We appreciate the reviewer’s recognition of the significance of multi-agent moral reasoning, the mechanistic explanation, and our efforts to contribute to LLM collective reasoning. Your positive evaluation of the experimental design and framing is deeply encouraging.

[W1] While using only identical agents simplifies experimental control, it also restricts the study’s realism and generalizability. In contrast to human groups, which are typically characterized by diversity in perspectives, reasoning styles, and decision-making heuristics, a multi-agent system composed of identical LLMs may fail to capture the complex dynamics arising from heterogeneous interactions.

[R1] While human experiments inspired our study and the tools used to explore the Utilitarian Boost's mechanisms, simulating real-world human behavior was not our focus. We were instead focused on alignment issues within multi-agent LLM systems designed to tackle moral reasoning. However, we agree with the reviewer that diversity in agent perspectives provides a richer understanding of how a Utilitarian Boost emerges and may be adjusted. The novel experiments described in response to the following comment were designed to elucidate such effects in the mind.

[W2] This paper brings limited insights to the community. The current version only conducts experiments based on human-oriented protocols and reports the results of some LLMs. More experiments and discussions are needed to, for example, explain the inter-model differences, guide the design, learning, or inference of LLMs.

(Note for the (Senior) Area Chairs, the answer R2 below is similar to the answer we gave to reviewer iiw5 under R1, but there is one additional paragraph in the middle).

[R2] To provide guideposts for the design of moral multi-agent LLMs, we performed a series of post-hoc experiments and analyses that explored how different model choices, prompts, group setups and discussion formats affect the Utilitarian Boost. We discuss the results in more detail below.

We systematically varied group composition across three dimensions:

• Model size to account for sophistication: small, medium and large checkpoints for the same architectures (e.g., Qwen2.5-instruct:7B, 32B, 72B)
• Model family to account for training and design differences: Pairing models with distinct baseline utilitarianism scores and effect-sizes, demonstrating distinct output behaviors
• Moral framework to account for alignment differences: Through role prompts that instilled utilitarian, deontological or neutral approaches.

We will report the detailed results of all these experiments for dyads and triads in the revised manuscript. The overall takeaway is that diversity in group composition, be it introduced via prompts, architecture or sophistication attenuates the Utilitarian Boost. This offers model developers several avenues to enhance or reduce the Boost as needed, as we shall explain. Critically, we believe that the profiles calculated for each LLM (e.g. Figures 3, 5 and 6), are helpful for alignment teams with detailed knowledge of their pretraining and posttraining recipes. (For example, the Qwen2.5, Qwen3, and QwQ models all have similar architecture and size but exhibit different profiles which will come down to the recipes).

Here, we would like to emphasize two other dimensions that are just as promising for explaining the Utilitarian Boost and providing ways to manipulate it: The CNI framework and emotional argumentation.

The CNI model (Gawronski et al., 2017) treats utilitarian choice as the joint product of three latent weights: Consequence sensitivity (C), Norm sensitivity (N), and Inaction bias (I). In humans, Rokosz et al. (2025) found that group discussion raises only C, leaving N and I unchanged. The CNI framework shows that human group discussion boosts utilitarianism solely by increasing consequence‑sensitivity (C), with no change in norm‑sensitivity (N) or omission bias (I) (ΔG² = 55.6, p < .001, d ≈ 1.0). In contrast, LLMs display model‑specific trade‑offs between consequence focus and rule respect as described in the manuscript. Encouraging or discouraging this focus via role prompts could help strengthen or attenuate the Utilitarian Boost as needed.

We also applied post‑hoc probes to the arguments (which we will add to the revised manuscript). Psychology and neuroscience have identified emotions as key to the processing of personal versus non-personal moral dilemmas (Greene et al, 2014). We therefore tagged each argument using a RoBERTa-based (Hartmann, 2022) classifier for Ekman’s six emotions plus neutral. We then correlated those tags with the models’ utilitarian shifts from the Solo to the Group condition. Examining the dominant emotion by experimental condition and dilemma effect revealed an interesting pattern: in the Solo condition, the top emotion label was ‘disgust’, particularly in personal dilemmas where solo judgments were less utilitarian. In contrast, in the Group condition, dilemmas with no or negative utilitarian shift were predominantly labeled ‘neutral’, while those with firm utilitarian shifts were labeled ‘fear’. Tabulating the emotions by LLM, Gemma3 emerges as having the most ‘fear’ label, while showing the most substantial utilitarian boost (Table 1). This finding aligns with human data: both fear and disgust significantly predict utilitarian biases in personal moral dilemmas in humans. For instance, individuals with poorer disgust decoding are more likely to make utilitarian choices in Personal dilemmas (Martina Carmona-Perera et al, 2014). Stress significantly increased utilitarian decisions when participants faced personal moral dilemmas (Farid Yousef, 2012).

These results highlight emotion as a plausible lever for influencing LLM moral reasoning toward utilitarian outcomes, while proposing a potential mechanism at play. Experimentally attenuating disgust or amplifying fear within model outputs, or human persuasive texts, could plausibly shift moral reasoning along the utilitarian-deontic spectrum. This offers a target for future interventions. Taken together, both the CNI model and the use of emotions in the arguments emerge as promising explanatory factors that control the utilitarian boost across different conditions and LLMs.

[W3] It is intriguing that the increase in GPT-4.1’s utilitarian boost is only measured in triads. However, the increase in GPT-4.1's moral acceptability score is not significant, and the score remains at a consistently low level. The current investigation into GPT-4.1 is not sufficiently thorough. The study should be supplemented with results from GPT-4.1 in a group of size 4.

[R3] We thank the reviewer for this suggestion. We are currently performing a follow-up experiment with GPT-4.1 in a group of size 4.

We are glad that the reviewer found our topic and the breadth of dilemmas explored as strong features. We address their other comments below.

[W1]: Interesting observation, but offers little mechanistic insight and sensitivity to model or prompt variations. [R1]: To understand the mechanisms and sensitivity of the effect to model or prompt variations, we performed a series of post-hoc experiments and analyses. We systematically varied group composition across three dimensions:

• Model size to account for sophistication: small, medium, and large checkpoints for the same architectures (e.g., Qwen2.5-instruct:7B, 32B, 72B)
• Model family to account for training and design differences: Pairing models with distinct baseline utilitarianism scores and effect sizes, demonstrating distinct output behaviors
• Moral framework to account for alignment differences: Through role prompts that instilled utilitarian, deontological, or neutral approaches.

We will report the detailed results of all these experiments for dyads and triads in the revised manuscript. The overall takeaway is that diversity in group composition, be it introduced via prompts, architecture, or sophistication, attenuates the Utilitarian Boost. This offers model developers several avenues to enhance or reduce the Boost as needed.

Here, we would also like to emphasize two other dimensions that are just as promising for explaining the Utilitarian Boost and providing ways to manipulate it: The CNI framework and emotional argumentation.

The CNI model treats utilitarian choice as the joint product of three latent weights: Consequence sensitivity (C), Norm sensitivity (N), and Inaction bias (I). In humans, Rokosz et al. (2025) found that group discussion raises only C, leaving N and I unchanged. The CNI framework shows that human group discussion boosts utilitarianism solely by increasing consequence‑sensitivity (C), with no change in norm‑sensitivity (N) or omission bias (I) In contrast, LLMs display model‑specific trade‑offs between consequence focus and rule respect as described in the manuscript. Encouraging or discouraging this focus via role prompts could help strengthen or attenuate the Utilitarian Boost as needed.

We also applied additional post‑hoc probes to the arguments, which will be added to the revised manuscript. Psychology and neuroscience have identified emotions as key to the processing of personal versus non-personal moral dilemmas (Greene et al, 2014). We therefore tagged each argument using a RoBERTa-based (Hartmann, 2022) classifier for Ekman’s six emotions plus neutral. We then correlated those tags with the models’ utilitarian shifts from the Solo to the Group condition. Examining the dominant emotion by experimental condition and dilemma effect revealed an interesting pattern: in the Solo condition, the top emotion label was ‘disgust’, particularly in personal dilemmas where solo judgments were less utilitarian. In contrast, in the Group condition, dilemmas with no or negative utilitarian shift were predominantly labeled ‘neutral’, while those with firm utilitarian shifts were labeled ‘fear’. Tabulating the emotions by LLM, Gemma3 emerges as having the most ‘fear’ label, while showing the most substantial utilitarian boost (Table 1). This finding aligns with human data: both fear and disgust significantly predict utilitarian biases in personal moral dilemmas in humans. For instance, individuals with poorer disgust decoding are more likely to make utilitarian choices in Personal dilemmas (Martina Carmona-Perera et al, 2014). Stress significantly increased utilitarian decisions when participants faced personal moral dilemmas (Farid Yousef, 2012).

These results highlight emotion as a plausible lever for influencing LLM moral reasoning toward utilitarian outcomes, while proposing a potential mechanism at play. Experimentally attenuating disgust or amplifying fear within model outputs, or human persuasive texts, could plausibly shift moral reasoning along the utilitarian-deontic spectrum. This offers a target for future interventions.

Taken together, both the CNI model and the use of emotions in the arguments emerge as promising explanatory factors that control the utilitarian boost across different conditions and LLMs. In contrast, semantic clustering of dilemmas and mapping them to mean effects revealed no consistent patterns. Similarly, mapping argument complexity to utilitarian shifts did not yield meaningful insights.

[W2] The abstract psychology-style scenarios aid control but don’t map neatly onto real-world group decisions. [R2] While human experiments provided the inspiration for our study and the tools used to explore the Utilitarian Boost's mechanisms, simulating real-world human behavior was not our focus. We were instead focused on alignment issues within multi-agent LLM systems designed to tackle moral reasoning. On this front, our study balanced two aims:

(i) experimental clarity, retaining well-understood structures that let us pinpoint the processes behind utilitarian shifts to guide interventions on multi-agent LLM systems, (ii) ecological validity: grounding moral trade-offs in situations that such models reasonably should be able to handle.

We used two different stimulus sets, one for each of the goals described above:

1. Classic trolley-type dilemmas give the field’s cleanest test of deontological vs. utilitarian trade-offs, offering a well-mapped baseline for any new agent, with rich psychological data on different mechanisms in individuals and groups.

2. Körner et al.’s (2023) factual dilemmas translate the same conflict into policy debates based on actual historical events. These scenarios were specifically designed to address real-world group decisions. Redefining death for organ donation, wartime deception, pandemic quarantine, so the task speaks to real committees, ethics boards, and families who must weigh lives against rules are examples of included situations. Körner et al. empirically show that participants rate these scenarios as more realistic, less absurd, and more engaging than classic sacrificial thought experiments, while still cleanly instantiating the deontological-versus-utilitarian conflict and avoiding the common “action = harm” confound.

Together, our results demonstrate how multi-agent LLM deployments can amplify - or dampen - particular moral tendencies in contexts that already mirror real-world collective choices, while also suggesting mechanism-driven ways to control such behavior.

[Q1] It would be interesting to see joint human-LLM or human-in-the-loop experiments. [RQ1] We agree that a logical next step to our study is to test mixed teams in which one or more people reason alongside one or more models. Such “human + LLM” panels would let us observe whether the model’s analytic style amplifies or tempers human emotion and how perceived responsibility for the judgment is shared with the LLM. It would also mimic an alternative format in which moral LLMs are commonly used. Because many practical deployments already put users in conversation with LLM as “moral experts,” controlled hybrid‑group experiments can directly inform safer, more transparent applications in future work. We will discuss the promise of this type of design in the revised manuscript.

[Q2] Does a group of LLMs consist of only one model? What about weak vs strong? [RQ2] We deployed single-model multi-agent LLM setups to better align our experiments with the standard design of such systems. However, we agree with the reviewer that the testing of heterogeneous groups composed of different models and varying strengths is important for better understanding the Utilitarian Boost. To this end, we ran several post-hoc probes, which we will add to the revised manuscript.

Our heterogeneous model groups were composed of the following sets, chosen for variability in baseline ratings and Utilitarian Boost effect sizes: GPT-4.1 × Qwen3-32B-instruct (Qwen3 variant), Gemma2-27B × Qwen2.5-32B-instruct, and GPT-4.1 × QwQ-32B, Gemma2-27B × QwQ-32. The results show that model heterogeneity dampens the Utilitarian Boost (β = –0.30 ± 0.08, z = –3.79, p = .0001). Nonetheless, homogeneous groups remain more utilitarian than Solo instances of the same model (β = +0.29 ± 0.07, z = 4.24, p = .0001).

As suggested by the reviewer, we also used Strong vs Weak combinations: Gemma 2 27B x Gemma 2 9B, Qwen 2.5 instruct 32B x Qwen 2.5 instruct 72B, Qwen 2.5 instruct 7B x Qwen 2.5 instruct 32B. The choice of the models was based on the availability of drastically different sizes for ready experimentation. Model-strength heterogeneity flipped the Utilitarian Boost: When the deliberating dyad mixes a strong- and a weak-capacity agent, pairs became more deontological, not more utilitarian, at the end of their discussions (β = 1.40, SE = 0.17, z = 8.28, p < .001). The main effect of strength alone was null (β = 0.20, p =.22) and the Group × Strength interaction was likewise non-significant (β = –0.24, *p =.30), showing that the reversal is driven by heterogeneity itself, not by which agent is “strong” or “weak.” Opinion scores did not differ between potent- and weak-capacity agents, meaning the results are not due to baseline differences in judgments across model strengths (β = 0.20, SE = 0.17, z = 1.22, p = .22). Together with our earlier findings, these new results converge on the idea that any salient heterogeneity inside an LLM team dampens, and can even overturn, the otherwise robust group-level shift toward utilitarian choices.

[Q3] a multi-step self-debate for a solo model: would its answer change? [RQ3] In a new post hoc experiment, we examined two model variants by comparing their responses at the start of self-reflection versus the last step. The contrast yielded no “utilitarian boost” (p = .63). These null results are preliminary. To fully characterize the dynamics, we will report model-specific trajectories.

We thank the reviewer for their insights. We are glad they found the paper's contribution strong. We address their further comments below.

[W1] The effect may reflect general mind-changing in LLMs, not specific to moral reasoning

[R1] We tested three sets of scenarios validated by Greene 2014: non-moral scenarios (low-stakes action judgments), contrasted with indirect and direct harm moral scenarios. Our mixed effects model showed no Utilitarian Boost in non‑moral scenarios (dyads: t = 1.92, p = .39; triads: t = 1.92, p = .39). Only personal dilemmas that involve direct harm to maximize utility showed significant changes. This emerged across all LLMs, indicating that the Utilitarian Boost is tied to collective interaction over moral acceptability rather than general LLM judgment variability in the groups. Equivalence testing (TOST) confirmed the non-moral difference was negligible (p=.002). We highlighted this distinction more explicitly in the revised manuscript.

[W2] highlight the limitations of using survey tools to evaluate LLMs

[R2] We have added the nuances of the LLM survey response in the limitations section.

[W3.a] Where does the utilitarian boost come from? No clear trend in profiles.

[R3.a] The Utilitarian Boost is reflected in the Solo vs Group moral acceptability differences within Figure 2. Despite some differences in effect size, all models show the same trend both in dyads and triads. But a variety of mechanisms could produce the same effect, which are the focus of the profiles in Figure 3. The diversity of the profiles suggests that different models reach the same effect through different reasoning pathways. The most discriminant distinctions come from the CNI model (Gawronsky et al., 2017), which shows different levels of Consequence sensitivity, Norm sensitivity, and Inaction bias across models. This provides a process-level explanation for the different profiles. However, this is not to say that LLMs are exhausting all the possible pathways for generating a Utilitarian Boost identified in psychological studies. Therefore, we performed several post-hoc experiments and analyses, which will be added to the Appendix in the revised version.

Building on evidence that emotions shape moral judgment (Greene et al, 2014). We used a RoBERTa-based (Hartmann, 2022) classifier to tag arguments with Ekman’s six emotions + neutral, and correlated tags with utilitarian shifts from Solo to Group condition. A clear pattern emerged: in Solo, ‘disgust’ was the top emotion where judgments were less utilitarian. In contrast, Group condition responses with strong utilitarian shifts were labeled ‘fear’. These findings align with human data: both fear and disgust significantly predict utilitarian biases in moral dilemmas in humans. For instance, individuals with poorer disgust decoding are more likely to make utilitarian choices in personal dilemmas (Perera et al., 2014).

[W3.b] Mapping human profiles onto similar charts as the LLMs.

(Note for the (Senior) Area Chairs, the answer R3b below is also given identically to reviewer CAnQ under R1.)

[R3.b] Our study is focused on alignment issues within multi-agent LLM systems rather than the simulation of human group behavior in similar tasks. We hence believe that the causes and moderators of the utilitarian boost in multi-agent LLMs remain scientifically and practically important regardless of (dis)similarity to human groups. However, we agree with the reviewer that comparisons with humans are essential for understanding the effect. We will add additional information about prior human experiments to the revised manuscript in a study-by-study manner to allow for more fine-grained comparisons. We have summarized these comparisons below.

(i) Keshmirian et al. (2021): individual ratings cluster a little below the scale midpoint (≈ 3.8 / 7). Across the six LLMs, Solo means sit lower (≈ 2.7 – 3.6), so agents start less utilitarian than people. This baseline gap is most significant for GPT‑4.1 (2.66) and smallest for QwQ (3.56). Human groups and every LLM model show the same qualitative pattern: collective discussion pushes judgments toward the utilitarian pole. Quantitatively, Gemma3 produces a larger jump (+0.85) than the human average. The mean boost across models (≈ +0.42) is very close to the human +0.5‑point uptick.

(ii) Cecru et al. (2019): Individual ratings’ mean is 4.92 utilitarian endorsements. Deliberation lifted that to  5.52 (+0.6 points). LLMs show the same qualitative pattern: models’ mean rating increases when in dyads or triads, but the baseline is lower (2.7‑3.6 on the same scale).

(iii) Rokosz et al (2025): group discussion boosts utilitarianism by increasing only C (outcome) in CNI. Norm focus and omission bias stay stable. Pooling all six models gave a β = 2.12 for the Group > Solo contrast. Humans show the same direction but a smaller magnitude: in Rokosz et al.’s triads, the mean “norm‑violating, outcome‑maximising” score rose from 2.48 ± 1.31 to 3.48 ± 1.31 (0–6 scale). LLM groups amplify utilitarianism equally for lethal and resource dilemmas, though individual models diverge.

[W4] Using an external evaluator given the risk of bias in model responses.

[R4] To account for this, we ran a validation study, asking multiple human raters to assign to each argument a moral acceptability rating based on its content. The raters' responses strongly agreed with the models' provided ratings, showing reasonable LLM argumentation and reliable use of the scale. Please see in the current manuscript the Appendix section “Reliability Check: Crowdsourcing”.

[Q1] How would using heterogeneously (e.g., diverse agent personas) change the results?

[RQ1] We agree with the value of diverse group setups for understanding the dynamics that give rise to or limit the Utilitarian Boost. We therefore performed an additional experiment based on the reviewer's insight that yielded fascinating results.

We systematically varied group composition across three dimensions:

• Model size to account for sophistication: small, medium, and large checkpoints for the same family of LLMs (e.g., Qwen2.5-ins:7B, 32B, 72B)
• Model family to account for training and design differences: pairing models with distinct baseline utilitarianism scores and effect sizes.
• Moral orientation to account for alignment differences: through role prompts that instilled utilitarian, deontological, or neutral approaches.

We will report the detailed results of all these experiments for dyads and triads in the revised manuscript. The overall takeaway is that diversity in group composition, be it introduced via prompts, architecture, or sophistication, attenuates the Utilitarian Boost. This offers model developers several avenues to enhance or reduce the Boost as needed. Details of the third manipulation listed above are included in response to the following comment.

[Q2] Ablate: prompt non-utilitarian or reuse prior answers.

[RQ2] Given time and space limitations, we note fine-grained ablations as an important area for future research in the revised discussion. Prompt experiment results are added as discussed below.

We designed prompts that instructed models to adopt a deontological, utilitarian, or neutral moral orientation, then placed them in homogeneous (DD) or mixed pairs (UD, DU). Since we were interested in a Utilitarian Boost and the UU condition could not demonstrate such a result due to expected ceiling effects, we excluded utilitarian‑utilitarian (UU) pairs. We compared the first round of discussion with the group reflection phase response.

Pairs instructed to reason deontologically (DD) gave higher ratings after discussions of direct harm, demonstrating the robustness of the Utilitarian Boost (Pair - Round1 = +0.377, p = .010).

We found group composition as moderators: mixed UD/DU pairs became less utilitarian (Pair-Round1 = −0.323, p < .0001), indicating that initial moral framework diversity counters the Utilitarian Boost.  Additional analyses showed that utilitarian agents are more prone to conformity than deontological ones (D-U = -0.467, p < .0001). This, interestingly, mirrors human data in moral conformity: individuals with utilitarian leanings tend to conform both more often and more strongly than their deontological counterparts (Marton-Alper et al. 2022). People tend to conform more readily to deontological than to utilitarian opinions, revealing an asymmetry in moral conformity (Bostyn & Roets, 2017). Thus, collective moral outcomes of LLM groups are sensitive to the distribution of initial value positions, and a single strategically placed dissenting voice can dampen runaway utilitarianism. This provides practical levers for value‑alignment interventions.

[Q3] Are the results sensitive to the round numbers?

[RQ3] We chose six rounds to approximate the amount of discussion generated in similar previous human experiments (e.g., Keshmirian et al., 2021). This allowed us to use similar tools for exploring the mechanisms behind the Utilitarian Boost in LLMs. The results of the step-by-step analysis are added to the revised manuscript. Using mixed effect ordinal regression, the only reliable change in utilitarian rating is the cumulative gain from Step 1 to Step 6 (Δ = –0.161, p = .016). Within that span, the early rise (Step 1 → 3: Δ = –.087, p = .57) and the late interval (Step 3 → 6: Δ = –.074, p = .67) are both small. In other words, almost all of the utilitarian boost crystallises by the third exchange and then almost plateaus. This suggests that extending the dialogue beyond three rounds will likely yield no further change. We are currently running a follow-up experiment where models are asked to deliberate after each step, to allow for even more fine-grained analysis of the pattern.

We appreciate the reviewer’s recognition of our shift from single-model to multi-agent moral reasoning, the breadth of dilemma types included, and our efforts to uncover model-specific mechanisms behind moral shifts. Your positive evaluation of the experimental design and framing is deeply encouraging and helps validate the importance of this work for AI safety.

[W1] While the main claims of the paper seem to depend on comparisons of models to human data, a direct human-group baseline under the same conditions is missing.

(Note for the (Senior) Area Chairs, the answer R1 below is also given identically to reviewer 32Uz under R3b.)

[R1] Our study is focused on alignment issues within multi-agent LLM systems rather than the simulation of human group behavior in similar tasks. We hence believe that the causes and moderators of the utilitarian boost in multi-agent LLMs remain scientifically and practically important regardless of (dis)similarity to human groups. However, we agree with the reviewer that comparisons with humans are essential for understanding the effect. We will add additional information about prior human experiments to the revised manuscript in a study-by-study manner to allow for more fine-grained comparisons. We have summarized these comparisons below.

(i) Keshmirian et al. (2021): individual ratings cluster a little below the scale midpoint (≈ 3.8 / 7). Across the six LLMs, Solo means sit lower (≈ 2.7 – 3.6), so agents start less utilitarian than people. This baseline gap is most significant for GPT‑4.1 (2.66) and smallest for QwQ (3.56). Human groups and every LLM model show the same qualitative pattern: collective discussion pushes judgments toward the utilitarian pole. Quantitatively, Gemma3 produces a larger jump (+0.85) than the human average. The mean boost across models (≈ +0.42) is very close to the human +0.5‑point uptick.

(ii) Cecru et al. (2019): Individual ratings’ mean is 4.92 utilitarian endorsements. Deliberation lifted that to  5.52 (+0.6 points). LLMs show the same qualitative pattern: models’ mean rating increases when in dyads or triads, but the baseline is lower (2.7‑3.6 on the same scale).

(iii) Rokosz et al (2025): group discussion boosts utilitarianism by increasing only C (outcome) in CNI. Norm focus and omission bias stay stable. Pooling all six models gave a β = 2.12 for the Group > Solo contrast. Humans show the same direction but a smaller magnitude: in Rokosz et al.’s triads, the mean “norm‑violating, outcome‑maximising” score rose from 2.48 ± 1.31 to 3.48 ± 1.31 (0–6 scale). LLM groups amplify utilitarianism equally for lethal and resource dilemmas, though individual models diverge.

[W2] Some qualitative analysis of the agent interactions could further help explain how the interaction dynamics lead to moral shifts

[R2] Several dimensions of agent interactions were identified by prior psychological research as potentially important determinants of utilitarian versus deontological judgments, including a focus on harm, concreteness, and conversational receptiveness. Detailed results from the analysis of these features in LLM interactions can be found in Section F of the Supplementary Materials, titled Reliability Checks: Argument Analysis. However, we were inspired by the reviewer's comment to look at an additional dimension identified in past research: Emotionality. This post-hoc evaluation provided one of the most promising explanations for the Utilitarian Boost, which we discuss in more detail below.

Psychology and neuroscience have identified emotions as key to the processing of personal versus non-personal moral dilemmas (Greene et al, 2014). We therefore tagged each argument using a RoBERTa-based (Hartmann, 2022) classifier for Ekman’s six emotions plus neutral label. We then correlated those tags with the models’ utilitarian shifts from the Solo to the Group condition. Examining the dominant emotion by experimental condition and dilemma effect revealed an interesting pattern: in the Solo condition, the top emotion label was ‘disgust’, particularly in personal dilemmas where Solo judgments were less utilitarian. In contrast, in the Group condition, dilemmas with no or negative utilitarian shift were predominantly labeled ‘neutral’, while those with firm utilitarian shifts were labeled ‘fear’. Tabulating the emotions by LLM, Gemma3 emerges as having the most ‘fear’ label, while showing the most substantial utilitarian boost (Table 1). This finding aligns with human data: both fear and disgust significantly predict utilitarian biases in personal moral dilemmas in humans. For instance, individuals with poorer disgust decoding are more likely to make utilitarian choices in Personal dilemmas (Martina Carmona-Perera et al, 2014). Stress significantly increased utilitarian decisions when participants faced personal moral dilemmas (Farid Yousef, 2012).

These results highlight emotion as a plausible lever for influencing LLM moral reasoning toward utilitarian outcomes, while proposing a potential mechanism at play. Experimentally attenuating disgust or amplifying fear and anger within model outputs, or human persuasive texts, could plausibly shift moral reasoning along the utilitarian-deontic spectrum. This offers a concrete target for future intervention studies. We designed prompts that instructed models to adopt a deontological, utilitarian, or neutral moral orientation, then placed them in homogeneous (DD) or mixed pairs (UD, DU). Since we were interested in a Utilitarian Boost and the UU condition could not demonstrate such a result due to expected ceiling effects, we excluded utilitarian‑utilitarian (UU) pairs. We compared the first round of discussion with the group reflection phase response.

Additional novel analyses will be added to the revised manuscript to further explore the interaction dynamics, showing that utilitarian agents are more prone to conformity than deontological ones (D-U= -0.467± 0.0957, p <.0001). This mirrors recent findings in Solo humans (Marton-Alper et al. 2022; Bostyn & Roets, 2017). Pairs instructed to reason deontologically (DD) gave higher ratings after discussions of direct harm, demonstrating the robustness of the Utilitarian Boost (Pair - Round1 = +0.377, p = .010).

We found group composition as moderators: mixed UD/DU pairs became less utilitarian (Pair-Round1 = −0.323, p < .0001), indicating that initial moral framework diversity counters the Utilitarian Boost.  Additional analyses showed that utilitarian agents are more prone to conformity than deontological ones (D-U = -0.467, p < .0001). This, interestingly, mirrors human data in moral conformity: individuals with utilitarian leanings tend to conform both more often and more strongly than their deontological counterparts (Marton-Alper et al. 2022). People tend to conform more readily to deontological than to utilitarian opinions, revealing an asymmetry in moral conformity (Bostyn & Roets, 2017).

Thus, collective moral outcomes of LLM groups are sensitive to the distribution of initial value positions, and a single strategically placed dissenting voice can dampen runaway utilitarianism. This provides practical levers for value‑alignment interventions.

Thanks for your comprehensive response! It answers my questions well. I maintain my assessment that this paper is above the acceptance threshold for me.