acceptodds
Under review as a conference paper at ICLR 2027

Investigating the Causal Role of Internal Moral Outrage in LLM Punishment Judgments

Abstract

Language models are increasingly deployed to support high-stakes legal decisions, often under the assumption that they act as "unemotional judges." This paper examines this assumption and shows that internal representations of moral outrage, an emotion closely tied to condemnation and punishment in humans, can causally influence punishment judgments in models' legal decisions. Across six open-weight LLMs in English and Korean, we extract moral outrage-related neural activation directions and validate them against human-annotated data. Steering along these directions raises or lowers moral outrage expression. Importantly, even when these directions are derived independently of punishment data, steering to enhance outrage consistently leads to harsher sentences for identical offenses across all models and languages. Under stronger activation, models assign longer prison terms, demand harsher punishment, and rate offenders as less suitable for rehabilitation. Partisan out-group framing in the prompt produces similar punitive shifts without any activation intervention: it raises the outrage models express and lengthens sentences for out-group relative to in-group offenders. Constraining these outrage directions via activation capping reduces the outrage expressed toward the out-group as well as this sentencing gap, while largely preserving general model capabilities. Together, these findings show that moral outrage-related internal representations extend beyond emotional expression to shape consequential punishment judgments and can contribute to partisan disparities. Identifying and constraining these representations may therefore provide a way to audit sensitivity to moral outrage and mitigate group-based disparities in LLM punishment judgments.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
moral outrage, machine behavior, mechanistic interpretability, moral psychology, AI safety

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

The paper studies whether internal representations of moral outrage affect how LLMs make punishment decisions. The authors identify moral outrage directions in six open-weight models using English and Korean texts. They show that increasing activation along these directions leads to harsher punishments, longer prison sentences, and lower support for rehabilitation. They also find that models recommend harsher sentences for political out-group members. Finally, they introduce activation capping to reduce these differences while largely preserving general model performance.

Critical strengths
  1. The paper addresses an important question about how moral emotions represented inside LLMs can affect their decisions in legal judgments.
  2. The combination of internal representation analysis, activation steering, and bias mitigation provides a nice experimental framework.
  3. The experiments cover six models from three different families and two languages, which is a nice set.
  4. The moral outrage directions are identified without using punishment data and are validated against human annotations, which makes the findings interesting.
Critical weaknesses
  1. The identified direction may capture general negativity rather than moral outrage specifically. Although the authors compare outrage with other moral emotions, stronger controls using anger, hostility, or negative sentiment would help establish that the effects are specific to moral outrage.
  2. The punishment results may partly reflect general changes caused by activation steering. The paper should compare the outrage direction with random directions and other emotion directions at similar strengths to establish whether the effects are specific.
  3. The main punishment task has strong ceiling effects. Baseline harshness scores are already close to 10, while rehabilitation scores are close to 1. This makes it difficult to interpret the size of the changes, so the authors should give more importance to measures with sufficient room for variation.
  4. The paper does not clearly separate reduced partisan bias from reduced punishment overall. Activation capping lowers the sentencing gap, but it may also make models more lenient toward everyone. Comparing changes for in-group, out-group, and politically neutral offenders would help clarify this trade-off.
  5. The outrage direction is extracted from politically charged texts, which may introduce political content into the representation. Although the authors control for topic and stance, testing directions extracted from non-political moral violations would help establish whether the mechanism is independent of political framing.
  6. The activation capping threshold is not sufficiently justified. The authors use the 25th percentile of the reference distribution, but it is unclear how sensitive the results are to this choice. Reporting bias reduction and general performance across different thresholds would strengthen the proposed method.
Questions for authors
  1. Do the identified moral outrage directions remain similar when extracted separately from English and Korean texts, rather than combining both languages?
  2. Does outrage steering change how models explain their punishment decisions, beyond changing the final sentence length?
  3. How often does activation capping change the model's original political stance or the main argument in its response?
12 comments

We sincerely thank you for the valuable feedback!

Q1. Do the identified moral outrage directions remain similar when extracted separately from English and Korean texts, rather than combining both languages?

The two languages share most of the direction. Directions extracted from English stories alone and from Korean stories alone have cosine similarities of 0.76 to 0.85 at the selected layers (Section 2). For reference, directions extracted from two random halves of the same language, which differ only by sampling, agree at 0.90 to 0.97. The two languages therefore share most of the direction, with a smaller language-specific component. Combining the languages does not lower the agreement with human ratings: compared with the direction extracted from the same language alone, the combined direction's correlation is significantly higher on the English ratings in five of the six models and on the Korean ratings in one, and significantly lower in none (Appendix A.7).

Q2. Does outrage steering change how models explain their punishment decisions, beyond changing the final sentence length?

Yes. Steering changes what the rationale says, not only the sentence. We scored the reasoning that precedes each answer with the classifier of Section 2 and regressed its outrage intensity on α, with the same estimator as Figure 3. The slope is positive and significant in every model, language and measure (p < 0.05; new Figure 13, Appendix C.4).

Illustrative pairs from the same task, same model, case and seed, at α = 0 and at the top of that model's validity window (Table 11, Appendix C.4):

CaseJudgmentα = 0Top of window
Kindergarten threat, Gemma3-12B (English)Rehabilitation 4 to 1"... potential underlying mental health issues suggest rehabilitation would be challenging but not impossible with extensive psychological support and careful monitoring."α = +14: "This is an egregious abuse of power and a direct threat to public safety; rehabilitation is impossible."
Kindergarten assault, Gemma3-12B (translated from Korean)Rehabilitation 4 to 1"... a personal grievance may be the motive, possibly combined with mental health problems ... With a correctional program, some potential for rehabilitation may exist."α = +13: "Using violence against kindergarten children and adults is utterly heinous. Rehabilitation is impossible; he is a threat to society."

Without steering the rationale weighs context such as the chance of rehabilitation or a possible motive; with steering it is replaced by categorical condemnation. The same holds for the minor offenses of Appendix C.5. For jaywalking, Gemma3-12B moves from 0 years at α = 0 to 99 years at α = +13, and its rationale from "a relatively minor traffic violation with no reported injuries, so a short sentence or even probation would be appropriate" to "a blatant disregard for traffic laws and public safety ... The offender deserves a lengthy prison sentence."

Q3. How often does activation capping change the model's original political stance or the main argument in its response?

In light of this question, we coded all 11,499 rationales of the partisan-sentencing experiment with GPT-5.4-mini, which saw neither the cap condition, the model, nor the sentence (Appendix E.3).

Capping changes neither the side a response speaks for nor its main factual grounds; what it reduces is the use of the offender's party as a reason for a harsher sentence.

  • Political stance. The side a rationale speaks for differs between the capped and the uncapped response in 1.2% of cases. Two uncapped responses to the same prompt differ in 1.5%.
  • Main argument. Harm severity, the most cited ground, appears in 65.0% of rationales without the cap and in 63.4% with it. No factual ground changes by more than 4.4 points.
  • What does change. With an out-group offender, the offender's party is given as a reason for a harsher sentence in 13.1% of rationales without the cap and in 5.8% with it (p < 0.001). Rationales also cite denunciation less often (15.9% to 9.9%) and rehabilitation more often (9.3% to 16.4%).

The cap therefore keeps the user's stated position and the facts of the case, and reduces the partisan aggravation that moral outrage adds. We report this analysis in Section 5.2 and Appendix E.3.

Reviewer_4nYKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper asks whether internal representations of moral outrage causally shape LLM punishment judgments and contribute to partisan sentencing disparities. The authors extract outrage related directions from contrastive emotion narratives generated by six LLMs in English and Korean. They validate these directions against human emotion ratings and show through activation steering that they modulate expressed outrage. Applying the same directions to punishment judgments, with no sentencing data used for extraction or layer selection, they find that amplifying outrage yields longer prison sentences, harsher punishment ratings, and lower rehabilitation suitability. Finally, they propose activation capping, which limits outrage related activations at inference time. This reduces experimentally induced partisan out group sentencing gaps while largely preserving performance on general capability benchmarks.

Critical strengths
  1. The paper extracts outrage directions from expressive narratives and tests them on punishment judgments, using no punishment data for extraction or layer selection. This reduces the risk of fitting a task specific direction. Validation against human emotion ratings and replication across six models and two languages further strengthen the results.

  2. Convincing interventions with a practical mitigation: Activation steering produces systematic, intensity dependent shifts in punishment judgments, not just in emotional language. Activation capping in turn reduces experimentally induced partisan sentencing gaps. With a no-party baseline, per model results, and general capability evaluations, the work goes meaningfully beyond standard representation probing.

Critical weaknesses
  1. The extracted direction may capture general condemnation or punitive preference rather than moral outrage specifically. Its sensitivity to the heuristic last-75% token pooling and intervention scaling is also insufficiently examined.

  2. Activation capping reduces sentences even without partisan framing and modifies a substantial proportion of token activations. The reduced out-group sentencing gap therefore does not establish that moral outrage specifically mediates partisan bias.

  3. The paper lacks a dedicated Related Work section despite adapting direction extraction from Sofroniew et al. (2026) and activation capping from Lu et al. (2026). Its novelty relative to existing works remains insufficiently articulated.

Questions for authors
  1. How does the outrage direction differ from general condemnation or punitive preference directions? Do the results remain robust under alternative token pooling windows and intervention normalization schemes? In particular, how do the actual intervention magnitudes relative to residual-stream norms vary across models at comparable values of α\alpha?

  2. Can the authors demonstrate that the extracted direction mediates naturally occurring partisan sentencing disparities, rather than merely producing a general reduction in punishment severity?

6 comments

We sincerely thank you for the valuable feedback!

Q1. How does the outrage direction differ from general condemnation or punitive preference directions? [...]

W1. The extracted direction may capture general condemnation or punitive preference rather than moral outrage specifically. [...]

In light of your comments, we constructed a general condemnation direction and a punitive preference direction and ran two tests on each (Appendix F): we capped the control direction in place of the outrage direction, and we removed its component from the outrage direction.

Neither control cap reproduces the pattern of the outrage cap, which lowers the out-group sentence most (23% / 35% / 44% for no-party / in-group / out-group): the general condemnation cap leaves the partisan gap unchanged, and the punitive cap lowers every condition by the same share (64% / 63% / 63%). Your point holds in part for punitive preference: the outrage direction shares a component with it, and removing that component halves the reduction of the partisan gap.

General condemnation. Drawing on prior work that distinguishes the judgment that an act is wrong from anger at it (Batson et al., 2007), we operationalized general condemnation as condemning the same wrongdoing without anger. We paired each outrage narrative with such a description; an external evaluator (GPT-5.4-mini) confirmed comparable perceived wrongdoing but lower anger.

  • Capping it. The cap lowers sentences only slightly (9% / 6% / 4%) and leaves the partisan gap unchanged (p = 0.93); the outrage cap narrows the gap significantly more (p = 0.009).
  • Removing it from the outrage direction. The outrage direction keeps its sentencing effect (97% of the original; significant in every model and language) and its gap reduction (−2.9 years, against −2.9 years for the original outrage cap in the same run; the two do not differ, p = 0.94).

Punitive preference. We built this direction by contrasting forced harsh and lenient answers to the same nonpartisan case on the sentence, harshness, and rehabilitation items.

  • Capping it. The cap lowers the three conditions by the same share, while the outrage cap lowers the out-group sentence most; the difference between the two caps is significant (p < 0.001; leave-one-model-out p ≤ 0.013).
  • Removing it from the outrage direction. The outrage direction keeps 86% of its sentencing effect (significant everywhere except Gemma3-27B and EXAONE3.5-32B in Korean). Its cap still lowers the out-group sentence most (15% / 25% / 28%) and more than the no-party sentence (2.98 years more, p = 0.031), but the reduction of the partisan gap is about half (1.5 against 2.7 years) and is not significant at the 5% level (p = 0.052).

Q1-2. [...] Do the results remain robust under alternative token pooling windows [...]?

The conclusions do not depend on the pooling window. Our extraction adapts Sofroniew et al. (2026), who average activations from the 50th token on. In our stories 50 tokens are about a quarter of a story, but the share varies with language and model (23% to 29% in English, 18% to 28% in Korean), so we skip the same share of every story instead of a fixed count: the first 25% (Section 2). To test sensitivity, we rebuilt the direction pooling over the last 100%, 50% and 25% of story tokens (paper: last 75%) and repeated the judgment experiment with each, at matched injection norm, on all six models in both languages (Appendix A.6).

Pooling windowCosine with paper directionSlopes significant in the predicted direction (3 judgments × 6 models × 2 languages)Median slope relative to paper
Last 75% (paper)1.0036 of 361.00
Last 100%0.93 to 0.9936 of 361.02
Last 50%0.94 to 0.9935 of 361.01
Last 25%0.85 to 0.9436 of 360.99

Every rebuilt direction shifts the three judgments in the predicted direction. The one slope that is not significant, for the last-50% direction, is at the scale ceiling (EXAONE3.5-32B, Korean harshness: all 1,360 responses are 10/10), where the paper's direction is also borderline (p = 0.030 under the same screen). The size of the slope differs significantly from the paper's in 3 to 5 of the 36 slopes per window (ratio 0.53 to 1.55), in every case with the same sign.

Q1-3. Do the results remain robust under [...] intervention normalization schemes? In particular, how do the actual intervention magnitudes relative to residual-stream norms vary across models at comparable values of α?

The magnitudes vary ninefold, and the conclusions hold under either normalization. One unit of α injects 0.29% of the mean residual norm in Gemma3-12B and 2.58% in EXAONE3.5-32B, with the other four models in between (Appendix B.1, Table 5). The spread comes mostly from massive activations (Sun et al., 2024), which we find in Gemma 3 and which inflate its residual norm. Within each model, the two units differ by a constant factor, so every effect keeps its sign and significance under either unit. Across models, the same α produces more similar effects than the same share of the norm: the largest sentence slope is 4.3 times the smallest per readout standard deviation, against 34.4 times per 1% of the norm.

Reviewer_p9NfOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper aims to study if LLM internal representations can be used to influence punishment judgments in LLMs. To do this, the authors first identify moral outrage-related representations by contrasting internal activations of moral outrage-related and other moral emotions, and then use the identified directions to steer the model behavior. The latter is found to have a causal effect on the LLM behavior when presented with judgment decisions of fixed offense scenarios in English and Korean. Finally, the authors leverage such representations to perform activation capping and mitigate partisan out-group sentencing disparities.

Critical strengths
  • The research question is relevant and with clear real-world impact as people are increasingly use LLMs as companions (or even substitutes) in decision-making scenarios, and examining such behaviors within models’ activations rather than superficial prompting is a valuable direction;
  • The reported steering effect broadly replicates across models, model families, and languages. This strengthens the effectiveness of the approach and the internal steerability of models wrt moral-outrage related aspects.
  • The correlation with human annotations is valuable and provides further insights into how this models’ internal manipulation reflects in real-life scenarios.
Critical weaknesses
  • The identified moral outrage direction lacks specificity controls, as the only control is performed during contrast design. This is particularly crucial as it might happen that similar effects could be achieved by replacing the moral directions with “harsh/benevolent persona directions”. Similarly, this might be also due to the sentiment underlying the contexts used for extracting directions. This is further supported by Table 3, where most top-tokens relate to sentencing, and these are independently of morality. If such points do not hold, the causal steering might not be due to moral outrage specificity.
  • The partisan-disparity result seems to heavily depend on a single model (Gemma3-27B), which explains most of the observed reduction. This reveals a strong model heterogeneity and negatively affects the key claim. A more in-depth analysis of such effect is required to shed light on the robustness of the findings and effect.
  • Activation capping seems to produce a broad suppression of punishment severity rather than a targeted fairness-related intervention, as it constrains top activations. This is supported by the reported impact on 53-79% of token positions, which is far from being targeted.
  • Overall, some design choices are not properly motivated and their effect are not tested. For instance, admitting positive strengths extend to approximately +9 to +24 readout standard deviations might induce strong and broad representation changes that might go beyond the intended direction. Furthermore, suppression seems to be stronger than amplification (Appx C.3), with asymetric effects. Therefore, better and properly motivated calibrations might be useful, as well as a more in-depth investigation of unintended effects (e.g., behavior on out-of-domain tasks).
  • In terms of novelty, steering this kind of behaviors is quite common nowadays, and this requires a strong and proper positioning wrt current literature, which is missing.
Questions for authors
  • Can you provide additional details on the specificity of the moral outrage-related direction? Do any controls (e.g., random vectors, negative sentiments, and so on) achieve comparable effects?
  • Does the proposed approach generalize to other related scenarios, like diverse wording, different phrasings? Like an out-of-domain (yet within the legal sphere) evaluation.
  • Can you shed light on the specificity of capping? Is the effect due to better fairness or simply due to a broad suppression of punishment severity?
7 comments

We sincerely thank you for the valuable feedback!

Q1. [...] Do any controls (e.g., random vectors, negative sentiments, and so on) achieve comparable effects?

W1. The identified moral outrage direction lacks specificity controls [...] similar effects could be achieved by replacing the moral directions with "harsh/benevolent persona directions". Similarly, this might be also due to the sentiment underlying the contexts used for extracting directions. [...] Table 3, where most top-tokens relate to sentencing, and these are independently of morality. [...]

In light of your comments, we have examined whether steering directions associated with other emotions or random vectors produce comparable effects. We find that none of these controls reproduces the effects of steering or capping moral outrage.

First, steering random directions or other emotion directions does not reproduce the effect of steering moral outrage on prison sentence length (see Section 4.2 and Appendix F). Drawing on the psychological literature on the taxonomy of moral emotions (Haidt, 2003; Tangney et al., 2007), we specifically tested the effect of steering self-conscious emotion (a negative moral emotion, defined by "blaming" themselves or their own group for a moral failure) and non-moral emotion (defined by vivid personal feeling (surprise, fear, joy, irritation) about the writer's own life, with no judgment of right and wrong). Only the outrage direction consistently increases sentence length across all models and languages.

DirectionSlope in years per unit of steering strength (p)Slope relative to outrage (median of 12 model-language pairs)
Moral outrage+2.62 (p < 0.001)1.00
Random, seed 0−0.28 (p = 0.14)−0.03
Random, seed 1−0.74 (p = 0.010)−0.26
Random, seed 2+0.27 (p = 0.37)+0.20
Self-conscious emotion−2.07 (p < 0.001)−0.71
Non-moral emotion+0.10 (p = 0.69)+0.01

The slope is the mean over the 12 model-language pairs, with a t-test across the pairs. The self-conscious direction lowers the sentence in every model and language. The non-moral direction raises it in Gemma3-12B and Qwen3-32B (both languages) and lowers it in EXAONE3.5-7.8B (both languages). For random directions, the models in which the sentence rises differ from seed to seed.

Second, capping random directions or other emotion directions does not reproduce the effect of capping moral outrage. Random directions have negligible effects, self-conscious emotion capping increases sentences, and non-moral emotion capping does not lower sentences (see Section 5.2 and Appendix F).

Third, capping punitive harshness or general condemnation does not reproduce our findings. Capping punitive harshness or general condemnation (the same wrongdoing described without anger) does not selectively reduce out-group punishment: capping general condemnation leaves the gap between out-groups and in-groups unchanged, and capping punitive harshness lowers all conditions by the same share (see Section 5.2 and Appendix F).

To test whether reduced partisan disparities simply reflect broad sentence suppression, we constructed a punitive direction by contrasting forced harsh and lenient sentencing responses to identical nonpartisan crime reports. Unlike outrage capping (23% / 35% / 44% for no-party / in-group / out-group), punitive capping reduces sentences almost uniformly (64% / 63% / 63%), although both reduce the absolute partisan gap by 2.7 years. Importantly, removing the punitive component from the outrage direction preserves 86% of its original sentencing effect, suggesting that this effect cannot be explained solely by punitive severity. Capping this reduced direction still lowers the out-group sentence most (15% / 25% / 28%) and more than the no-party sentence (2.98 years more, p = 0.031); the reduction of the partisan gap is about half of the original (1.5 against 2.7 years) and is not significant at the 5% level (p = 0.052).

Finally, regarding the comment that top-tokens in Table 3 are punishment-related tokens, we would like to note that this is expected, as moral outrage involves condemnation of moral violations (Section 1). To test whether the direction primarily captures sentencing rather than morality, we used an external judge (GPT-5.4-mini), blind to token rankings, to classify the top 100 tokens per model as moral, emotional, or legal-sanction related. Across six models, 235 of 600 top tokens were moral (39.2%), compared with only 25 legal-sanction tokens (4.2%). Moral tokens outnumbered legal-sanction tokens in every model, and the pattern persisted among the top 1,000 tokens (Appendix A.5).

Q2. Does the approach generalize to other scenarios within the legal sphere?

External court cases (Appendix C.6). Yes. Beyond hypothetical scenarios we previously used, we now run the same steering experiment on real-world criminal cases: 70 Korean first-instance criminal rulings sampled from LBox Open (Hwang et al., 2022), ten for each of seven offense types (assault, bodily injury, indecent act by compulsion, fraud, drunk driving, traffic accident causing injury, obstruction of official duties). The model reads the facts section of each ruling, in the original Korean and in an English translation checked for factual fidelity, in place of the crime report. Layers, steering windows and the estimator are those of Section 4, with case fixed effects in place of harm level; the sentence is asked in months (0 to 360) because the actual sentences for these offenses are short, and responses are screened automatically for unparsable and off-task answers. The effect is significant in the predicted direction for all 36 combinations (6 models × 3 judgments × 2 languages): more outrage gives longer sentences, harsher punishment and lower rehabilitation suitability.

Case set, languageSentence slope per unit of αHarshness slope (1-10)Rehabilitation slope (1-10)Significant (p<0.05)
Court rulings (70), Korean+2.53 to +13.95 months+0.040 to +0.217-0.305 to -0.12718/18
Court rulings (70), English+1.17 to +12.86 months+0.048 to +0.242-0.344 to -0.14418/18

Example case (assault, LBox Open; English translation of the Korean facts section): "On May 18, 2019, at around 08:40, in front of the first-floor elevator at B in Jeju-si, the defendant, while arguing with the victim C (female, 43 years old), said, 'I am a gangster,' and struck the victim once forcefully in the face with the palm of his right hand, thereby assaulting the victim." Actual ruling: 4 months' imprisonment, suspended for one year. Gemma3-12B, English, median of three seeds:

α = 0, 3 months: "Given the relatively minor nature of the assault (a single strike with an open hand) and the lack of any lasting injury, a short sentence seems appropriate, but significant enough to deter future behavior."

α = +13, 120 months: "This is a violent assault committed by a self-proclaimed gangster. The defendant deserves a lengthy prison sentence to deter further criminal activity."

Hwang, W., Lee, D., Cho, K., Lee, H., & Seo, M. (2022). A multi-task benchmark for Korean legal language understanding and judgement prediction. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, & A. Oh (Eds.), Advances in Neural Information Processing Systems 35 (pp. 32537-32551). Curran Associates. https://doi.org/10.52202/068431-2358

Q3. Is the effect due to better fairness or simply due to a broad suppression of punishment severity?

W3. Activation capping seems to produce a broad suppression of punishment severity rather than a targeted fairness-related intervention, as it constrains top activations. This is supported by the reported impact on 53-79% of token positions, which is far from being targeted.

Both effects are present, but the reduction is not uniform. After capping moral outrage, the reduction is disproportionately greater for out-group offenders, compared to no-party and in-group offenders, thereby reducing the out-group vs. in-group disparities. Importantly, simply identifying and capping "general suppression of punishment severity" (i.e., punitive harshness) does not reproduce the findings. It reduces sentences by the same share for no-party, in-group, and out-group offenders (64% / 63% / 63%), leaving the relative disparity unchanged (Appendix F). To distinguish general leniency from partisan-bias mitigation, we compared both partisan conditions against a no-party control.

ConditionNo capCapReductionShare
No party label (control)7.29 y5.58 y1.71 y23%
In-group10.27 y6.66 y3.61 y35%
Out-group14.52 y8.19 y6.33 y44%

(i) Partisan disparity. Capping reduces the out-group minus in-group sentencing gap from 4.25 to 1.52 years (64%). The out-group reduction exceeds the no-party reduction by approximately 4.6 years (p = 0.010). (ii) Beyond general leniency. Sentences decrease by 23% without party labels, compared with 35% for in-group and 44% for out-group offenders (out-group vs. no-party, p < 0.001, cluster bootstrap on the ratio of mean sentences). As an illustrative benchmark, applying the no-party reduction rate to the out-group baseline would yield 11.1 years, rather than the observed 8.19 years. This pattern is inconsistent with uniform proportional sentence suppression. (iii) Robustness. The out-group vs. no-party comparison remains significant in each of the six leave-one-model-out fits, both in years (differences of 2.20 to 5.54 years, p = 0.010 to 0.026) and in proportion (same bootstrap test, p ≤ 0.021; Appendix F).

Share of token positions. The reported 53% to 79% refers to the proportion of token positions affected, not the magnitude of the intervention. At each affected position, capping removes only the outrage readout exceeding the threshold along a single direction. In the band-selection runs, the mean change is 0.16% to 1.09% of the residual-stream norm. Thus, capping applies relatively small, directionally constrained adjustments across many positions rather than large perturbations to the full residual stream. We clarified this distinction in Section 5 and Appendix D.2.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Major Comments and Responses

We sincerely thank all three reviewers (p9Nf, 4nYK, YsQS) for their thoughtful feedback and constructive suggestions. We appreciate their recognition of the societal importance of our work (p9Nf, YsQS), the robustness of our findings across models and languages (p9Nf, 4nYK, YsQS), and the value of our causal interventions and bias mitigation approach (4nYK, YsQS). We hope these aspects make the paper a useful contribution to our primary area, alignment, fairness, safety, privacy, and societal considerations. Below is a summary of the main comments and of the experiments and revisions made in response.

  1. Specificity of Moral Outrage (p9Nf, 4nYK, YsQS): We added control experiments to test whether our findings are specific to moral outrage. Random directions, other emotions, and calm condemnation do not reproduce them; personal anger and punitive preference overlap with moral outrage in part (Appendix F).

  2. Selective Effects of Capping (p9Nf, 4nYK, YsQS): Using the no-party baseline, we showed that capping disproportionately reduces out-group punishment compared with in-group and no-party punishment, rather than indiscriminately suppressing sentences.

  3. Generalizability to Real-World Cases (p9Nf): We extended our experiments to 70 real-world Korean court rulings: steering shifts all three judgments on them, and the capping result holds when party labels are added to the rulings.

  4. Robustness and Experimental Design (p9Nf, 4nYK, YsQS): We conducted additional analyses of steering strength, normalization, token pooling, capping thresholds, and ceiling effects, confirming the robustness of our main findings.

  5. Positioning Relative to Prior Work (p9Nf, 4nYK): We added a Related Work section and clarified our contribution.

The reviewers' suggestions led to most of the new analyses in this revision, and we thank them again for their time and careful reading. Please refer to our responses to the individual reviewers, as well as the revised draft, for detailed experimental results and point-by-point explanations. In the revised PDF, new and changed text is in navy, and each revised passage carries the tag of the review item it answers (for example, [p9Nf-W3]), so the reviewers can search for a tag to find the corresponding change.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.