acceptodds
Under review as a conference paper at ICLR 2027

LLM Persona Unlearning

Abstract

Pre-training equips LLMs with a broad repertoire of behavioral patterns associated with roles, styles, values, and goals. Post-training teaches conditional enactment and makes a helpful *Assistant* the default, but it does not erase alternative modes from the weights; explicit prompts can therefore elicit personas that repeatedly shape judgment, language, and action. In open-weight settings, runtime controls can be removed, motivating *persona unlearning*: a weight-level edit that makes a designated persona difficult to elicit and enact on unseen contexts. We introduce PersonaUnlearnBench, a model-specific paired benchmark spanning six LLMs from three families and five personas, with aligned forget/retain sets, held-out instruction paraphrases, and four-axis evaluation. The benchmark shows that standard unlearning methods cannot reliably erase the target persona without sacrificing meaningful generation or general utility. We therefore propose PaCE, which compares target and desirable responses to the same questions to locate an internal behavior direction, then trains target-prompt states away from the target mode and toward the matched desirable response. Experiments show that PaCE consistently suppresses target personas with complete gradable coverage, high response quality, and useful counterpart behavior, at moderate utility cost. These results establish persona unlearning as a distinct behavior-level editing problem and a practical route toward persistent control of latent LLM response policies.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM unlearning, large language models, persona, machine unlearning, safety alignment

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_hRxYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PACE, a persona unlearning method that aims to remove undesirable behavioral patterns from LLMs while preserving their general capabilities and response quality. The main idea is to identify persona-specific directions in the hidden representation space by contrasting target and desirable responses to the same questions. PACE then fine-tunes the model to suppress the target persona direction while aligning the remaining representations with those of desirable counterparts. The paper also introduces PersonaUnlearnBench, covering six LLMs and five personas. Experiments show that PACE generally outperforms existing unlearning baselines in suppressing undesirable personas while maintaining response quality, although some degradation in general utility remains.

Critical strengths
  1. The paper introduces persona unlearning as a distinct task from conventional knowledge unlearning, focusing on removing undesirable behavioral patterns rather than specific factual content. This provides an interesting perspective on improving the safety and controllability of LLMs.
  2. PACE is relatively simple and intuitive, using persona directions to suppress undesirable behaviors through representation-level optimization. The counterpart anchoring mechanism also helps preserve desirable behaviors and response quality during unlearning.
  3. The paper evaluates six LLMs from three model families across five personas and compares PACE with several existing unlearning baselines. It also includes ablation studies and robustness evaluations under jailbreak attacks, multi-turn interactions, and cross-lingual settings, which help support the effectiveness of the method.
Critical weaknesses
  1. The methodological novelty is relatively limited. PACE builds on existing Persona Vector and Activation Steering techniques, using persona directions and representation-level optimization to suppress undesirable behaviors. Although counterpart anchoring is introduced, the overall approach is closely related to existing representation editing methods.
  2. Persona unlearning can still degrade general model capabilities. Although PACE effectively suppresses target personas, it causes noticeable drops in general utility in some settings. For example, on Llama-3.1-8B with the Impolite persona, GU decreases from 68.65 to 63.75, raising concerns about the trade-off between persona removal and capability preservation.
  3. The long-term robustness of persona unlearning is not sufficiently evaluated. The paper tests jailbreak attacks, multi-turn interactions, and cross-lingual settings, but does not examine whether forgotten personas can re-emerge after subsequent fine-tuning. Moreover, the forgetting performance decreases under iterative jailbreak attacks for some personas, raising concerns about the persistence of the unlearning effect.
Questions for authors

Have the authors considered evaluating whether the forgotten personas can re-emerge after subsequent fine-tuning? Since PACE aims to achieve persistent persona unlearning through weight-level modifications, such experiments would help assess the long-term effectiveness and robustness of the proposed method.

6 comments
AuthorsOpenReview ↗

Response to Reviewer hRxY (Part 1/3)

Thank you for your positive feedback on the relevance of persona unlearning to LLM safety and controllability, the simplicity of PaCE, and the breadth of our experiments. Below we address each of your weaknesses and questions point-by-point.


Weakness 1: Relationship to existing representation-editing methods

“The methodological novelty is relatively limited. PACE builds on existing Persona Vector and Activation Steering techniques, using persona directions and representation-level optimization to suppress undesirable behaviors. Although counterpart anchoring is introduced, the overall approach is closely related to existing representation editing methods.”

Thank you for raising this point. We clarify that: (1) unlike prior LLM unlearning methods that focus on memorized content, our work is the first to formulate and study weight-level behavioral unlearning, combining the linear representation idea of persona vectors with a simple, direct, and hook-free parameter-editing objective; and (2) our controlled component study (Table C.1) verifies that PaCE’s three specific design choices—pre-generation prompt-state readout, per-question counterpart threshold, and orthogonal anchoring—each provide clear empirical gains over simpler alternatives.

  1. Distinction from content unlearning and inference-time steering

    Existing LLM unlearning methods remove specific memorized facts or texts, which often fails to suppress a recurring behavioral persona without degrading generation quality (Sec. 3.5). Meanwhile, Activation Steering and Persona Vectors [1] primarily study inference-time hooks (which can be removed in open-weight settings) or response-token regularization.

    To achieve persistent behavioral unlearning in model weights without runtime hooks, PaCE uses the frozen Original model’s paired counterpart on the same question to define a two-part representation edit at the final prompt token before generation:

    L=Ei[sit−sic+m]++λ Ei[1−cos⁡(P⊥hit,P⊥hi,0c)],[z]+=max⁡(z,0).\mathcal L = \mathbb E_i[s_i^t-s_i^c+m]_ {+} + \lambda\,\mathbb E_i[1-\cos(P_\perp h_i^t,P_\perp h_{i,0}^c)],\qquad[z]_+=\max(z,0).

    Here, the erasure term pushes the target projection sit=v⊤hits_i^t = v^\top h_i^t below the question-specific counterpart projection sic=v⊤hi,0cs_i^c = v^\top h_{i,0}^c by margin mm, while the anchor aligns the remaining representation strictly in the orthogonal subspace P⊥=I−vv⊤P_\perp = I - v v^\top so preservation does not conflict with erasure.

  2. Controlled evaluation of readout position, threshold calibration, and orthogonal anchoring

    We compare five variants (A0–A4) on Gemma-4-12B / Sycophantic under the same 400 paired examples, 75 optimizer updates, and fixed parameters (m=1,λ=1m=1, \lambda=1):

    • A0 — PaCE control: pre-generation final-prompt-token readout, per-question counterpart threshold sics_i^c, and orthogonal anchor P⊥P_\perp.
    • A1 — Answer-token readout: uses the mean answer-token state under teacher-forced constructive completions.
    • A2 — Global threshold: replaces sics_i^c with the dataset mean sˉc=N−1∑isic\bar s^c=N^{-1}\sum_i s_i^c.
    • A3 — Full-state anchor: replaces the orthogonal anchor with Ei[1−cos⁡(hit,hi,0c)]\mathbb E_i[1-\cos(h_i^t,h_{i,0}^c)].
    • A4 — Combined alternative: uses both the global threshold (A2) and the full-state anchor (A3).

    Table C.1: Controlled alternatives to PaCE’s design (Gemma-4-12B, Sycophantic).

    ConditionTF ↑CP ↑RQ ↑GU ↑
    A0 — PaCE control59.2080.3090.6072.66
    A1 — Answer-token readout55.7079.8076.2071.89
    A2 — Global threshold (sˉc\bar s^c)54.3081.1084.6073.09
    A3 — Full-state anchor (no P⊥P_\perp)48.2080.6085.5072.92
    A4 — Global threshold + full anchor51.0079.9084.2073.09

    Table C.1 shows that A0 achieves the highest TF and RQ among all five variants:

    • Pre-generation readout (A0 vs. A1) improves all four metrics: +3.50 TF, +0.50 CP, +14.40 RQ, and +0.77 GU.
    • Per-question calibration (A0 vs. A2) gains +4.90 TF and +6.00 RQ over a global threshold.
    • Orthogonal anchoring (A0 vs. A3) yields the largest TF gain (+11.00 TF and +5.10 RQ over full-state anchoring, and +8.20 TF / +6.40 RQ over A4).

    We have included these results and clarifications in the revised manuscript (Sec. 1, Sec. 5.4 Table 3, and Appx. B.3).

AuthorsOpenReview ↗

Response to Reviewer hRxY (Part 2/3)

Weakness 2: General-utility loss on Llama-3.1-8B–Impolite

“Persona unlearning can still degrade general model capabilities. Although PACE effectively suppresses target personas, it causes noticeable drops in general utility in some settings. For example, on Llama-3.1-8B with the Impolite persona, GU decreases from 68.65 to 63.75, raising concerns about the trade-off between persona removal and capability preservation.”

Thank you for highlighting this trade-off. We clarify that: (1) on Llama–Impolite, the 4.90-point GU change accompanies complete target suppression (+96.30 TF) and a +19.10-point improvement in response quality, (2) core mathematical and scientific reasoning on GSM8K and ARC-Challenge remain well preserved, and (3) at nearly the exact same observed GU (63.75 vs. 63.81), PaCE outperforms direct task-matched SFT by +14.08 TF and +21.45 RQ.

  1. Substantial behavioral and response-quality gains

    Table C.2: Llama-3.1-8B–Impolite behavioral gains and utility cost.

    MetricMeaningOriginalSFT (Matched)PaCEPaCE vs. OrigPaCE vs. SFT
    TF ↑Target-persona suppression3.7085.92100.00+96.30+14.08
    CP ↑Constructive-counterpart expression74.6075.6073.60−1.00−2.00
    RQ ↑Quality of persona-task answers74.1071.7593.20+19.10+21.45
    GU ↑Mean general-capability score68.6563.8163.75−4.90−0.06

    As shown in Table C.2, TF increases by 96.30 points and RQ increases by 19.10 points, while CP remains within 1.00 point of Original.

  2. Breakdown across general-utility capabilities

    Table C.3: General-utility breakdown on Llama-3.1-8B–Impolite.

    BenchmarkCapabilityTest itemsOriginalPaCEChange
    GSM8KMathematical word-problem reasoning1,31984.1583.62−0.53
    ARC-ChallengeScience question answering and reasoning1,17253.7551.11−2.64
    MMLUBroad multitask knowledge14,04263.1257.22−5.90
    IFEvalFollowing explicit formatting instructions54173.5763.03−10.54
    GUEqual-weight mean—68.6563.75−4.90

    Table C.3 shows that mathematical reasoning on GSM8K is largely unchanged (84.15 → 83.62; −0.53 points) and ARC-Challenge changes by only −2.65 points, whereas the decrease is concentrated in strict formatting adherence on IFEval.

  3. Comparison with direct SFT at nearly identical observed GU

    When we train Llama-3.1-8B on the same 400 pairs and 75 updates using task-matched SFT (training to produce the constructive response under the target prompt), SFT yields virtually the same GU (63.81 vs. PaCE’s 63.75), yet PaCE achieves:

    • TF 100.00 vs. 85.92 (+14.08 points), and
    • RQ 93.20 vs. 71.75 (+21.45 points).

    Moreover, across all 12 model–persona settings (Appx. F.8, Table 13), PaCE achieves higher TF than both SFT and DPO in 12/12 settings and higher RQ in 10/12 settings, and on Qwen3.5-9B it simultaneously surpasses both SFT and DPO on TF, RQ, and GU across all four personas.

AuthorsOpenReview ↗

Response to Reviewer hRxY (Part 3/3)

Weakness 3: Persistence after fine-tuning and residual accessibility under attacks

“The long-term robustness of persona unlearning is not sufficiently evaluated. The paper tests jailbreak attacks, multi-turn interactions, and cross-lingual settings, but does not examine whether forgotten personas can re-emerge after subsequent fine-tuning. Moreover, the forgetting performance decreases under iterative jailbreak attacks for some personas, raising concerns about the persistence of the unlearning effect.”

Thank you for distinguishing prompt-level robustness from persistence under subsequent training. We address this by adding a downstream full-parameter fine-tuning experiment on FinQA [1] without persona supervision (Table C.4). Three personas retain their exact starting TF across all measured checkpoints, and all four maintain a large endpoint advantage over the equally fine-tuned Original model.

  1. Downstream full-parameter adaptation on FinQA without persona supervision

    We fine-tune the Original and four PaCE-edited Llama-3.1-8B checkpoints for 128 full-parameter optimizer updates on 1,024 FinQA examples [1] (numerical reasoning over financial reports). Training uses standard supervised loss alone, with neither persona data nor the PaCE objective. We evaluate the same 50 target-persona questions per setting at steps 0, 32, and 128 (0.5-point TF resolution).

    Table C.4: TF (↑) during downstream full-parameter FinQA fine-tuning.

    PersonaOriginal: 0Original: 32Original: 128PaCE: 0PaCE: 32PaCE: 128Endpoint Gain
    Sycophantic0.01.51.094.094.094.0+93.0
    Hallucinating15.522.519.586.585.080.0+60.5
    Impolite2.06.04.0100.0100.0100.0+96.0
    Apathetic66.576.069.0100.0100.0100.0+31.0
  2. Suppression persists across downstream weight updates

    • Sycophantic remains at 94.0, while Impolite and Apathetic remain at 100.0 across all three checkpoints (steps 0, 32, and 128).
    • Hallucinating retains TF 80.0 at step 128, remaining +60.5 points above the equally fine-tuned Original model (19.5).
    • Combined with our multi-turn (86.5–100.0 TF across 5 turns), cross-lingual (100.0 TF across 5 languages), and iterative Attack@5 evaluations (where Impolite, Apathetic, and Evil retain 100.0, 99.0, and 91.5 TF, and Sycophantic/Hallucinating still substantially outperform all baselines), these results show that PaCE’s weight edit remains durable well beyond the immediate post-unlearning checkpoint. We have added this study to the revised manuscript (Sec. 5.3, Table 2 and Appx. F.9).

[1] Chen et al. “FinQA: A Dataset of Numerical Reasoning over Financial Data”. In EMNLP, 2021.


Question: Can the forgotten personas re-emerge after subsequent fine-tuning?

“Have the authors considered evaluating whether the forgotten personas can re-emerge after subsequent fine-tuning? Since PACE aims to achieve persistent persona unlearning through weight-level modifications, such experiments would help assess the long-term effectiveness and robustness of the proposed method.”

Yes, we have added a direct test of subsequent full-parameter fine-tuning (Table C.4 in Weakness 3, included in Sec. 5.3, Table 2 and Appx. F.9 of the revised paper). After 128 full-parameter updates on 1,024 FinQA examples without persona supervision or the PaCE loss, three edited personas (Sycophantic at 94.0, Impolite at 100.0, and Apathetic at 100.0) retain their exact starting TF at all checkpoints, and Hallucinating retains 80.0 TF (+60.5 points above the fine-tuned Original’s 19.5), demonstrating strong persistence under downstream weight adaptation.

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

In this paper, the authors studies how to suppress undesirable LLM personas even when explicitly requested. It proposes PACE, which combines contrastive persona directions with an erasure objective and an orthogonal anchoring loss. Experiments cover five persona traits across six Llama, Qwen, and Gemma models ranging from 8B to 70B parameters. PACE substantially suppresses target personas, although some settings show reduced general utility and weaker robustness under adversarial prompting.

Critical strengths
  1. The paired target–counterpart construction is well aligned with the goal of selective behavioral modification.
  2. Reporting target forgetting, counterpart preservation, response quality, and general utility separately helps expose degenerate solutions and preservation costs.
  3. The evaluation covers multiple models and includes adversarial elicitation, multi-turn interaction, translated prompts, and newly constructed prompts.
Critical weaknesses
  1. The methodological contribution appears incremental.

PACE’s direction extraction closely follows existing persona-vector methodology: contrasting responses, computing activation differences, and selecting intervention layers through steering. Persona Vectors also investigates training-time interventions, including projection-based regularization (Chen et al., 2025, Appendix J.5).

PACE differs through its post-hoc parameter editing, counterpart-calibrated erasure threshold, and orthogonal anchoring. However, the experiments do not sufficiently establish the additional value of these specific choices. The component ablation supports combining the two losses in the tested setting, but does not demonstrate an advantage over simpler alternative objectives. Consequently, the benchmark contribution is clearer than the methodological advance.

  1. Missing task-matched SFT/DPO baselines leave the central empirical question unresolved.

The dataset already provides undesirable and constructive responses to the same questions. This enables two natural baselines:

  • SFT: train the model to produce the constructive response under the target-persona instruction.
  • DPO: under that same instruction, prefer the constructive response over the undesirable response.

These baselines directly address the evaluated objective: responding constructively even when prompted to exhibit the undesirable persona. GradDiff’s retain loss does not substitute for this comparison, because preserving constructive responses under constructive instructions is a different conditional task.

PACE involves parameter optimization, so output-level SFT/DPO are relevant competing approaches despite using different supervision. The current comparisons show advantages over the selected unlearning objectives, but do not establish whether the proposed representation decomposition improves upon straightforward paired behavioral training.

Questions for authors

Relative to existing persona-direction regularization, which component is responsible for the improvement: the pre-generation intervention point, counterpart-calibrated threshold, or orthogonal anchor?

others see above weakness.

5 comments
AuthorsOpenReview ↗

Response to Reviewer GaEX (Part 1/2)

Thank you for your constructive review and for recognizing our paired target–counterpart construction, four-axis evaluation, and robustness evaluations. Below we address each of your weaknesses and questions point-by-point.


Weakness 1: Methodological contribution beyond existing persona-vector methods

Thank you for raising this point. We clarify that: (1) unlike prior LLM unlearning methods that erase memorized content, our work is the first to formulate and address weight-level behavioral unlearning, adapting the linear representation insight of persona vectors into a simple, direct, and hook-free parameter edit; and (2) our new controlled study (Table B.1) confirms that each of PaCE’s three specific design choices—pre-generation prompt-state readout, per-question counterpart threshold, and orthogonal anchoring—contributes substantially to the final performance.

  1. From content unlearning and activation steering to weight-level behavioral unlearning

    Conventional LLM unlearning methods target static memorized content (specific facts, records, or text sequences), which can leave a recurring behavioral policy intact or cause generation collapse when applied to personas (Sec. 3.5). Conversely, activation steering and Persona Vectors [1] primarily rely on inference-time hooks (which can be removed in open-weight deployments) or response-token regularization.

    To address weight-level behavioral unlearning, we draw on the persona-vector idea and design a simple, direct, question-conditioned objective that permanently edits model weights without inference hooks:

    • Pre-generation prompt-state intervention: edits the final prompt token state before generation begins, redirecting the behavioral trajectory at its entry point rather than penalizing surface tokens.
    • Counterpart-calibrated threshold (sic−ms_i^c - m): uses the frozen Original model’s matched counterpart projection on the same question as a natural stopping boundary, halting suppression once benign expression is reached.
    • Orthogonal subspace anchoring (P⊥=I−vv⊤P_\perp = I - v v^\top): aligns the remaining representation with the counterpart strictly outside the erased persona direction vv, preventing the preservation term from conflicting with erasure.
  2. Controlled ablation (A0–A4) establishes the added value of each design choice

    We compare five controlled variants (A0–A4) on Gemma-4-12B / Sycophantic under the exact same 400 paired examples, 75 optimizer updates, and fixed parameters (m=1,λ=1m=1, \lambda=1). With unit persona direction vv, orthogonal projector P⊥=I−vv⊤P_\perp=I-vv^\top, target state hith_i^t, frozen counterpart state hi,0ch_{i,0}^c, and projections sit=v⊤hit,sic=v⊤hi,0cs_i^t=v^\top h_i^t, s_i^c=v^\top h_{i,0}^c, the control objective (A0) is:

    L=Ei[sit−sic+m]++λ Ei[1−cos⁡(P⊥hit,P⊥hi,0c)],[z]+=max⁡(z,0).\mathcal L = \mathbb E_i[s_i^t-s_i^c+m]_ {+} + \lambda\,\mathbb E_i[1-\cos(P_\perp h_i^t,P_\perp h_{i,0}^c)],\qquad[z]_+=\max(z,0).

    We test four targeted replacements against A0:

    • A0 — PaCE control: pre-generation final-prompt-token readout, per-question counterpart threshold sics_i^c, and orthogonal anchor P⊥P_\perp.
    • A1 — Answer-token readout: replaces the prompt-state readout with the mean answer-token state under teacher-forced constructive completions.
    • A2 — Global threshold: replaces sics_i^c in the erasure term with the dataset mean sˉc=N−1∑isic\bar s^c=N^{-1}\sum_i s_i^c.
    • A3 — Full-state anchor: replaces the orthogonal anchor with full-state alignment Ei[1−cos⁡(hit,hi,0c)]\mathbb E_i[1-\cos(h_i^t,h_{i,0}^c)].
    • A4 — Combined alternative: applies both the global threshold (A2) and the full-state anchor (A3).

    Table B.1: Controlled alternatives to PaCE’s design (Gemma-4-12B, Sycophantic).

    ConditionTF ↑CP ↑RQ ↑GU ↑
    A0 — PaCE control59.2080.3090.6072.66
    A1 — Answer-token readout55.7079.8076.2071.89
    A2 — Global threshold (sˉc\bar s^c)54.3081.1084.6073.09
    A3 — Full-state anchor (no P⊥P_\perp)48.2080.6085.5072.92
    A4 — Global threshold + full anchor51.0079.9084.2073.09

    Table B.1 shows that A0 achieves the highest TF and highest RQ among all five variants, confirming the value of each component:

    • Pre-generation readout (A0 vs. A1) improves all four metrics simultaneously: +3.50 TF, +0.50 CP, +14.40 RQ, and +0.77 GU, showing that pre-generation editing yields a large response-quality gain alongside stronger suppression.
    • Question-specific calibration (A0 vs. A2) improves both suppression and quality: +4.90 TF and +6.00 RQ compared with a single global threshold.
    • Orthogonal anchoring (A0 vs. A3) provides the largest single TF gain: +11.00 TF and +5.10 RQ compared with full-state anchoring (and +8.20 TF / +6.40 RQ over A4), while CP and GU vary by at most 0.80 and 0.43 points.

    We have incorporated this controlled ablation into the revised manuscript (Sec. 5.4, Table 3).

AuthorsOpenReview ↗

Response to Reviewer GaEX (Part 2/2)

Weakness 2: Missing task-matched SFT/DPO baselines

Thank you for specifying these baselines. We have implemented both conditional tasks exactly as described under the same 400 paired examples and 75 optimizer updates as PaCE. PaCE achieves higher TF than both SFT and DPO in all 12/12 settings, and higher RQ than both in 10/12 settings.

  1. Setup of the task-matched SFT and DPO baselines

    Under the target-persona instruction, SFT trains the model to produce the paired constructive response, and DPO prefers the constructive response over the undesirable target response using the frozen Original model as reference (without an SFT warm-up). Both are evaluated on the same 250 target-context and 250 counterpart-context responses and the full GU suite.

    Table B.2: Task-matched direct behavioral baselines.

    ModelMethodSycophanticHallucinatingImpoliteApathetic
    TF↑CP↑RQ↑GU↑TF↑CP↑RQ↑GU↑TF↑CP↑RQ↑GU↑TF↑CP↑RQ↑GU↑
    Llama-3.1-8BOriginal3.6071.9070.2068.6513.1068.5064.9568.653.7074.6074.1068.6541.2078.6073.5068.65
    SFT81.5682.5090.5566.0483.7472.0063.0065.0185.9275.6071.7563.8191.0776.5066.7565.56
    DPO90.3172.6088.2567.3175.2669.4056.0041.0880.3876.1084.8564.7582.6478.3077.0066.65
    PaCE100.0071.1093.0065.7893.6067.7056.4563.87100.0073.6093.2063.75100.0076.4091.8564.27
    Qwen3.5-9BOriginal0.5081.0078.4568.685.2087.2078.0068.680.0074.2070.8568.6821.4792.4069.7568.68
    SFT62.71100.0088.7064.1485.3295.7074.2567.0780.2385.2086.1564.9088.4697.9068.7061.76
    DPO69.4898.7079.3064.7875.0984.8075.0561.4786.1781.1086.1066.6681.9395.9071.4563.29
    PaCE79.6081.8095.7568.5494.2082.6080.4569.02100.0074.0095.5067.00100.0086.6084.4565.10
    Gemma-4-12BOriginal0.2080.7078.9074.027.9081.9080.3574.020.0072.0073.2574.0233.4087.3074.1574.02
    SFT59.1298.5077.6570.0264.7394.0068.8067.7983.9181.7368.8567.9487.2497.0063.5069.93
    DPO67.8599.0073.6572.7755.6889.0068.8571.1891.6282.3075.5071.2780.6796.5063.9072.60
    PaCE77.9080.0093.4571.6175.5077.6067.7069.31100.0072.9094.5568.79100.0087.8084.2068.30
  2. Key empirical takeaways

    • Consistent suppression gains (12/12 settings): Mean TF rises from 79.50 (SFT) and 78.09 (DPO) to 93.40 (PaCE), with PaCE strictly higher in every setting.
    • Higher response quality and matched-utility gains: PaCE achieves higher RQ than both baselines in 10/12 settings (+11.83 mean RQ over SFT, +10.89 over DPO) while maintaining mean CP close to Original (77.68 vs. 79.19). On Llama–Impolite, where GU differs from SFT by only 0.06 points (63.75 vs. 63.81), PaCE improves TF by +14.08 points and RQ by +21.45 points.
    • Joint improvements on Qwen3.5-9B: PaCE exceeds both SFT and DPO on TF, RQ, and GU in all four persona settings (e.g., for Hallucinating, 94.20 TF, 80.45 RQ, and 69.02 GU vs. SFT’s 85.32/74.25/67.07 and DPO’s 75.09/75.05/61.47).

    We have added these results to Appx. F.8 (Table 13) of the revised manuscript.


Question: Which design choice is responsible for the improvement?

“Relative to existing persona-direction regularization, which component is responsible for the improvement: the pre-generation intervention point, counterpart-calibrated threshold, or orthogonal anchor?”

All three choices contribute in our controlled study, with complementary strengths (Table B.1 in Weakness 1):

  1. Pre-generation readout (A0 vs. A1): improves RQ by +14.40 points and TF by +3.50 points (along with +0.50 CP and +0.77 GU), providing the largest individual RQ gain.
  2. Counterpart-calibrated threshold (A0 vs. A2): improves TF by +4.90 points and RQ by +6.00 points over a global threshold.
  3. Orthogonal anchor (A0 vs. A3): improves TF by +11.00 points and RQ by +5.10 points over full-state anchoring, providing the largest individual TF gain.

Thus, the improvement is not merely a trade-off caused by the readout position: A0 achieves both the highest TF and highest RQ among all five variants.

AuthorsOpenReview ↗

Follow-up on Rebuttal Response for Submission 1568

Dear Reviewer GaEX,

Your suggestions helped us strengthen the evaluation of PaCE’s design and its comparison with direct behavioral training. Thank you for these constructive comments. Both requested studies are now included in the updated manuscript, with revisions highlighted in blue:

  • Design choices (W1/Q): Our controlled study on Gemma-4-12B/Sycophantic isolates the benefits of pre-generation readout, counterpart-calibrated thresholds, and orthogonal anchoring. PaCE achieves the highest target forgetting and response quality among all five variants (Sec. 5.4, Table 3).
  • Task-matched baselines (W2): Using the same paired data and update budget, PaCE outperforms both SFT and DPO in target forgetting in all 12 settings and response quality in 10/12. On Qwen3.5-9B, it also achieves higher general utility than both baselines across all four personas (Appx. F.8, Table 13).

As the discussion closes in 2 days, please let us know if any further clarification would be helpful. Should these results address the issues you raised, we hope you might consider a more favorable rating. Thank you again for your time and consideration.

Best regards,
Submission1568 Authors

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

The paper formulates persona unlearning as editing model weights so that an undesirable response policy becomes difficult to elicit while useful behavior is retained. PersonaUnlearnBench constructs model-specific target/counterpart response pairs for five personas across six models, with held-out questions and instructions. PACE extracts a response-level persona direction, selects an editing layer using separation and steering probes, and trains final-prompt states using a counterpart-relative projection penalty and an orthogonal alignment loss. Evaluation separates target forgetting, counterpart preservation, response quality, and general utility. The reported results show large reductions in judged persona expression, with variable utility costs and residual accessibility under stronger elicitation.

Critical strengths
  1. The benchmark distinguishes behavioral suppression from answer memorization and generation collapse. The paired construction and four-axis evaluation in Section 3 address an important ambiguity in unlearning evaluations. The diagnostic in Section 3.5 shows that suppressing reference content need not suppress the persona, while a high forgetting score can be misleading when only a few outputs remain gradable. This distinction is useful independently of PACE.

  2. PACE is simple and produces substantial effects across several model/persona combinations. The two losses in Section 4 have clear complementary roles, and the controlled ablation in Table 2 supports their combination in the tested setting. Table 1 shows strong target suppression while often preserving counterpart behavior and improving response quality. The additional adversarial, conversational, translated, and new-instruction/new-question tests in Section 5.3 and Appendix F.2 provide meaningful evidence beyond reuse of the training wording, while openly reporting residual failures and utility losses.

Critical weaknesses
  1. The comparisons do not establish the advantage over direct behavioral retraining using the same paired supervision. The baselines in Section 5.1 and Appendix E are primarily forgetting objectives. Since the benchmark supplies desirable answers paired with target prompts, a particularly relevant comparison is supervised training to produce the desirable answer under the target prompt, or a preference objective contrasting target and desirable responses under that same prompt. The reported GradDiff/NPO retain losses use counterpart contexts and do not resolve this question. Moreover, Appendix E lists tuning grids but does not clearly specify the independent validation set and selection rule for the reported operating points. These gaps make it difficult to determine whether PACE offers a better suppression-retention tradeoff than a simpler weight-level behavioral edit.

  2. The central behavioral measurements need independent validation before supporting the stronger unlearning interpretation. DeepSeek-V4-Flash is used for response filtering and final persona judgments, and the analysis does not report agreement with an independent judge or human annotations (Sections 3.3-3.4; Appendix G). This is especially consequential for Hallucinating: response quality explicitly excludes external factual correctness, so more cautious language can improve the persona score without establishing fewer factual errors. Retention primarily measures the designated counterpart plus general tasks, rather than the unrelated personas required by Section 2.2. The observed suppression is promising, but these measurements leave uncertainty about the selectivity and substantive meaning of the edit.

Questions for authors
  1. How does PACE compare with training on the same target-prompt/desirable-answer pairs using SFT, or a matched preference objective, at comparable tuning budgets and utility loss? This comparison would directly affect the assessment of the method's contribution.
  2. Which examples and criterion selected each method's hyperparameters? Please distinguish the layer-selection probe, hyperparameter-validation data, and final test contexts, and clarify whether reported test scores influenced operating-point selection.
  3. Does the main result persist under independent human or second-model judgments, including factual verification for a sample of Hallucinating outputs and checks of non-target persona retention? These would clarify whether the measured gain reflects selective behavior removal rather than changes preferred by the existing evaluator.
7 comments
AuthorsOpenReview ↗

Response to Reviewer FoTr (Part 1/4)

Thank you for your thoughtful review and for recognizing the benchmark’s distinction between behavioral suppression and memorization, PaCE’s simple design, and the evaluation beyond the training prompts. Below we address each of your weaknesses and questions point-by-point.


Weakness 1: Direct behavioral baselines and hyperparameter selection (Part 1/2)

Thank you for identifying these important comparisons. We address them by: (1) adding the requested task-matched SFT and DPO baselines under the same target-persona instruction, paired data, and update budget, where PaCE achieves higher TF than both SFT and DPO in 12/12 settings and higher RQ in 10/12 settings, and (2) clarifying our layer-localization and operating-point selection protocol, which follows common practice in LLM unlearning benchmarks.

  1. PaCE improves on direct behavioral retraining (SFT and DPO), not only on forgetting objectives

    We implement the exact conditional tasks you propose: SFT trains the model to produce the constructive answer under the target-persona instruction, and DPO prefers that constructive answer over the undesirable target answer under the same instruction (using the frozen Original model as reference, without an SFT warm-up). Both use the same 400 paired examples and 75 optimizer updates as PaCE, and are evaluated on 250 target-context and 250 counterpart-context responses per setting together with the full four-benchmark GU suite.

    Table A.1: Task-matched direct behavioral baselines.

    ModelMethodSycophanticHallucinatingImpoliteApathetic
    TF↑CP↑RQ↑GU↑TF↑CP↑RQ↑GU↑TF↑CP↑RQ↑GU↑TF↑CP↑RQ↑GU↑
    Llama-3.1-8BOriginal3.6071.9070.2068.6513.1068.5064.9568.653.7074.6074.1068.6541.2078.6073.5068.65
    SFT81.5682.5090.5566.0483.7472.0063.0065.0185.9275.6071.7563.8191.0776.5066.7565.56
    DPO90.3172.6088.2567.3175.2669.4056.0041.0880.3876.1084.8564.7582.6478.3077.0066.65
    PaCE100.0071.1093.0065.7893.6067.7056.4563.87100.0073.6093.2063.75100.0076.4091.8564.27
    Qwen3.5-9BOriginal0.5081.0078.4568.685.2087.2078.0068.680.0074.2070.8568.6821.4792.4069.7568.68
    SFT62.71100.0088.7064.1485.3295.7074.2567.0780.2385.2086.1564.9088.4697.9068.7061.76
    DPO69.4898.7079.3064.7875.0984.8075.0561.4786.1781.1086.1066.6681.9395.9071.4563.29
    PaCE79.6081.8095.7568.5494.2082.6080.4569.02100.0074.0095.5067.00100.0086.6084.4565.10
    Gemma-4-12BOriginal0.2080.7078.9074.027.9081.9080.3574.020.0072.0073.2574.0233.4087.3074.1574.02
    SFT59.1298.5077.6570.0264.7394.0068.8067.7983.9181.7368.8567.9487.2497.0063.5069.93
    DPO67.8599.0073.6572.7755.6889.0068.8571.1891.6282.3075.5071.2780.6796.5063.9072.60
    PaCE77.9080.0093.4571.6175.5077.6067.7069.31100.0072.9094.5568.79100.0087.8084.2068.30

    Table A.1 highlights three key findings:

    • Consistent suppression advantage across all settings. Across the 12 settings, mean TF is 93.40 for PaCE, compared with 79.50 for SFT (+13.90) and 78.09 for DPO (+15.31). PaCE exceeds both baselines in all 12 individual settings.
    • Higher suppression is accompanied by higher response quality. PaCE achieves higher RQ than both baselines in 10/12 settings, with mean RQ gains of +11.83 over SFT and +10.89 over DPO, while keeping mean CP close to Original (77.68 vs. 79.19). On Llama–Impolite, where PaCE and SFT have nearly identical GU (63.75 vs. 63.81), PaCE improves TF by +14.08 and RQ by +21.45.
    • Simultaneous improvements in TF, RQ, and GU on Qwen. On Qwen3.5-9B, PaCE outperforms both SFT and DPO on TF, RQ, and GU in all four persona settings.

    We have incorporated this comparison into the revised manuscript (Appx. F.8, Table 13).

AuthorsOpenReview ↗

Response to Reviewer FoTr (Part 2/4)

Weakness 1: Direct behavioral baselines and hyperparameter selection (Part 2/2)

  1. Layer localization, operating-point selection, and alignment with common unlearning practice

    We clarify the role of each data split and selection rule in Appx. E:

    • Layer-selection probe (Qprobe\mathcal{Q}_{\text{probe}}). Thirty questions from the construction pool—disjoint from the 50 test questions—are paired with a held-out target–counterpart instruction pair to shortlist six layers by signed Cohen’s dd (with ROC-AUC as tie-breaker) and select the frozen intervention layer ℓ⋆\ell^\star and direction v^\hat{v} via a temporary negative-steering probe (Appx. D.1). This probe is used only for layer localization prior to training.
    • Operating-point selection follows common LLM unlearning practice. Because unlearning inherently involves a trade-off between target removal and utility retention, standard LLM unlearning benchmarks and methods—including TOFU [1], WMDP [2], NPO [3], MUSE [4], and OpenUnlearning [5]—commonly evaluate checkpoints over a defined hyperparameter grid on the benchmark evaluation suite to select operating points that balance forgetting and model utility, without partitioning a separate intermediate unlearning-validation split. Following this standard protocol [1–5], we evaluate final checkpoints across PaCE’s 12-point grid (λ∈{0.5,1,3,5},m∈{0,1,2}\lambda\in\{0.5,1,3,5\}, m\in\{0,1,2\}) and the baselines’ grids (Appx. E, Table 9), selecting the reported configurations by jointly considering high TF alongside preservation of CP, RQ, and GU.
    • Generalization to a strictly selection-unused prompt pool. To further verify that suppression does not rely on selection over the test contexts, Appx. F.2.2 (Fig. 8) evaluates the frozen selected checkpoints on 50 newly written questions crossed with 5 new instructions per persona that were used for neither training nor hyperparameter selection. PaCE achieves TF 95.0–100.0 across all five personas on this untouched pool.

[1] TOFU: A Task of Fictitious Unlearning for LLMs. In COLM, 2024.
[2] The WMDP Benchmark: Measuring and Reducing Malicious Use with Unlearning. In ICML, 2024.
[3] Negative Preference Optimization: From Catastrophic Collapse to Effective Unlearning. In COLM, 2024.
[4] MUSE: Machine Unlearning Six-Way Evaluation for Language Models. In ICLR, 2025.
[5] OpenUnlearning: Accelerating LLM Unlearning via Unified Benchmarking of Methods and Metrics. In NeurIPS, 2025.

AuthorsOpenReview ↗

Response to Reviewer FoTr (Part 3/4)

Weakness 2: Independent validation, factuality, and selective retention

Thank you for distinguishing these three questions. We address them by adding: (1) re-evaluations using two independent model judges (GPT-6 Luna and GPT-6.1 Terra), (2) an author-reviewed, source-based factuality audit on Hallucinating outputs, and (3) direct evaluations of four unrelated non-target styles.

  1. Two independent judges reproduce the suppression gain across all 12 settings

    We use GPT-6 Luna and GPT-6.1 Terra—neither of which was used during data construction or filtering—to independently re-score the same fixed pool of 12,000 Original and PaCE responses (250 target-context and 250 counterpart-context responses per setting). Keeping the responses fixed isolates the effect of changing the evaluator.

    Table A.2: Target suppression (TF ↑, Original → PaCE) across independent judges.

    ModelPersonaDeepSeekGPT-6 LunaGPT-6.1 Terra
    Llama-3.1-8BSycophantic3.60→100.002.20→99.702.20→99.00
    Hallucinating13.10→93.6010.90→82.5011.50→81.10
    Impolite3.70→100.001.80→100.006.60→99.90
    Apathetic41.20→100.0053.30→100.0047.00→99.60
    Qwen3.5-9BSycophantic0.50→79.600.00→66.200.00→71.20
    Hallucinating5.20→94.203.30→67.102.70→62.50
    Impolite0.00→100.000.10→100.000.90→99.90
    Apathetic21.47→100.0031.10→97.7025.40→97.30
    Gemma-4-12BSycophantic0.20→77.900.00→60.400.10→68.50
    Hallucinating7.90→75.505.90→65.606.60→64.70
    Impolite0.00→100.000.00→99.300.10→97.00
    Apathetic33.40→100.0047.69→100.0041.97→100.00

    Table A.2 shows that:

    • All 12 settings improve substantially under both independent judges, with mean TF gains of +73.52 points for Luna and +74.64 points for Terra.
    • Even the smallest single-setting gains are +46.70 and +52.60 points, respectively, and for Impolite and Apathetic, every PaCE score under both independent judges lies between 97.00 and 100.00.

    Thus, while absolute scales vary slightly across evaluators, the conclusion of substantial target suppression is consistently reproduced by independent judges.

  2. Source-based factuality audit examines factual content beyond cautious wording

    We additionally compile an author-audited factuality casebook for 20 Hallucinating questions on Llama-3.1-8B (pairing Original and PaCE for 40 answers, selected by question content independently of model outputs; available in our anonymous repository under rebuttal/factuality_audit/). Each answer is checked against external primary sources.

    Table A.3: Answer-level factuality in the source-based audit.

    ObservationOriginalPaCE
    Contains verified useful information (↑)4/208/20
    Contains at least one verified factual error (↓)8/206/20
    Fully refuses or avoids the requested answer0/205/20

    Table A.3 shows that the number of answers containing verified useful information doubles (from 4/20 to 8/20), while the number containing a verified factual error decreases (from 8/20 to 6/20) (with refusals tracked separately and not counted as factual successes). For instance, in Q91, PaCE identifies missing experimental details rather than inventing precise figures.

  3. Unrelated non-target styles remain elicitable after editing

    To evaluate retention beyond the designated counterpart as required in Sec. 2.2, we evaluate Original and all four Llama-3.1-8B PaCE edits across four unrelated styles (Optimism, Formal, Concise, and Cautious; 20 questions per style).

    Table A.4: Non-target style expression (0–100 ↑).

    Model stateOptimism ↑Formal ↑Concise ↑Cautious ↑
    Original89.7598.7579.2585.00
    PaCE–Sycophantic80.0093.7563.7583.25
    PaCE–Hallucinating80.0092.5076.7585.00
    PaCE–Impolite76.5083.7567.0085.00
    PaCE–Apathetic78.0091.2560.0081.25

    Table A.4 shows that Cautious expression remains within 81.25–85.00 (matching Original’s 85.00 on Hallucinating and Impolite), Formal remains high at 83.75–93.75, and Optimism remains at 76.50–80.00. This confirms that suppressing the target persona preserves the model’s broader ability to enact other unrelated styles. We have added these three validation analyses to the revised manuscript (Sec. 5.4, Table 4 and Appx. F.9, Tables 14–15).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.