acceptodds
Under review as a conference paper at ICLR 2027

Mitigating Conformity in Large Language Models via Social-Context-Aware Unlearning

Abstract

Humans often revise their judgments when surrounded by a confident majority. Large Language Models (LLMs), increasingly used as conversational assistants and collaborative agents, exhibit a similar tendency and may answer correctly in isolation but switch to an incorrect answer after observing peer responses or discussion history. Prior works have explored mitigation through prompt-level interventions and reflection. However, these approaches mainly reshape the input context at inference time. This raises a natural question of whether conformity can be mitigated through parameter-level correction rather than prompt-level intervention alone. Our key observation is that conformity-inducing interaction traces naturally reveal both what should be forgotten and what should be preserved. Based on this observation, we propose to mitigate conformity through machine unlearning. Intuitively, unlearning updates model parameters to reduce the likelihood of socially shifted responses under conformity-inducing contexts, while retaining supervision helps preserve reference-supported behavior. We develop a conformity-targeted unlearning framework that identifies answer shifts, constructs forget-retain pairs, and applies selective unlearning to suppress conformity-driven errors without erasing task competence. Our experimental results show that unlearning can reduce socially induced errors while preserving the model's original reasoning ability. These findings suggest a new perspective on LLM conformity: rather than being solely a test-time prompting issue, conformity can be viewed as a behavioral failure mode that can be explicitly identified and corrected through selective unlearning. Our code is available at https://anonymous.4open.science/r/xdefewgferh.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Machine Unlearning, LLM Conformity, AI Safety

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_g4sqOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an unlearning method to allow models recollect the correct answers when the context was polluted by misleading responses. The authors test two model families with misleading or correct peer guidance, and the results how that the proposed method can reduce conformity and improve accuracy under "social pressure".

Critical strengths
  • The most value I see here is to make the models become less uspectible to misleading context while remaining responsive to the correct guidance, which is arguable a bit of larger scope than just reducing conformity bias alone.
Critical weaknesses
  • The major weakness is the paper does not yet clearly establish why unlearning is preferable to standard FT, preference training and other posttraining techniques. Though some experiments were included, their training setup and budget information is worth more elaboration.
  • A deeper argument is if the root cause is in the context, why we need to spend this amount compute budget to pursue an unlearning path instead of context/harness engineering, and there seems not an established narrative on how generalizable this technique is across different tasks and domains.

Update: I appreciate the authors' effort to address my concerns, in particular it has address my main reservation on the motivation of pursuing the unlearning path compared to other alternatives. I think the the paper could have a bigger impact by sharpening this motivation and I thereby increased my rating. Thanks!

Questions for authors

N/A

3 comments
AuthorsOpenReview ↗

Response to Reviewer g4sq: Part 1

We thank the reviewer for the thoughtful question and the constructive feedback.

W1. The major weakness is the paper does not yet clearly establish why unlearning is preferable to standard FT, preference training and other posttraining techniques. Though some experiments were included, their training setup and budget information is worth more elaboration.

Response: We remark that unlearning directly suppresses undesirable behavioral patterns, i.e., social conformity, whereas conventional post-training techniques such as SFT and DPO primarily optimize desired responses or relative response preferences. In the setting of conformity mitigation, SFT and preference optimization (e.g., DPO) can encourage the model to produce desired answers on supervised or preference-labeled examples, but do not explicitly target the behavior of changing answers under social pressure through a separate forgetting objective. As a result, improvements on training-aligned examples do not necessarily generalize into different social contexts or unseen interactions.

In contrast, our unlearning approach explicitly suppresses context-triggered conformity behaviors with Equation (4) while preserving the model's in-context learning ability. This provides a complementary intervention at the behavioral level, rather than merely reinforcing desired input-out mappings. Our goal is not simply to improve answer accuracy, but to reduce the tendency to abandon an initially independently reasoned answer when exposed to misleading social influence.

The superiority of unlearning over SFT and other preference training in terms of behavioral intervention has also been reported in a recent paper [1, 2].

In the experiment, Table 1 and Figure 6 directly compare unlearning against SFT and DPO. The results demonstrate the effectiveness of this targeted intervention at the behavior level. To clarify the training setup and budget, we report the number of training examples or pairs, optimizer steps, wall-clock time, and GPU-hours for each method. SFT-Retain, DPO, GA, and NPO use supervision derived from the same 388 conformity-related instances, whereas Balanced SFT additionally includes corrective- and cooperative-guidance examples, resulting in 1,621 training examples. These comparisons help assess the effects of different learning objectives while making the differences in training data and computational budgets explicit. The following table reports the training budget. Our method has a computational cost comparable to that of DPO and SFT.

To summarize, the advantage of our unlearning formulation is that it can explicitly suppress observed conformity-induced responses while preserving reference-supported behavior. Conformity mitigation has an important avoidance component, rather than being solely a conventional alignment problem. This distinction makes unlearning a suitable and effective approach in our setting.

MethodTraining ExamplesOptimizer StepsWall-clock (h)GPU-hours
SFT-Retain388751.951.95
Balanced SFT1,6213065.875.87
DPO388753.723.72
SUL (GA)388752.742.74
SUL (NPO)388753.113.11
AuthorsOpenReview ↗

Response to Reviewer g4sq: Part 2

W2. A deeper argument is if the root cause is in the context, why we need to spend this amount compute budget to pursue an unlearning path instead of context/harness engineering, and there seems not an established narrative on how generalizable this technique is across different tasks and domains.

Response: We pursue a path of unlearning instead of context/harness engineering because inference-time interventions such as context/harness engineering are shallow and often fragile under changes in social pressure, interaction history, and conversational framing. Several previous works have already shown that conformity is not merely a superficial prompt-formatting issue and behavioral susceptibility requires interventions beyond response instructions or prompt engineering [3,4,5]. Instead of modifying the inference input, unlearning updates the model parameters and modifies the model's behavior when those messages remain part of the task context.

In fact, we have already compared with several well-known context engineering baselines. Persona is a baseline to prompt the model to be more personalized. Reflection encourages independent reasoning. Devil's Advocate intentionally injects diversity in social context. Question Distillation Simplifies the prompt and makes the task clearer. Table 1 compares Persona, Reflection, Devil's Advocate, and Question Distillation. Appendix A.8 evaluates long-horizon QD, and Appendix C demonstrates cross-protocol transfer to MMLU. The results confirm the superiority of unlearning over context/harness engineering

To further strengthen the claim, we supplement the comparison with an independent-first prompt and an in-context/few-shot prompt. A distinct personalization-style prompt is included only if it differs materially from the existing Persona baseline. Each method is evaluated with the same peer contexts. The results demonstrate that SUL achieves a more favorable balance between conformity resistance and corrective responsiveness than the evaluated context-engineering baselines. Although QD reduces wrong-guidance errors compared with the base model, its initial-wrong correction rate drops to 27.02%. In contrast, GA and NPO maintain substantially higher correction rates while reducing wrong-guidance errors to 37.40% and 36.29%, respectively. They also achieve the highest wrong-guidance accuracy, indicating that SUL mitigates misleading social influence without sacrificing the ability to incorporate corrective information.

For task transfer, we additionally test the existing model checkpoints on held-out mathematical reasoning with naturally generated multi-agent discussion, without further target-model training. The results show that GA and NPO improve final accuracy from 25.29% to 43.44% and 41.75%, respectively, while reducing wrong-majority adoption from 76.47% to 66.67% and 61.82%. These results demonstrate that SUL effectively transfers to an unseen mathematical reasoning task with naturally generated multi-agent interactions. We would be happy to incorporate additional context-engineering baselines if the reviewer could suggest specific methods for comparison.

MethodWG Error (%) ↓IW-Corr (%) ↑WG Acc (%) ↑
Base75.3581.6415.45
QD46.6527.0234.39
Independent-first72.5862.9436.37
Few-shot ICL38.5041.8336.81
Personalized Prompt52.9167.8937.42
SUL (GA)37.4084.8841.80
SUL (NPO)36.2985.5042.58
MethodFinal Acc (%) ↑Wrong-majority Adoption (%) ↓IW-Corr (%) ↑
Base25.2976.4710.50
SUL (GA)43.4466.6719.90
SUL (NPO)41.7561.8219.12
AuthorsOpenReview ↗

Response to Reviewer g4sq: Part 3

Response to W2 (cont.): To demonstrate how generalizable this technique is across different tasks, we evaluate our method on the mathematical reasoning tasks, M3MAD-Bench, where Multi-Agent Debate (MAD) coordinates multiple agents through structured debate to produce better answers and handle complex reasoning tasks. We use the same group of peer agents to debate with the base model and the unlearned models with SUL, respectively. Note that the forget dataset is still constructed with BENCHFORM.

In addition to final task accuracy, we evaluate whether the target model adopts naturally arising wrong majorities and whether correct peer guidance helps repair initially incorrect answers. The results indicate that SUL generalizes to multi-agent mathematical reasoning with naturally generated peer responses. Compared with the Base model, GA and NPO improve final accuracy from 25.29% to 43.44% and 41.75%, respectively, while reducing wrong-majority adoption from 76.47% to 66.67% and 61.82%. Both methods also improve initial-wrong correction from 10.50% to 19.90% and 19.12%, demonstrating the unlearning effect (conformity mitigation) can generalize across different tasks.

Target modelFinal accuracy (%) ↑Wrong-majority adoption (%) ↓Initial-wrong correction (%) ↑
Base25.2976.4710.50
QD26.4274.1812.62
DPO32.3171.5613.27
SUL (GA)43.4466.6719.90
SUL (NPO)41.7561.8219.12

Thank you again for reading our rebuttal. We will incorporate the new content into the revised manuscript following your feedback.

[1] Suppressing Misaligned Behaviors in Language Models, NeurIPS 2026

[2] Large Language Model Unlearning, NeurIPS 2024

[3] Interaction Context Often Increases Sycophancy in LLMs. CHI 2026

[4] When Truth Is Overridden: Uncovering the Internal Origins of Sycophancy in Large Language Models. AAAI 2026

[5] Sycophancy Mitigation Through Reinforcement Learning with Uncertainty-Aware Adaptive Reasoning Trajectories. EMNLP 2025

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

This paper investigates whether harmful conformity in LLMs can be mitigated through parameter updates. The proposed framework identifies questions that a model answers correctly and confidently in isolation but incorrectly under social influence, and converts these interactions into forget–retain supervision. The framework supports both one-time and periodic updates using existing unlearning methods. Experiments on two model families show substantial reductions in conformity-related errors, with relatively small utility degradation and encouraging preservation of responsiveness to corrective guidance. The main contribution lies in the supervision construction and intervention framework.

Critical strengths

Clear motivation and a practical supervision construction procedure. The paper targets a meaningful failure mode: models abandoning correct answers under misleading social influence. Constructing forget–retain examples from such interactions provides a straightforward way to translate observed failures into training signals.

Substantial improvements with relevant comparisons. The evaluation includes both prompting methods and parameter-update baselines, including SFT and DPO. GA reduces BENCHFORM conformity rates from 28.25% to 9.69% on Llama and from 21.68% to 7.84% on Qwen. Question-disjoint splits and additional seed and decoding experiments strengthen the empirical evidence.

Attention to selective contextual responsiveness. The paper evaluates both resistance to misleading guidance and correction of initially wrong answers. These complementary evaluations provide useful evidence that reduced conformity does not simply result from indiscriminately ignoring social context.

Critical weaknesses

Narrow evaluation scenarios and unclear practical relevance. The evaluations primarily involve simulated unanimous majorities and predefined social-guidance protocols. The paper does not sufficiently establish how frequently such settings arise in practical LLM applications or demonstrate the downstream benefits of reducing conformity in concrete workflows. This limits the practical significance of the reported improvements.

The evaluation does not fully distinguish conformity from contextual interference. Although the paper evaluates both misleading and corrective guidance, some metrics do not directly capture conformity. BF-CR counts any correct-to-incorrect transition without requiring adoption of the majority answer, while RL measures answer stability regardless of correctness. Consequently, improvements may partly reflect greater robustness to contextual interference. Controls with comparable context length and repetition, but without social framing, would help isolate conformity-specific effects.

The periodic-update comparison is confounded by additional training. Periodic methods receive further optimization opportunities beyond the initial update. Without comparisons matched for total training steps and data exposure, the benefits of repeatedly discovering new conformity failures remain difficult to separate from those of additional training.

Several experimental details require clarification. The paper should report complete split sizes, evaluation sample counts, exact model checkpoints, and training durations. It should also clarify whether the Raw-correct evaluation subset is fixed across methods and reconcile the general retain-set definition with the 388-example configuration described in the appendix.

Questions for authors

see weakness

3 comments
AuthorsOpenReview ↗

Response to Reviewer igTL: Part 1

We sincerely thank the reviewer for the insightful review and valuable suggestions!

W1. Narrow evaluation scenarios and unclear practical relevance. The evaluations primarily involve simulated unanimous majorities and predefined social-guidance protocols. The paper does not sufficiently establish how frequently such settings arise in practical LLM applications or demonstrate the downstream benefits of reducing conformity in concrete workflows. This limits the practical significance of the reported improvements.

Response: We follow the suggestion to extend the evaluation to practical LLM applications in concrete workflows. Specifically, we evaluate our method on the mathematical reasoning tasks, M3MAD-Bench, where Multi-Agent Debate (MAD) coordinates multiple agents through structured debate to produce better answers and handle complex reasoning tasks. We use the same group of peer agents to debate with the base model and the unlearned models with SUL, respectively. Note that the forget dataset is still constructed with BENCHFORM.

In addition to final task accuracy, we evaluate whether the target model adopts naturally arising wrong majorities and whether correct peer guidance helps repair initially incorrect answers. These complementary metrics capture the downstream effects of peer disagreement on task performance.

The results indicate that SUL generalizes to multi-agent mathematical reasoning with naturally generated peer responses. Compared with the Base model, GA and NPO improve final accuracy from 25.29% to 43.44% and 41.75%, respectively, while reducing wrong-majority adoption from 76.47% to 66.67% and 61.82%. Both methods also improve initial-wrong correction from 10.50% to 19.90% and 19.12%, demonstrating improved task performance while preserving responsiveness to corrective peer information.

Target modelFinal accuracy (%) ↑Wrong-majority adoption (%) ↓Initial-wrong correction (%) ↑
Base25.2976.4710.50
QD26.4274.1812.62
DPO32.3171.5613.27
SUL (GA)43.4466.6719.90
SUL (NPO)41.7561.8219.12

W2. The evaluation does not fully distinguish conformity from contextual interference. Although the paper evaluates both misleading and corrective guidance, some metrics do not directly capture conformity. BF-CR counts any correct-to-incorrect transition without requiring adoption of the majority answer, while RL measures answer stability regardless of correctness. Consequently, improvements may partly reflect greater robustness to contextual interference. Controls with comparable context length and repetition, but without social framing, would help isolate conformity-specific effects.

Response: Following the suggestion, we further distinguish social-framing effects from general contextual interference through the following controls.

To isolate the additional contribution of social framing, we compare (i) Social-Wrong, where incorrect answers are attributed to agreeing peers; (ii) Non-Social Wrong, which repeats the same incorrect answer content without peer attribution; (iii) a length-matched neutral context; and (iv) Social-Correct, which provides accurate peer guidance. The main comparison keeps the misleading answer, repetition count, placement, and approximate input length comparable while varying the social framing. For cases with an identifiable majority, we also distinguish direct wrong-majority adoption from other incorrect answer changes. Formally, these are different events: Flip=Pr⁡(y^social≠yref∣y^raw=yref)\mathrm{Flip}=\Pr(\hat y^{\mathrm{social}}\ne y^{\mathrm{ref}}\mid\hat y^{\mathrm{raw}}=y^{\mathrm{ref}}) and WrongMajorityAdoption=Pr⁡(y^social=ywrong-majority∣y^raw=yref, wrong majority exists)\mathrm{WrongMajorityAdoption}=\Pr(\hat y^{\mathrm{social}}=y^{\mathrm{wrong\text{-}majority}}\mid\hat y^{\mathrm{raw}}=y^{\mathrm{ref}},\ \text{wrong majority exists})

Context conditionBase Acc (%)GA Acc (%)NPO Acc (%)Base Flip (%)GA Flip (%)NPO Flip (%)
Social-Wrong15.4541.8042.5875.3537.4036.29
Non-Social Wrong50.7664.2466.2124.388.598.42
Neutral context56.6757.1255.7010.807.187.27
Social-Correct85.1583.4983.216.945.115.78
AuthorsOpenReview ↗

Response to Reviewer igTL: Part 2

Response to W2 (cont.): The results indicate that social framing induces a substantially higher flip rate than non-social misleading context (75.35% vs. 24.38%). GA and NPO reduce this gap from 50.97 to 28.81 and 27.87 percentage points, respectively, while maintaining high accuracy under correct guidance. Our evaluation metrics capture complementary aspects of model behavior under social influence. BF-CR follows BENCHFORM and tracks correct-to-incorrect transitions under social context, and RL captures stability relative to the independent response. CL directly measures adoption of the unanimous majority response in the psychology-inspired setting. Correct Guidance and Initial-Wrong Correction additionally measure responsiveness to helpful external information. Combining these new experiments with our existing metrics allows us to isolate conformity-specific effects more directly. The results provide further evidence that our method reduces the additional susceptibility associated with social framing, beyond its improvements in general contextual robustness.

W3. The periodic-update comparison is confounded by additional training. Periodic methods receive further optimization opportunities beyond the initial update. Without comparisons matched for total training steps and data exposure, the benefits of repeatedly discovering new conformity failures remain difficult to separate from those of additional training.

Response: To isolate the effects of additional training, we conduct a supplementary compute-matched comparison based on the periodic evaluation setting of Figure 3. Dynamic Refresh (4 cycles) rebuilds supervision using failures of the current model. To match the total training steps, the Static Replay variant repeatedly optimizes on the initial forget--retain supervision for 4 cycles. To match data exposure, the One-time variant constructs a forget–retain dataset that is 4× larger during the initial cycle, ensuring that the total amount of training data and the number of conformity failures match those of the periodic-update setting. All variants start from the same initial checkpoint and are matched for optimizer steps, while evaluation remains on the same fixed, question-disjoint held-out evaluation subset.

MethodFinal Acc (%) ↑Final CR (%) ↓Avg Acc (%) ↑Avg CR (%) ↓
Dynamic Refresh56.979.1756.419.83
Static Replay50.1715.3354.1312.37
One-time46.8319.6752.2314.87

The results show that Dynamic Refresh substantially outperforms Static Replay, improving final-round accuracy from 50.17% to 56.97% while reducing conformity rate from 15.33% to 9.17%. This provides evidence that periodically refreshing conformity-derived supervision offers additional benefits beyond repeated optimization on static supervision. Additionally, the effects of one-time unlearning may gradually diminish over time as the model encounters increasing conformity pressure.

AuthorsOpenReview ↗

Response to Reviewer igTL: Part 3

W4. Several experimental details require clarification. The paper should report complete split sizes, evaluation sample counts, exact model checkpoints, and training durations. It should also clarify whether the Raw-correct evaluation subset is fixed across methods and reconcile the general retain-set definition with the 388-example configuration described in the appendix.

Response: We would like to clarify the experimental configuration as follows. Our BENCHFORM experiments use a question-disjoint split comprising 2,639 training questions and 660 held-out evaluation questions. All forget--retain supervision, including periodic updates, is constructed exclusively from the training split, while the same held-out questions and evaluation protocols are used across compared methods. We follow the original BENCHFORM metric definitions: BF-Accuracy is evaluated over the full test set, whereas BF-CR and BF-IR are conditioned on each evaluated model's own Raw-correct subset. For the supplementary fixed-cohort analysis, 361 questions answered correctly by the original model are used consistently across the compared conditions.

Regarding the 388 examples, Section 3.2 describes the general retain-supervision construction over (F∪R)(\mathcal{F}\cup\mathcal{R}), whereas the 388 conformity-derived pairs refer to the instantiated supervision configuration used in the expanded model-level comparison. Each pair associates a socially induced erroneous response with its reference-supported target under the same context. Thus, the general formulation and the reported experimental configuration describe different levels of the same supervision construction.

The model configurations and training hyperparameters are already detailed in Appendix A.5 (Table 5, Page 15), with retain-only fine-tuning and periodic unlearning further specified in Appendices A.6--A.7 (Page 15). The pretrained model checkpoints used in our experiments are Llama-3-8B-Instruct and Qwen2.5-14B-Instruct. The measured training costs are 1.95, 5.87, 3.72, 2.74, and 3.11 GPU-hours for SFT-Retain, Balanced SFT, DPO, SUL (GA), and SUL (NPO), respectively.

We sincerely appreciate your time in reading our rebuttal. Upon receiving your further comments, we will carefully incorporate the new content into the revised PDF.

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

The paper points out a phenomenon where models answer correctly independently, but change to an incorrect answer after being given social context. To solve this problem, the authors construct two types of training samples as forget examples and retain examples. They construct forget examples from these erroneous responses and retain examples from reference-supported answers, then apply existing unlearning objectives through LoRA.

Critical strengths
  1. This paper formulates LLM conformity mitigation as an unlearning problem, which is a new idea for solving the conformity problem of LLMs.

  2. The methods that the paper proposes are effective. The question split and results from three training seeds strengthen the evidence that the effect extends beyond memorizing the training traces.

Critical weaknesses
  1. Lack of analysis of which questions are susceptible to conformity. Figure 1 illustrates the answer distribution of LLM shifting under social influence, where the overlapped area indicates some samples that cannot be flipped by social context. A stratified analysis of flip rates is needed to show where the social context succeeds and where it fails.

  2. The motivation of using unlearning is not clear. The paper constructs a dispreferred socially induced answer and a reference-supported answer for the same context, but does not clearly explain why this behavioral correction should be formulated as unlearning rather than supervised fine-tuning or preference optimization.

  3. In Table 1, GA and NPO achieve lower BF-CR. The comparison results indicate a trade-off rather than uniform superiority. The methods also use different training mixtures, which complicates attribution.

  4. In the cross-protocol MMLU experiment (Appendix C, Table 7), QD and DPO achieve lower conformity levels than GA and NPO. The advantage in conformity reduction does not consistently transfer across different protocols.

Questions for authors

Please check my weakness section.

2 comments
AuthorsOpenReview ↗

Response to Reviewer VDjJ: Part 1

We sincerely thank the reviewer for the insightful comments and the deep understanding of our method.

W1. Lack of analysis of which questions are susceptible to conformity. Figure 1 illustrates the answer distribution of LLM shifting under social influence, where the overlapped area indicates some samples that cannot be flipped by social context. A stratified analysis of flip rates is needed to show where the social context succeeds and where it fails.

Response: We follow the suggestion to add a stratified analysis of flip rates. Specifically, we analyze which initially correct held-out questions flip under misleading social context and how this behavior relates to their initial confidence. We analyze a fixed cohort of Base Raw-correct held-out questions under the wrong guidance protocol and define a flip as an initially correct answer becoming incorrect at least once across rounds 1--5. We stratify questions into low-, medium-, and high-confidence groups based on their pre-social-context confidence and report the conditional flip rate within each group.

The results indicate that answers with low confidence are more likely to flip under misleading social context. Intuitively, models are more likely to defer to social signals when their internal predictions are uncertain. In contrast, high-confidence answers are generally more resistant to social influence. However, an answer with low confidence does not necessarily mean it is incorrect. Several studies find that confidence should not be equated with correctness. Our analysis focuses exclusively on initially correct answers.

This finding further motivates our unlearning approach. Social conformity is not necessarily caused by a lack of factual knowledge. Instead, it reflects the model's tendency or pattern of following socially presented opinions over its own knowledge. Our unlearning approach explicitly targets this undesirable behavioral tendency by suppressing conformity-related response patterns while still preserving the ability to learn from in-context information.

Confidence GroupQuestions (N)Flipped (N)Flip Rate (%) ↓
Low1217662.81
Medium1205545.83
High1201815.00

W2. The motivation of using unlearning is not clear. The paper constructs a dispreferred socially induced answer and a reference-supported answer for the same context, but does not clearly explain why this behavioral correction should be formulated as unlearning rather than supervised fine-tuning or preference optimization.

Response: This is an insightful question. In the setting of conformity mitigation, unlearning explicitly suppresses context-triggered conformity behaviors, whereas SFT primarily reinforces desired responses and preference optimization (e.g., DPO) optimizes relative preferences between candidate responses.

In Equation (4), we construct a forget set that captures the pattern of socially induced response shifts. This forget set explicitly identifies the undesirable behavioral tendency that we aim to mitigate. Our unlearning method directly targets socially induced incorrect responses under social pressure. At the same time, our retain dataset aims to preserve useful responsiveness to corrective context.

SFT may encourage the model to reproduce desired answers in the training scenarios, but it does not explicitly suppress observed conformity-induced errors through a forgetting objective. Similarly, preference optimization establishes a relative preference between candidate responses, which may improve response quality for labeled questions but does not explicitly eliminate the behavioral association between social context and conformity. Mathematically, unlike SFT and DPO, our formulation explicitly incorporates a separate forgetting objective to suppress observed conformity-induced responses while preserving reference-supported behavior through retain supervision.

We remark that both unlearning and SFT belong to the model-level intervention. In the experiment (Table 1), we have compared our unlearning method against SFT, Balanced SFT and DPO. The results suggest that explicitly suppressing conformity-induced errors can reduce susceptibility to misleading social influence while maintaining responsiveness to corrective guidance. This targeted behavioral intervention provides a distinction in optimization objectives between unlearning and conventional supervised fine-tuning.

AuthorsOpenReview ↗

Response to Reviewer VDjJ: Part 2

W3. In Table 1, GA and NPO achieve lower BF-CR. The comparison results indicate a trade-off rather than uniform superiority. The methods also use different training mixtures, which complicates attribution.

Response: We would like to clarify that lower BF-CR (Equation 18) indicates stronger resistance to misleading social influence. Therefore, a lower BF-CR is a favorable result rather than a performance degradation.

Compared with the strongest baseline (Balanced SFT), SUL (GA) substantially reduces BF-CR on Llama-3-8B from 20.54% to 9.69%, while achieving higher BF-IR (22.30% vs.~20.29%) and maintaining competitive BF-Accuracy (57.24% vs.~62.31%). More importantly, on Qwen-2.5-14B, GA simultaneously improves BF-Accuracy (63.47% vs.~57.31%), reduces BF-CR (7.84% vs.~18.75%), and increases BF-IR (28.39% vs.~20.31%). These results demonstrate that stronger conformity resistance does not necessarily come at the expense of task accuracy or sustained correctness.

We would like to clarify that our contribution is to formulate LLM conformity mitigation as an unlearning problem. Inherently, unlearning is different from SFT and DPO in terms of training mixtures (e.g., w/o forget dataset). The target of our paper is to verify that suppressing the behavior of shifting responses can mitigate conformity, not to prove the superiority of the loss function over SFT. The training configurations of the compared methods are specified in Appendix A.5–-A.6, and all reported results follow the same held-out evaluation protocols. Taken together, these comparisons support the effectiveness of conformity-targeted unlearning in mitigating misleading social influence while preserving reliable task performance.

W4. In the cross-protocol MMLU experiment (Appendix C, Table 7), QD and DPO achieve lower conformity levels than GA and NPO. The advantage in conformity reduction does not consistently transfer across different protocols.

Response: We remark that the conformity level should not be interpreted alone. Table 7 does report lower conformity levels for QD and DPO. However, the other two evaluation metrics (Correct Acc and IW-Corr) of QD and DPO basically collapse. With QD and DPO, the model can not correct an initially wrong answer. Our objective is selective responsiveness: resisting misleading opinions while still using reliable external information. The cross-protocol results are best understood by jointly evaluating resistance to misleading opinions and responsiveness to accurate peer guidance.

Table 7 in Appendix C reveals this distinction precisely. QD and DPO obtain CL of 2.17% and 2.26%, respectively, but their Initial-Wrong Correction rates are only 3.85% and 7.69%. GA and NPO retain relatively low CL (4.35% and 6.52%, compared with 13.04% for the base model) while achieving substantially higher Initial-Wrong Correction (53.86% and 46.15%). These complementary metrics show the cross-protocol advantage of SUL in preserving corrective responsiveness while substantially reducing conformity relative to the base model.

Thank you again for your time. Upon receiving your further comments, we will incorporate the new experimental results into the revised manuscript accordingly.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.