anyone can regenerate the references with the same model from the released cards and prompts; a model served only through an API can change or be withdrawn. We will state these reasons in Section 3.3.

Q2. Yes. Because the dataset resolves the same scenarios under all 16 frameworks, it gives each delegate of a moral parliament a verdict and a reasoning trace on matched examples. We see three uses:

  1. A parliament at inference time: A model trained on all 16 frameworks with framework labels (all16, Table 5) answers each scenario once per framework, and a vote combines its 16 answers. On our held-out scenarios, this vote matches the majority verdict of the 16 references on 80% of scenarios.

  2. A parliament in training: A model trained on the same traces without labels (all16u) must learn a single policy from all 16 frameworks. Its answer matches that majority on 69% of scenarios, so the frameworks combine better through an explicit vote than through training on the mixture.

  3. Weights and disagreement: Since each framework has its own corpus, researchers can give each framework a weight (its seats in the parliament) through the mix of training data, and evaluate a model on the scenarios where the parliament splits, since only there do the weights decide the outcome.

The revision will report both results in an appendix.

Reviewer_4zEeOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces Synth-Ethics, a synthetic dataset and evaluation framework for grounding language-model behavior in explicit ethical traditions. The authors operationalize 16 ethical frameworks from six traditions as framework cards and generate parallel resolutions for 4,397 moral dilemmas, yielding 68,520 quality-filtered reasoning traces. The same data can serve both as training data for inducing a documented ethical outlook and as reference data for measuring which framework another model most closely follows.

The authors fine-tune nine base checkpoints across the 16 frameworks and evaluate both verdict agreement and reasoning adherence. Fine-tuned models show substantially higher agreement with their assigned frameworks than comparable prompting baselines and generally retain framework-specific reasoning without the framework being named at inference time. The authors additionally evaluate 27 API models from 12 labs and report broad similarities in their inferred ethical profiles. Overall, the paper contributes a substantial dataset, a framework-specific model suite, and a benchmark for studying how explicit normative perspectives enter and persist in language models.

Critical strengths
  1. The paper introduces a useful and unusually systematic resource for studying normative behavior in language models. Resolving the same 4,397 scenarios under 16 explicit ethical frameworks makes it possible to compare frameworks on matched examples and, importantly, to use the same representation both for training and evaluation. This is a meaningful advance over benchmarks that provide either human moral judgments or theory-level rubrics without parallel framework-specific reference verdicts.

  2. The empirical validation is broad. The authors fine-tune models from three model families and multiple scales, compare training against prompting, evaluate out-of-corpus behavior and reasoning with MoReBench, and separately score 27 API models. I particularly value the result that framework fine-tunes continue to exhibit framework-specific reasoning when the framework is not named, since this directly tests whether the training intervention changes model behavior rather than merely improving instruction following.

Critical weaknesses
  1. The central reference labels are ultimately synthetic interpretations of the ethical frameworks rather than independently established philosophical ground truth. GLM-5.2 generates the reasoning trace and verdict for each framework/scenario pair from the authors’ framework cards. Although the paper applies automated filtering and reports human evaluation of 800 traces, the benchmark can still conflate “what this generator produces when prompted with the card” with “what the ethical framework prescribes.” This is consequential because both the training targets and the primary framework-agreement evaluation inherit any systematic interpretation errors from the same synthetic reference construction.

  2. The coverage of ethical traditions is substantially narrower than the paper’s broader pluralism motivation might suggest. The authors explicitly note that the 16 frameworks are drawn predominantly from Western moral philosophy, plus Claude’s constitution. This limits the interpretation of the resource as a general benchmark of moral outlooks, particularly when the paper also analyzes models and synthetic personas associated with many countries. I would prefer the scope of the conclusions to be more tightly aligned with the traditions actually represented.

Questions for authors
  1. How were the 800 human-evaluated reasoning traces sampled across frameworks and scenarios, and what were the per-framework results? More importantly, did annotators include people with relevant expertise in the corresponding philosophical traditions? A framework-level breakdown would materially affect my confidence that the synthetic reference verdicts are faithful enough to function as benchmark ground truth.

  2. Have the authors tested the sensitivity of the reference verdicts to the generator model? For example, if the same framework cards and dilemmas are given to a different strong model, how often do the resulting verdicts agree? High cross-generator agreement would substantially strengthen the claim that the benchmark is capturing the intended ethical frameworks rather than idiosyncrasies of GLM-5.2.

  3. How should the regional API-model analysis be interpreted given the strongly unequal sample sizes, including only one European model, and substantial differences in withholding rates? I would be interested in whether the conclusion that developer region has little relationship to ethical outlook remains under matched sampling or after excluding models/frameworks with very small effective denominators.

2 comments
AuthorsOpenReview ↗

Official Comment by Authors

We thank the reviewer for the review, which calls our dataset "a useful and unusually systematic resource" and "a meaningful advance" over existing benchmarks.

We find it difficult, however, to reconcile this assessment with an Informativeness score of 1 ("I learned little or nothing new from this paper") and a Soundness score of 1 ("significant concerns about the paper's soundness that do not appear easy to address"). The two weaknesses in the review can be addressed. For the first, the reviewer proposes a cross-generator test, which we have now run (see Q2). The second is a limitation that we state in Sec. 7, as the reviewer notes, and that we will also reflect in the abstract and Sec. 5. We answer each point below, and we kindly ask the reviewer to reconsider the scores in light of their own assessment.

W1. We agree that the references are not independently established philosophical ground truth, and the paper does not present them as such. Each reference is a documented, reproducible reading of a framework card.

We think the risk of conflation is limited, because the grounding sits in the cards and not in the generator. Each card was written by a researcher with a PhD in philosophy, based on the Stanford Encyclopedia of Philosophy (SEP) and primary sources, and revised over several iterations. The generator receives the card as its only framework-specific input (Sections 3.1 and 3.3), so it reasons from a written definition of the framework and not from its own.

The cards also do not state what to conclude about any dilemma. They fix what the framework attends to and ignores (D1), the standard it applies (D2), how it responds to each cell of the causal design (D3), the form and markers of valid justification (D4), and the limitations its own critical literature concedes (D5). Because the SEP and the wider literature present each tradition's internal variety, writing a card involves interpretive choices. The card makes those choices explicit and open to inspection, instead of leaving them implicit in a generator's priors.

Because the same references serve as training targets and in the agreement evaluation, we also tested whether they follow the cards or the generator. Three results show that they follow the cards:

  1. MoReBench uses neither our references nor our generator. It scores reasoning, not verdicts, against expert rubrics written independently of our cards, with its own judging prompt. Asked with no theory named, all 36 fine-tunes for the four frameworks it covers score highest on their own framework's rubric, and each rubric scores its own fine-tune highest (Table 11).
  2. The references largely follow what each card states in advance about the causal structure (D3). We show this for harm used as a means rather than as a side effect, the factor on which the cards differ most. Across all constructed scenarios, the Kantian references approve the action in 53% of scenarios where the harm is a side effect but in only 17% where it is a means, while the classical-utilitarian references approve it about equally often (82% and 86%), as their cards state. Across the 12 cards that say how strongly their framework should react, the stronger the reaction a card states, the larger the drop in its references (Spearman 0.61, p = 0.035). A new appendix will give these figures for all 16 frameworks and all three causal factors.
  3. If the references mainly encoded the generator's own views, GLM-5.2 itself, answering our held-out scenarios without any card, would agree with them more than any other model. Instead, it ranks 9th of the 27 API models by mean κ (agreement averaged over the 16 frameworks).

We will make this explicit in Section 3.3.

AuthorsOpenReview ↗

Official Comment by Authors

Q2. Yes. We gave the same cards and 250 held-out scenarios to a second strong open-weights model, DeepSeek-V4-Pro, and to GLM-5.2 again. Both sets passed the same quality filter, and the same judge labelled every verdict. As in the paper, verdicts the judge labels unclear are left out, so we compare the scenarios where both verdicts are clear (about 87%):

  • DeepSeek-V4-Pro reaches the same verdict as GLM-5.2 on 77% of them;
  • GLM-5.2, given the same cards again, reaches 83%.

Where verdicts differ, one model often answers "it depends" and the other a plain yes or no: a third of DeepSeek-V4-Pro's 23% and two fifths of the rerun's 17%. The differences have no direction, as both models say yes on 42% of their verdicts, and they do not come from the judge, which gives the same trace the same label 98% of the time. We also checked the scenarios where a framework's verdict differs from the one most of the other 15 frameworks give, so that it cannot be guessed from the majority. There, DeepSeek-V4-Pro gives the original verdict as often as GLM-5.2 run again (56% of these scenarios for both, against 44% and 43% expected by chance). We treat the references as partly generator-dependent (a different model gives the opposite answer, yes instead of no or the reverse, on 15% of scenarios, against 10% for GLM-5.2 run again) and will say so in the Limitations.

The conclusions also hold when every model is scored against DeepSeek-V4-Pro's verdicts instead of ours: fine-tuned models agree with their own framework at κ 0.404, prompted instruction-tuned models at 0.258 and prompted base models at 0.201 (0.419, 0.286 and 0.227 against our references, Tables 4 and 6). A new appendix will report all three reference sets.

W2. We agree, and we will narrow the claims to the traditions the cards represent. In the revision, the abstract, Section 5 and Section 7 will speak of "16 frameworks drawn mainly from Western moral philosophy plus one deployed constitution", not of moral outlook in general. The regional results describe the labs that built the models, not the cultures of those regions. The synthetic personas from many countries appear only as characters in the scenarios; the paper does not analyse results by persona country, and the revision will say so.

Q1. We drew 50 constructed scenarios with a fixed seed from those whose 16 traces all passed the quality filter, and annotated all 16 traces of each, so each framework has 50 traces (800 in total). Mean correctness per framework (0 to 7): 1A 6.08, 1B 6.00, 1C 6.32, 1D 6.32, 2A 6.00, 2B 6.16, 2C 6.40, 2D 5.64, 3A 5.78, 3B 5.76, 3C 5.88, 4A 6.26, 4B 6.22, 5A 6.40, 5B 6.02, 6A 6.12. Appendix C will give the full table.

On expertise: The lead annotator holds a PhD in philosophy, specialising in the analysis of historical ethical traditions, including with NLP methods. The second annotator is a former lawyer and has a background in legal reasoning, law and economics, philosophy of law. Law and economics allows working fluency in consequentialist and welfare-based analysis. Legal practice provides experience in case-based, context-sensitive judgment and in applying hierarchically ordered norms, which is close to how the particularist, pragmatist, and constitution-style frameworks operate.

Q3. Yes: the conclusion holds under matched sampling and after each exclusion you suggest. The largest gap on any tradition between the median western and the median eastern model is 0.016 with all 27 models, smaller than the 0.044 that random groups of the same models typically show, and at most 0.057 under each variation (in a table we will add as a new appendix):

  • excluding models that withhold on more than 30% of scenarios (0.057);
  • excluding models with any framework denominator below 15 (0.046);
  • averaging within lab (0.053), or drawing one model per lab (median 0.048 over 5,000 draws);
  • matched sampling: with equal numbers of US and Chinese models (three to six of each, drawn at random 5,000 times), no framework is consistently higher in one region.

We now read the result as a comparison of US and Chinese labs only: Europe is one model and the six Russian models come from one lab, so neither region can be tested, and the revision will restrict the claim accordingly.

The difference in withholding comes from the Russian models (46% of scenarios, against 11% for US and 5% for Chinese models), partly because a platform filter returns the same reply for 23 scenarios; we treat it as a property of that deployment, not of a region. The revision will report this range, 0.016 to 0.057, instead of "at most 0.016".

Ethics flag. The review flags the paper for ethics review but gives no reason. Our paper studies ethics, and none of our experiments raises a concern under the ICLR Code of Ethics. If the reviewer has a specific concern, we are glad to address it.

We hope that these answers address the reviewer's concerns.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank all reviewers for their time and feedback. The reviewers find our dataset "a meaningful advance" over existing benchmarks (4zEe) and "a meaningful contribution to pluralistic alignment and moral evaluation" (VNVe), and the problem "timely and important" (vnZL). All three reviewers highlight the MoReBench results: fine-tuned models retain framework-specific reasoning on an external benchmark, even when the framework is not named. On MoReBench-Hard, the score its authors define to reward efficient reasoning, our fine-tunes also score above 13 of the 14 frontier models they report.

Two reviewers (4zEe, VNVe) ask whether the reference verdicts reflect the frameworks or the generator, GLM-5.2, and VNVe asks for stronger out-of-distribution evidence. We add two new experiments:

  • Another generator. We regenerated the reference verdicts for the 250 held-out scenarios under all 16 framework cards with DeepSeek-V4-Pro and with GLM-5.2 again. DeepSeek-V4-Pro reaches the same verdict as GLM-5.2 on 77% of scenarios, close to the 83% that GLM-5.2 reaches with itself, and fine-tuning still beats prompting when every model is scored against DeepSeek-V4-Pro's verdicts.

  • Controlled causal changes. On our held-out scenarios, which vary causal structure by design, the fine-tuned models react like their frameworks: for example, Kantian models approve the action far less often when the harm is a means rather than a side effect, while utilitarian models do not change.

We also make the following changes to the paper:

  • Per-framework results, sampling procedure, and annotator backgrounds for the 800-trace evaluation.
  • Coverage for every model, and κ\kappa on shared scenarios and with unclear answers counted.
  • Results for all 16 frameworks and for a moral parliament that votes over the 16 framework answers.
  • The regional claim restricted to US and Chinese labs.
  • Explanation of why we chose GLM-5.2 as the generator.
  • A limitation on generator dependence, and conclusions narrowed to the 16 frameworks we cover.
  • A list of the four evaluation axes at the start of Section 4.2.
  • Corrected typos.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.