- 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
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.
| Method | WG Error (%) ↓ | IW-Corr (%) ↑ | WG Acc (%) ↑ |
|---|---|---|---|
| Base | 75.35 | 81.64 | 15.45 |
| QD | 46.65 | 27.02 | 34.39 |
| Independent-first | 72.58 | 62.94 | 36.37 |
| Few-shot ICL | 38.50 | 41.83 | 36.81 |
| Personalized Prompt | 52.91 | 67.89 | 37.42 |
| SUL (GA) | 37.40 | 84.88 | 41.80 |
| SUL (NPO) | 36.29 | 85.50 | 42.58 |
| Method | Final Acc (%) ↑ | Wrong-majority Adoption (%) ↓ | IW-Corr (%) ↑ |
|---|---|---|---|
| Base | 25.29 | 76.47 | 10.50 |
| SUL (GA) | 43.44 | 66.67 | 19.90 |
| SUL (NPO) | 41.75 | 61.82 | 19.12 |
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 model | Final accuracy (%) ↑ | Wrong-majority adoption (%) ↓ | Initial-wrong correction (%) ↑ |
|---|---|---|---|
| Base | 25.29 | 76.47 | 10.50 |
| QD | 26.42 | 74.18 | 12.62 |
| DPO | 32.31 | 71.56 | 13.27 |
| SUL (GA) | 43.44 | 66.67 | 19.90 |
| SUL (NPO) | 41.75 | 61.82 | 19.12 |
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
Response to Reviewer g4sq: Part 1
We thank the reviewer for the thoughtful question and the constructive feedback.
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.