acceptodds
Under review as a conference paper at ICLR 2027

RapidUn: Influence-Driven Parameter Reweighting for Efficient Large Language Model Unlearning

Abstract

Machine unlearning for large language models (LLMs) remains challenging because full retraining is costly, while approximate methods often struggle to remove targeted behaviors without degrading retained utility, especially under limited post-deployment supervision. We consider a practical PEFT setting for targeted behavioral contamination removal with a small forget set, a limited retain buffer, and LoRA-only updates, and propose RapidUn, an influence-guided framework that converts cross-sample influence estimates into fixed sample-specific weights for weighted LoRA unlearning. Across Llama-3-8B on Dolly-15k and Alpaca-57k, with cross-model validation on Mistral-7B + Dolly-15k, RapidUn achieves lower seen-trigger and OOD-trigger-family ASR than Fisher, GA, and LoReUn while maintaining competitive clean utility. On Llama-3-8B + Alpaca-57k, it achieves a wall-clock speedup over the clean-corpus LoRA retraining reference. Complementary TOFU, semantic LLM-judge, and IFEval evaluations further support the effectiveness of influence-guided sample reweighting beyond the controlled trigger benchmark.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
machine unlearning, large language models, influence estimation, parameter-efficient fine-tuning, LoRA, sample reweighting

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

RapidUn studies behavioral contamination removal when only a small forget set and retain buffer are available. It converts four types of forget/retain sample influence into fixed, normalized weights for LoRA ascent and descent. Experiments cover synthetic trigger contamination on Dolly and Alpaca, cross-model validation on Mistral, and complementary TOFU and IFEval evaluations.

Critical strengths
  • The method addresses a practical unlearning setting with a small forget set and limited retain data.
  • Influence-guided weighting improves the forgetting–utility trade-off and speedup the retraining process.
  • The ablations, three-seed Dolly experiments, TOFU evaluation, and IFEval checks provide useful supporting evidence.
Critical weaknesses
  • only 7-8B models are evaluated
  • 77x speedup compared with retrain might be unfair as the training corpus are not the same.
Questions for authors
  • How does the k ~= 3 comes in line 63-64. Is it a magic number?
  • In the paper, we only test the generalizability in the 7-8B model, it would be nice the authors can provide the 1B, 4B models from different families to check the generalizability of the models. It would strength the approach. I see it would be lots of work to retrain the unlearning approaches, but 1 or 2 smaller models would be nice.
  • what's the epoch in retrain? is it the same as the rapidUn since this is connect to the training effciency?
  • Does the speedup 77x comes from few training corpus? If yes, in the table 5, the delta seen ASR in rapid is 29, for the retrain, it's about 96.7. So we can control the delta_ASR seen during retrain phase, and monitor if the ASR of retrain cross the rapidUn method can check the wallclock of retrain used. I'm happy to improve my score if the authors resolve my concern.
Reviewer_K4rZOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies approximate unlearning for LLMs in a parameter-efficient setting. The full training corpus is unavailable, and the practitioner has only a small forget set (~40 examples), a retain buffer about three times larger, and LoRA-only updates. The target use case is removing behavior introduced by contaminated fine-tuning data. The paper instantiates this as a trigger-based backdoor that maps triggered instructions to off-topic science-fiction-style responses.

The proposed method, RapidUn, is a weighted gradient-ascent/descent scheme with precomputed, fixed per-sample weights. The weights are built in three steps:

  1. Influence estimation. The RapidIn token-wise influence estimator computes four directional influence matrices over the forget and retain sets: forget→forget, forget→retain, retain→forget, and retain→retain.
  2. Directional fusion. The matrices are aggregated row-wise and fused into per-set scores. Forget samples are favored when they are representative of the forget set and have low influence on the retain buffer; retain samples are favored analogously. On the trigger benchmark, optional terms restrict influence to the poisoned answer span.
  3. Influence-to-weight mapping. Scores are mapped to mean-one weights via median/MAD robust scaling, temperature smoothing, and log-space clipping.

The weights then scale the per-sample losses in the LoRA objective Lr−αFALfL_r - \alpha_{FA} L_f.

Claimed contributions:

  • the formulation of this small-buffer PEFT unlearning setting;
  • the four-way influence-to-weight mapping, with released code;
  • an improved forgetting–utility trade-off over GA, Fisher Forgetting, and LoReUn.

Validation. The main experiments use Llama-3-8B-Instruct on a poisoned Dolly-15k benchmark. Forgetting is measured by keyword-based attack success rate (ASR) on seen and held-out trigger variants, and utility by clean perplexity. Cross-model results are reported on Mistral-7B. Scalability results are reported on a larger Alpaca corpus, where the authors claim a 77× wall-clock speedup over full-corpus LoRA retraining. Supporting analyses include:

  • ablations on influence directions and on forget- vs. retain-side weighting;
  • a proxy-swap experiment replacing RapidIn with alternative scoring signals;
  • a retain–forget semantic-overlap stress test;
  • TOFU Forget05 comparisons against SimNPO and BalDRO-DV;
  • an LLM-judge semantic ASR check and an IFEval instruction-following evaluation.
Critical strengths
  1. Practical, clearly scoped setting. The constraints (a small forget set, a small retain buffer, LoRA-only updates) reflect a realistic post-deployment scenario. The paper is careful about its claims: it repeatedly states that the results are behavioral and imply neither certified deletion nor privacy guarantees, and that "OOD" refers to held-out trigger variants within the same contamination design.
  2. Simple and modular method. Weights are computed once and kept fixed, so the method adds little overhead to standard GA. The influence estimator is treated as a pluggable signal source, as the proxy-swap experiment illustrates.
  3. Ablations target the core claim. The uniform / self-only / full comparison, the forget-only vs. retain-only weighting ablation, and the proxy-swap experiment directly probe whether the cross-set influence signal, rather than generic reweighting, is responsible for the gains.
  4. Broad supplementary evaluation. Multiple seeds, forget-set-size sensitivity, αFA\alpha_{FA} sensitivity, a retain-overlap stress test, a second architecture, TOFU, IFEval, and an LLM-judge check together give a wider picture than the main trigger benchmark alone.
  5. Reproducibility. Hyperparameters and fusion/mapping defaults are fully reported and fixed across settings, and code is released.
Critical weaknesses

Weaknesses

  1. Weak baselines and single-point comparison. On the main benchmark, Fisher Forgetting (seen ASR 0.83) and Retain-Only (0.86) barely move from the poisoned base (0.844). The effective comparison therefore reduces to GA, which is RapidUn with uniform weights, and LoReUn. Stronger and more recent unlearning methods (e.g., NPO/SimNPO, RMU, WAGLE) and backdoor-removal defenses are absent from the main benchmark; SimNPO and BalDRO-DV appear only on TOFU, where the margin is 0.01 and no variance is reported. In addition, every method is evaluated at a single forgetting strength. Table 13 shows that αFA\alpha_{FA} trades ASR against PPL, so a fair comparison would present ASR–PPL Pareto curves for all methods, with each baseline's strength tuned.

  2. The Alpaca results and the 77× claim are presented misleadingly. Table 5 reports only ΔASR. Since Retrain reduces seen ASR by 96.7 pp, the poisoned base's seen ASR must be at least 0.967. RapidUn's reduction of 29.0 pp then implies a residual seen ASR of roughly 0.68 or higher. The OOD split shows the same pattern: Retrain reduces OOD ASR by 41.4 pp, so the base OOD ASR is at least 0.414, and RapidUn's 21.0 pp reduction leaves a residual of roughly 0.20 or higher. In other words, on the larger corpus most of the backdoor behavior appears to remain. The 77× speedup, highlighted in the abstract, comes from the small-buffer setting rather than from the method: GA (~111×) and LoReUn (~91×) are faster still. The "ΔASR per hour" metric also favors methods that are cheap regardless of how much they remove. Please report absolute ASR for all methods on Alpaca.

  3. Unclear or possibly inconsistent method details.

    • Score fusion. With the stated defaults, Eq. 1 becomes Sf=FF−FR−FFh−FRhS_f = FF - FR - FF_h - FR_h, so the token-restricted forget→forget term enters with a negative sign, opposite to FFFF. Moreover, the loss is already answer-only. If forget examples are trained on poisoned answers, it is unclear how FFhFF_h differs from FFFF; if the two are nearly equal, the FFFF contribution largely cancels.
    • Ascent target. The paper does not state whether the ascent on forget examples uses the poisoned answer or the clean answer (Appendix A allows either).
    • Weight distribution. The example retain weights in Table 19 (0.885, 0.885, 2.655) correspond exactly to pre-normalization values of 1 (the median) and 3 (the clipping cap). This suggests the scores may contain many ties or near-zero MAD. Please show the weight histograms.
    • Possible double weighting. It is unclear whether the influence weights also enter the weighted sampler, which would apply them twice.
  4. Questionable metric reliability. Clean PPL is reported as a mean of per-example perplexities, which is highly sensitive to outliers. This is a likely explanation for the gap between the final validation PPL in Fig. 3 (~6) and the test PPL in Table 2 (~44.6). Under this metric, the 0.3-PPL differences between methods are probably within noise. Retain-Only fine-tuning on clean data increases clean PPL from 50.5 to 54.6 and slightly increases ASR, which suggests an issue with the training setup. The LLM-judge check names neither the judge model nor its agreement with human or keyword labels, and it produces numbers nearly identical to keyword ASR (0.153 vs. 0.153 for RapidUn), which limits its value as an independent check.

  5. Forgetting may be shallow, and the evaluation is narrow. In Fig. 3(b), RapidUn's PPL on the poisoned forget targets rises only to about 1.44. The model therefore still assigns high likelihood to the poisoned responses, even though greedy decoding no longer emits them. No relearning, sampling-based, or extraction attacks are evaluated. Combined with the highly synthetic setup (stylistically extreme sci-fi responses, a keyword detector derived from the training data, and "OOD" triggers drawn from the same families), it remains unclear whether the behavior is removed or merely suppressed at the surface.

Questions for authors
  1. Clarifying the method.

    • With the default coefficients, Eq. 1 gives Sf=FF−FR−FFh−FRhS_f = FF - FR - FF_h - FR_h. Why does FFhFF_h enter with a negative sign, and how does it differ numerically from FFFF, given that the loss is already answer-only?
    • Does forget-side ascent use the poisoned answer or the clean answer?
    • Could you show histograms of wfw_f and wrw_r, and report what fraction of weights hit the clipping bounds?
    • Do the influence weights also affect the weighted sampler?
  2. Absolute results and comparison protocol.

    • Could you report absolute seen/OOD ASR for all methods on Alpaca?
    • Could you provide ASR–PPL Pareto curves for RapidUn and the baselines, with each baseline's forgetting strength tuned?
    • How does RapidUn compare on the trigger benchmark against NPO [1], RMU [2], and a backdoor-removal defense such as BEEAR [3] or Fine-Pruning [4]?
  3. Robustness of forgetting. Forget-set poisoned PPL remains around 1.44 after unlearning. Does the backdoor resurface under (a) temperature sampling, (b) a small amount of benign or related fine-tuning, as in the relearning attacks of [5], or (c) paraphrased triggers outside the three families?

  4. Metrics.

    • Could you report clean PPL as exp(mean token NLL) in addition to the mean of per-example PPL?
    • Why does Retain-Only increase both clean PPL and ASR relative to the base model?
    • Which model is used as the LLM judge, and what is its agreement rate with human or keyword labels?
  5. Role of RapidIn. On a single checkpoint, RapidIn largely reduces to projected gradient similarity, similar in spirit to TracIn [6], and the Grad-dot proxy in Table 9 is reasonably close. What specifically does RapidIn's token-wise estimation contribute? Do the learned weights correlate with trigger family or with any interpretable sample property?

Missing related work. Several closely related LLM unlearning methods are not discussed. They need not all be added as experimental baselines, but they should be positioned in the related work:

  • Influence-based unlearning. SOUL [7] explicitly connects second-order optimization with influence unlearning and extends one-shot influence-function updates to an iterative LLM unlearning procedure. Because RapidUn's central claim is an influence-driven design, the paper should discuss how its influence-derived sample weights relate to this line of work.
  • Reweighting-based objectives. WGA [8] and SatImp [9] reweight the gradient-ascent objective at the token level based on the model's own prediction confidence. FLAT [10] performs unlearning with only forget data and assigns importance weights through an f-divergence formulation. RapidUn also reweights GA, but at the sample level using influence signals. The paper should clarify how its weighting differs from, or complements, these schemes; SatImp is currently cited only in passing in Section 2. FLAT is also directly relevant to the paper's small-retain-buffer motivation.
  • Spurious forgetting and robustness. BS-T/BS-S [11] show that GA-style objectives push probability mass onto semantically related rephrasings (the "squeezing effect"), yielding spurious unlearning. EUA [12] similarly argues that token-level suppression leads to biased deletion and proposes distribution-level, energy-based unlearning. Fan et al. [13] show that unlearned models are vulnerable to relearning from a few forget examples and propose smoothness optimization as a defense. TRU [14] argues that the untargeted nature of GA objectives causes incoherent outputs and incomplete removal. All four are directly relevant to whether RapidUn's ASR reduction reflects genuine removal or surface-level suppression (Weakness 5, Question 3).
  • Behavioral unlearning and efficiency. Yao et al. [15] formulate LLM unlearning as removing harmful behaviors using only negative examples, which is close to RapidUn's behavioral-contamination setting. LUNAR [16] restricts updates to a single down-projection matrix via activation redirection and reports large efficiency gains, which is relevant to the paper's efficiency and PEFT claims.

References

[1] Ruiqi Zhang, Licong Lin, Yu Bai, Song Mei. Negative Preference Optimization: From Catastrophic Collapse to Effective Unlearning. COLM 2024.

[2] Nathaniel Li, Alexander Pan, Anjali Gopal, et al. The WMDP Benchmark: Measuring and Reducing Malicious Use with Unlearning. ICML 2024.

[3] Yi Zeng, Weiyu Sun, Tran Ngoc Huynh, Dawn Song, Bo Li, Ruoxi Jia. BEEAR: Embedding-based Adversarial Removal of Safety Backdoors in Instruction-tuned Language Models. EMNLP 2024.

[4] Kang Liu, Brendan Dolan-Gavitt, Siddharth Garg. Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks. RAID 2018.

[5] Shengyuan Hu, Yiwei Fu, Zhiwei Steven Wu, Virginia Smith. Unlearning or Obfuscating? Jogging the Memory of Unlearned LLMs via Benign Relearning. ICLR 2025.

[6] Garima Pruthi, Frederick Liu, Satyen Kale, Mukund Sundararajan. Estimating Training Data Influence by Tracing Gradient Descent. NeurIPS 2020.

[7] Jinghan Jia, Yihua Zhang, Yimeng Zhang, Jiancheng Liu, Bharat Runwal, James Diffenderfer, Bhavya Kailkhura, Sijia Liu. SOUL: Unlocking the Power of Second-Order Optimization for LLM Unlearning. EMNLP 2024.

[8] Qizhou Wang, Jin Peng Zhou, Zhanke Zhou, Saebyeol Shin, Bo Han, Kilian Q. Weinberger. Rethinking LLM Unlearning Objectives: A Gradient Perspective and Go Beyond. ICLR 2025.

[9] Puning Yang, Qizhou Wang, Zhuo Huang, Tongliang Liu, Chengqi Zhang, Bo Han. Exploring Criteria of Loss Reweighting to Enhance LLM Unlearning. ICML 2025.

[10] Yaxuan Wang, Jiaheng Wei, Chris Yuhao Liu, Jinlong Pang, Quan Liu, Ankit Parag Shah, Yujia Bao, Yang Liu, Wei Wei. LLM Unlearning via Loss Adjustment with Only Forget Data. ICLR 2025.

[11] Kemou Li, Qizhou Wang, Yue Wang, Fengpeng Li, Jun Liu, Bo Han, Jiantao Zhou. LLM Unlearning with LLM Beliefs. ICLR 2026.

[12] Puning Yang, Junchi Yu, Qizhou Wang, Philip Torr, Bo Han, Xiuying Chen. Distinguishable Deletion: Unifying Knowledge Erasure and Refusal for Large Language Model Unlearning. arXiv preprint arXiv:2605.16776, 2026.

[13] Chongyu Fan, Jinghan Jia, Yihua Zhang, Anil Ramakrishna, Mingyi Hong, Sijia Liu. Towards LLM Unlearning Resilient to Relearning Attacks: A Sharpness-Aware Minimization Perspective and Beyond. ICML 2025.

[14] Junfeng Liao, Qizhou Wang, Shanshan Ye, Xin Yu, Ling Chen, Zhen Fang. Explainable LLM Unlearning Through Reasoning. ICLR 2026.

[15] Yuanshun Yao, Xiaojun Xu, Yang Liu. Large Language Model Unlearning. NeurIPS 2024.

[16] William F. Shen, Xinchi Qiu, Meghdad Kurmanji, Alexandru-Andrei Iacob, Lorenzo Sani, Yihong Chen, Nicola Cancedda, Nicholas D. Lane. LLM Unlearning via Neural Activation Redirection. NeurIPS 2025.

Reviewer_E8aoOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes an LLM unlearning with small number of forget set. The paper explains that existing unlearning methods use unanimous weights for every forget and retain samples, despite their individual contribution on forgetting, retaining or cross-interference differs among them. To address this, the authors propose RapidUN. It uses RapidIN to compute gradient interactions between each sample. For a forget sample, the method computes the sample influence of the forget sample to another forget and retain sample and aggregate the influence score to measure how the forget sample affects other forget or retain samples. The paper assigns high score for forget sample whose influence to other forget sample is higher and whose influence to retain samples are lower. Vice versa for retain samples. RapidUN conducts PEFT with the loss from the weighted samples. For experiments, the authors constructed poisoned samples using a fraction of Dolly-15k. Each sample is equipped with a trigger and an out-of-domain suffix following the trigger (sci-fi topic texts). The experiment unlearns the trigger and demonstrate that the unlearning removes the trigger and hence the model does not reproduce the memorized suffix. The result shows that RapidUN achieves the lowest ASR with high model utility (low perplexity on clean set).

Critical strengths

None

Critical weaknesses
  1. The problem setting is not clearly stated. The paper repeatedly mentions behavioral removal, however the paper never discusses nor defines what that is, and does not differentiate how that is different from other settings of LLM unlearning.

  2. From the experiments, it seems that the goal of the paper is to unlearn a backdoor, to mitigate data poisoning attack. However, the paper never discusses any related works regarding backdoor attack. The paper should compare their method against the baselines tailored for backdoor unlearning. Am example can be OSFT (Li et al, 2025), and W2SDefense (Zhao et al. 2025)

  3. The method requires excessive number of gradient computation to measure the per-sample influence. The authors should provide the actual runtime of the algorithm to demonstrate its feasibility.

  4. If measuring per-sample influence requires significant amount of computation, then the comparison to the baselines is not fair as proposed method is given computational advantage over the baselines. The authors should address this by offering fair amount of computational resource to baselines as well.

  5. The paper does not discuss any scalability analysis of the proposed method. If the forget corpus is bigger, how does the computation and performance scale?

  6. The backdoor leads to the poisoned response that is off-topic. Since the output is lexically distinctive, the model simply could learn to not generate any sci-fi related words, instead of removing the backdoor. Thus the paper should demonstrate the effectiveness of unlearning with poisoning the data with in-domain response.

  7. The paper lacks any robustness analysis (relearning attack, etc.)

  8. Per-sample weighting is not a novel idea and has been explored widely. The paper should discuss them and if possible, include them as baselines (Niu et al. 2025, Liu et al. 2026).

  9. The per-sample weighting naturally assigns lower score to forget samples correlated to the retain samples. The lower score on them could unlearn them insufficiently. This potentially poses a security risk. The paper should provide an analysis on them to demonstrate those samples are sufficiently unlearned even with the lower score.


Li, Haoran, et al. "Simulate and eliminate: Revoke backdoors for generative large language models." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 39. No. 1. 2025.

Zhao, Shuai, et al. "Unlearning backdoor attacks for llms with weak-to-strong knowledge distillation." Findings of the Association for Computational Linguistics: ACL 2025. 2025.

Niu, Peizhi, et al. "Guard: Guided unlearning and retention via data attribution for large language models." arXiv preprint arXiv:2506.10946 (2025).

Liu, Zhihao, et al. "Towards Mitigating Excessive Forgetting in LLM Unlearning via Entanglement-Guidance with Proxy Constraint." IEEE Transactions on Dependable and Secure Computing (2026).

Questions for authors
  1. Please see the weakness.
  2. Why did you decide not to use benchmark dataset for backdoors of LLM? such as BackdoorLLM (Li et al, 2026) or PoisonBench (Fu et al., 2024) ?

I am willing to increase the score depending on the authors' response.


Li, Yige, et al. "Backdoorllm: A comprehensive benchmark for backdoor attacks and defenses on large language models." Advances in neural information processing systems 38 (2026).

Fu, Tingchen, et al. "Poisonbench: Assessing large language model vulnerability to data poisoning." arXiv preprint arXiv:2410.08811 (2024).

Comments on the paper

1 comment

Dear reviewers Thank you for your thoughtful review. The authors have submitted their rebuttal, and the discussion period is now open. Please take a moment to read the authors' response and engage with them and your fellow reviewers to address any remaining questions or clarify key points. Your active participation during this phase is vital to reaching a fair and well-informed final recommendation.

Best regards, area chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.