iShumei-Chinchunmei at SemEval-2025 Task 4: A balanced forgetting and retention multi-task framework using effective unlearning loss

Tian Li; Yujian Sun

REVIEW 4 major objections 5 minor 16 references

iShumei-Chinchunmei at SemEval-2025 Task 4: A balanced forgetting and retention multi-task framework using effective unlearning loss

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that a reciprocal of the standard next-token prediction loss lets gradient descent erase targeted knowledge from an LLM, and that combining it with retention fine-tuning and data augmentation balances forgetting with…

desk verdict A competent SemEval system paper whose proposed loss reduces to reweighted gradient ascent and is never tested against that baseline. read the letter →

arxiv 2507.16263 v1 pith:PCWCKZIJ submitted 2025-07-22 cs.CL

Yujian Sun , Tian Li This is my paper

classification cs.CL

keywords machineunlearningeffectivelossmulti-tasklearninglargelanguagemodelsdataaugmentationnegativeresponsereplacementLoRAfine-tuningprivacypreservation

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Machine unlearning aims to delete specific sensitive content from a trained language model without wrecking what the model otherwise knows. This paper claims that a single change to the training objective—using the multiplicative inverse of the usual next-token prediction loss on the texts to be forgotten—makes ordinary gradient descent push the model away from that content, in place of the usual but unstable gradient-ascent step. The inverse loss is paired with supervised fine-tuning on a retained set and with a data-augmentation scheme that re-segments long outputs, and the authors report that this multi-task recipe balances forgetting with general performance. The system placed 5th in the shared task, and the ablation study shows retention fine-tuning is essential: without it, forgetting scores stay high while general knowledge falls to near-random.

What carries the argument

The load-bearing object is EUL, the reciprocal of the standard next-token prediction loss on the forget data, scaled by a factor $\alpha$. Since $\nabla(1/L) = -\nabla L/L^2$, descending on EUL is mathematically ascending on the original loss with an adaptive step size that shrinks as the loss grows; this identity is what lets the method claim stable forgetting without an optimizer change. The surrounding framework alternates forget-set EUL batches with retain-set SFT batches, and its data-augmentation scheme re-segments long outputs to counter a long-tail distribution of text lengths.

What would settle it

A decisive test is to run the same reciprocal-loss objective on data the model should keep; if the forgetting metric rises there too, the loss is damaging the model generally rather than erasing the targeted content.

Watch

Extended reading notes

Core claim

The central claim is that the Effective Unlearning Loss (EUL), $L_{\mathrm{EUL}} = \alpha / L_{\mathrm{ntp}}(x_{\mathrm{input}}, y_{\mathrm{forget}})$, lets the model forget during ordinary gradient descent: when the model's predictions align with the text to be forgotten, the reciprocal loss grows, so a descent step moves the parameters away from that alignment. The authors embed EUL in a two-task framework that alternates forget batches with retain batches optimized by the standard next-token prediction loss, and they add a data-augmentation step that splits long outputs into shorter input/output pairs. Experiments on OLMo-1B and OLMo-7B with LoRA show the best balanced configuration to be EUL plus retain-set fine-tuning plus data augmentation, while the actual competition submission, constrained by deadline, used EUL plus retain fine-tuning plus negative-response replacement. The paper concludes that retention fine-tuning is indispensable and that EUL is a more controllable alternative to gradient ascent.

Load-bearing premise

The load-bearing premise is that making the model worse at predicting the next word of the target text means it has genuinely unlearned that text rather than simply become damaged overall; the paper's own ablation shows the forgetting metric can stay near perfect while general knowledge collapses to near random.

Editorial extensions

If this is right

Gradient descent can drive the forgetting step, so unlearning no longer requires the optimizer instability of gradient ascent.
Retain-set supervised fine-tuning is a necessary component of the recipe; without it, forgetting comes at the cost of near-random general knowledge.
Re-segmenting long outputs into shorter input/output pairs improves the forgetting metric more than simply adding more training steps.
The whole pipeline can be completed under tight compute: LoRA fine-tuning on a single GPU finishes within an hour.
On small forget sets, replacing answers with safety phrases can improve general capability but harms unlearning effectiveness due to overfitting.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A direct mathematical corollary of the definition is that $\nabla(1/L) = -\nabla L/L^2$, so EUL is gradient ascent with an adaptive step size; a natural follow-up is to test whether a hand-tuned adaptive ascent schedule reproduces EUL's stability.
The near-random general-knowledge scores under EUL-only suggest that benchmark scores which average forgetting, task, and capability metrics can reward damaging the model; future unlearning evaluations may need a capability floor or per-sample verification.
The length-augmentation result suggests that failed unlearning on long documents is partly a length-generalization problem, so similar re-segmentation could transfer to unlearning in summarization or long-context tasks.
A control the paper does not report—applying EUL to the retain set—would directly test whether the loss is target-specific; this is the cheapest experiment that would sharpen the central claim.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

You can read this one quickly. It's a solid SemEval-2025 Task 4 system paper with a weak central claim. The proposed Effective Unlearning Loss (EUL) is just alpha divided by the next-token loss. Minimizing that equals gradient ascent on the NTP loss with step size alpha/L^2. The authors acknowledge gradient ascent as prior work but never compare EUL against plain gradient ascent, so the "more controllable" claim is unsupported. That's the main thing.

What the paper does well: it's clearly written, the ablations are extensive (Tables 1-6), and the authors report configurations that fail, not just the winners. The data augmentation analysis (Table 3) is careful, controlling for training steps. The alternating-batch multi-task scheme is simple and sensible. A 5th-place leaderboard finish is a real result.

The soft spots are proportionate to the claim. No gradient ascent baseline, no error bars, single runs everywhere. The configuration selection is odd: the submitted RD&NR&EUL run gives MIA=0.000 (no measured forgetting) on OLMo-1B in Table 1, while the high-scoring EUL-only rows have MMLU 0.229, below the ~0.28 baseline — that's near-random general knowledge. So the evidence supports the alternative reading that EUL's high MIA comes from degrading the model, not targeted forgetting. The MIA metric is also closely tied to the NTP loss EUL maximizes, so some circularity is there. The one "balanced" point, RD+DA+EUL, is a single run with no variance information.

I'm not saying the paper is fraudulent or incoherent. It's an honest system description. But the central methodological novelty — that EUL is a better, more controllable forgetting loss — does not survive contact with the math and the missing baseline. The "balanced forgetting and retention" framing is doing a lot of work that the data don't fully support.

Who is this for? Practitioners entering SemEval-2025 Task 4, or people who want a tidy example of a competition system with thorough ablations. Not for someone looking for a new unlearning principle. I'd accept it for peer review if I were the editor: the experiments are real, the task is important, and a referee could push for a gradient ascent comparison and error bars. Without those, it's a workshop-level system description, not a methods paper.

My recommendation: send it to review, but expect a revision that adds a gradient ascent baseline and confronts the model-damage interpretation head-on.

Referee Report

4 major / 5 minor

Summary. This paper describes the iShumei-Chinchunmei submission to SemEval-2025 Task 4, which asks systems to unlearn sensitive content from OLMo-1B and OLMo-7B while preserving general capabilities. The authors propose Effective Unlearning Loss (EUL), defined in Eq. (1) as alpha divided by the next-token prediction loss on the forget set, and combine it with supervised fine-tuning on the retain set (RD), negative-response replacement (NR), and a re-segmentation data augmentation (DA) within a multi-task training framework. The paper reports extensive ablations over these components and hyperparameters on both models, and states that the system ranked 5th on the official leaderboard using the RD&NR&EUL configuration, even though the authors identify RD&DA&EUL as their best balanced result.

Significance. The paper is an honest and thorough competition report: it provides a large ablation matrix covering dozens of configurations across two model sizes, and it explicitly discloses that the submitted configuration was not the best-performing configuration. If the claimed advantage of EUL over plain gradient ascent were substantiated, the inverse-loss formulation would be a simple and practical addition to the unlearning toolbox. However, the central technical claim is currently unsecured: EUL is mathematically a reweighted gradient ascent, the paper never trains a gradient-ascent baseline, and the paper's own tables show that high MIA scores are accompanied by near-random MMLU scores. The official submission, with MIA=0.000, also undercuts the title claim of a balanced framework.

major comments (4)

[Section 3.1, Eq. (1)] The proposed EUL is a reweighted gradient ascent, not a distinct mechanism. Minimizing alpha/L_ntp by gradient descent yields an update of the form +eta*alpha*grad(L_ntp)/L_ntp^2, which is gradient ascent on the next-token prediction loss with an adaptive step size. The paper claims that EUL is 'more stable' and 'more controllable' than gradient ascent, but no plain gradient-ascent baseline is trained in any of Tables 1 through 6. Since the distinguishing property of the method is precisely its relationship to gradient ascent, this missing baseline is load-bearing for the central claim. Additionally, the inverse-square scaling can produce very large updates when L_ntp is small, so the asserted stability advantage is not self-evident and should be demonstrated empirically or analytically.
[Section 5, Table 1] Table 1 does not support the 'balanced forgetting and retention' characterization in the title. The highest Final Scores are achieved by EUL-only (MIA 0.993, MMLU 0.229, Final 0.543) and EUL+DA (MIA 0.989, MMLU 0.229, Final 0.547), but these rows have MMLU near chance and far below the RD-only row's 0.283, which is exactly the generic-degradation pattern the paper says it abandons. The submitted configuration (RD&NR&EUL) has MIA=0.000 and Final=0.127 on OLMo-1B, and the analogous OLMo-7B row in Table 6 also shows MIA=0.000, i.e., no measurable forgetting on the official submission. The only configuration that approaches balance, RD+DA+EUL (MIA 0.593, MMLU 0.275), is a single unreplicated run, and the paper does not explain why the official leaderboard result, which used a configuration with MIA=0, should be taken as evidence for the framework.
[Section 5 and Section 6, Tables 1-6] All experiments are single runs without error bars, multiple seeds, or confidence intervals. This is particularly concerning because the key comparisons are highly sensitive to hyperparameters: in Table 2, the RD&DA&EUL configuration moves from MIA 0.593 at learning rate 1e-4 to MIA 0.001 at 1e-5 for the same epoch count. The difference between the advertised balanced configuration (MIA 0.593) and the submitted configuration (MIA 0.000) could therefore be within run-to-run variance or small training-detail changes. The authors should report repeated runs with standard deviations, or at least a sensitivity analysis, before claiming that one configuration is reliably balanced.
[Section 3.1 and Section 4] The MIA metric, which measures whether the model still produces forget-target content, is closely aligned with the next-token prediction loss that EUL directly increases; high MIA is thus partly by construction. The only external capability check, MMLU, collapses exactly in the high-MIA rows (e.g., MMLU 0.229 for EUL-only). The evidence is therefore consistent with the alternative explanation that EUL achieves high MIA by degrading the model to near-random outputs rather than by targeted, selective forgetting. To rule this out, the paper should compare EUL against plain gradient ascent at matched step budgets and report a control metric on retained or unrelated content, such as retain-set perplexity, in addition to MMLU.

minor comments (5)

[Eq. (1)] The notation 'LEU L' should be 'L_EUL' or 'L_EUL' consistently; the current subscript is a typo.
[Section 5] The sentence 'The best performance is achieved when using EUL to process forgotten data while incorporating data augmentation and fine-tuning retained data' conflicts with Table 1, where the highest Final Score belongs to EUL-only and EUL+DA configurations. Please clarify that this is the best configuration after applying the MMLU-preservation criterion, and state that criterion explicitly.
[Section 6.1] The statement 'as the epoch number and learning rate increase, the performance increases' is not supported by Table 2: at LR 1e-5 and 1e-6, MIA remains at or near 0 for all epochs, and at LR 1e-4 the MIA jumps from 0.215 to 0.593 between 4 and 5 epochs. Please describe the non-monotonic and threshold-like behavior accurately.
[Figure 1] The paper references 'figure 1' and 'Figure 1: The Overview of Our System,' but no actual figure appears in the text. Either include the figure or remove the reference.
[Section 4] The paper reports MMLU scores but never gives the pre-unlearning MMLU baseline for OLMo-1B or OLMo-7B. Reporting these baselines would let readers quantify capability preservation directly instead of inferring it from no-EUL rows.

Circularity Check

1 steps flagged · score 4.0 of 10

EUL's high MIA is partly built into its reciprocal-NTP-loss objective, but the multi-task framework retains independent external evaluation.

self definitional [Eq. (1) in Section 3.1; Section 4 metric definitions; Table 1 EUL-only row]
"L_EUL = α × 1 / L_ntp(x_input, y_forget) (1). ... When the model's output closely aligns with the distribution of the information to be forgotten, the loss increases significantly; conversely, it remains low when the output deviates from that information. ... Membership Inference Attack Score (MIA): This evaluates the extent to which the relevant knowledge is retained or effectively forgotten."

Minimizing Eq. (1) is a reweighted gradient ascent on the next-token prediction loss for the forget targets, since ∇θ(α/L_ntp) = -α ∇θ L_ntp / L_ntp^2. MIA is the benchmark's measure of whether the model still retains or forgets that same forget-target knowledge. Therefore the EUL-only row's high MIA (0.993) is reporting the direction of the very quantity being optimized, not an independent confirmation of targeted unlearning. The improvement is built into the loss definition. MMLU and TAS are external, which is why the framework-level balance claim retains independent content, but the EUL-specific forgetting evidence is partially circular.

full rationale

The paper is a self-contained empirical system paper with no load-bearing self-citations and no imported uniqueness theorems. The central methodological novelty, EUL, is defined as the reciprocal of the next-token prediction loss on forget data, making the observed MIA improvement structurally aligned with the optimization objective rather than an independent outcome. This is a real but partial issue: no parameter is fitted directly to MIA, the submitted configuration actually reports MIA=0 for several rows, and the multi-task framework's value is assessed through external TAS and MMLU metrics not optimized by EUL. The absence of a gradient-ascent baseline weakens the 'more controllable' claim, but that is a comparison and validity concern rather than circularity. Overall, one construction-level alignment exists, but the central balanced-forgetting-and-retention claim does not fully reduce to its inputs.

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

The framework rests on hand-chosen hyperparameters and on the assumption that the evaluation metrics measure what they claim. No additional entities are introduced.

free parameters (6)

alpha (EUL scaling factor) = 1
Hand set to 1 in Section 4. Controls the magnitude of the inverse loss and thus the strength of forgetting.
learning_rate = 1e-4
Chosen by hand; Table 2 shows strong sensitivity, with lower learning rates yielding near-zero MIA.
epochs = 5
Chosen because of computational limits; Section 6.1 notes performance was still increasing at 5 epochs.
LoRA rank = 8
Standard LoRA configuration stated in Section 4.
LoRA alpha = 32
Standard LoRA configuration stated in Section 4.
batch_size = 32
Chosen by hand as part of the training setup in Section 4.

assumptions (3)

standard math The inverse of a positive loss is a monotone decreasing function, so gradient descent on EUL is equivalent to gradient ascent on the original loss with an adaptive step size.
This equivalence is implicit in Section 3.1, where EUL is claimed to induce forgetting during gradient descent; it is not stated or proven.
domain assumption The MIA score is a valid measure of unlearning, comparable with TAS and MMLU in the Final Score.
The paper interprets Table 1's MIA values as forgetting effectiveness. If MIA reflects global model degradation (as EUL-only's MMLU collapse suggests), the central evaluation is compromised.
domain assumption Fine-tuning on retain data preserves general capabilities without interfering with forgetting.
Section 5 concludes that RD is essential based on MMLU changes; this causal claim is asserted without a controlled decomposition beyond the ablation table.

how reviews work

0 comments

Cite this review

Pith. "Pith review of iShumei-Chinchunmei at SemEval-2025 Task 4: A balanced forgetting and retention multi-task framework using effective unlearning loss." pith.science (2026). https://pith.science/paper/PCWCKZIJ

@misc{pith2026250716263,
  author       = {Pith},
  title        = {Pith review of: iShumei-Chinchunmei at SemEval-2025 Task 4: A balanced forgetting and retention multi-task framework using effective unlearning loss},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PCWCKZIJ}},
  note         = {Machine review of arXiv:2507.16263}
}

read the original abstract

As the Large Language Model (LLM) gains widespread adoption, increasing attention has been given to the challenge of making LLM forget non-compliant data memorized during its pre-training. Machine Unlearning focuses on efficiently erasing sensitive information from LLM under limited computational resources. To advance research in this area, SemEval 2025 Task 4: "Unlearning Sensitive Content from Large Language Models" introduces three unlearning datasets and establishes a benchmark by evaluating both forgetting effectiveness and the preservation of standard capabilities. In this work, we propose a more controllable forgetting loss, Effective Unlearning Loss, and explore its integration with various techniques to achieve more efficient and controlled unlearning. Our system ultimately ranked 5th on the competition leaderboard.

Figures

Figures reproduced from arXiv: 2507.16263 by the authors.

**Figure 1.** The Overview of Our System. secures 5th place on the final leaderboard, demonstrating its effectiveness. 2 Background Research on knowledge unlearning in large language models (LLMs) is an emerging field, with fine-tuning-based unlearning being one of the most common method. This involves retraining the model on datasets containing specific target knowledge to weaken its memory of the unwanted information. One in… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

16 extracted references · 5 canonical work pages

[1]

Minseok Choi, Daniel Rim, Dohyun Lee, and Jaegul Choo. 2024. Snap: Unlearning selective knowledge in large language models with negative instructions. arXiv preprint arXiv:2406.12329

arXiv 2024
[2]

Ronen Eldan and Mark Russinovich. 2023. Who's harry potter? approximate unlearning in llms. arXiv preprint arXiv:2310.02238

arXiv 2023
[3]

Dirk Groeneveld, Iz Beltagy, Pete Walsh, Akshita Bhagia, Rodney Kinney, Oyvind Tafjord, Ananya Harsh Jha, Hamish Ivison, Ian Magnusson, Yizhong Wang, et al. 2024. Olmo: Accelerating the science of language models. arXiv preprint arXiv:2402.00838

arXiv 2024
[4]

Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. 2022. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3

2022
[5]

Joel Jang, Dongkeun Yoon, Sohee Yang, Sungmin Cha, Moontae Lee, Lajanugen Logeswaran, and Minjoon Seo. 2022. Knowledge unlearning for mitigating privacy risks in language models. arXiv preprint arXiv:2210.01504

arXiv 2022
[6]

Na Li, Chunyi Zhou, Yansong Gao, Hui Chen, Zhi Zhang, Boyu Kuang, and Anmin Fu. 2025. Machine unlearning: Taxonomy, metrics, applications, challenges, and prospects. IEEE Transactions on Neural Networks and Learning Systems

work page 2025
[7]

Lipton, and J

Pratyush Maini, Zhili Feng, Avi Schwarzschild, Zachary C. Lipton, and J. Zico Kolter. 2024. Tofu: A task of fictitious unlearning for llms

work page 2024
[8]

Anmol Mekala, Vineeth Dorna, Shreya Dubey, Abhishek Lalwani, David Koleczek, Mukund Rungta, Sadid Hasan, and Elita Lobo. 2024. Alternate preference optimization for unlearning factual knowledge in large language models. arXiv preprint arXiv:2409.13474

arXiv 2024

Show all 16 references

[9]

Youyang Qu, Xin Yuan, Ming Ding, Wei Ni, Thierry Rakotoarivelo, and David Smith. 2023. Learn to unlearn: A survey on machine unlearning. arXiv preprint arXiv:2305.07512

2023 arXiv
[10]

Anil Ramakrishna, Yixin Wan, Xiaomeng Jin, Kai-Wei Chang, Zhiqi Bu, Bhanukiran Vinzamuri, Volkan Cevher, Mingyi Hong, and Rahul Gupta. 2025 a . Lume: Llm unlearning with multitask evaluations. arXiv preprint arXiv:2502.15097

2025 arXiv
[11]

Anil Ramakrishna, Yixin Wan, Xiaomeng Jin, Kai-Wei Chang, Zhiqi Bu, Bhanukiran Vinzamuri, Volkan Cevher, Mingyi Hong, and Rahul Gupta. 2025 b . Semeval-2025 task 4: Unlearning sensitive content from large language models. arXiv preprint arXiv:2504.02883

2025 arXiv
[12]

Shaojie Shi, Xiaoyu Tan, Xihe Qiu, Chao Qu, Kexin Nie, Yuan Cheng, Wei Chu, Xu Yinghui, and Yuan Qi. 2024. Ulmr: Unlearning large language models via negative response and model parameter average. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language P...

2024
[13]

Akshaj Kumar Veldanda, Shi-Xiong Zhang, Anirban Das, Supriyo Chakraborty, Stephen Rawls, Sambit Sahu, and Milind Naphade. 2024. Llm surgery: Efficient knowledge unlearning and editing in large language models. arXiv preprint arXiv:2409.13054

2024 arXiv
[14]

Weiqi Wang, Zhiyi Tian, Chenhan Zhang, and Shui Yu. 2024. Machine unlearning: A comprehensive survey. arXiv preprint arXiv:2405.07406

2024 arXiv
[15]

online" 'onlinestring :=

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...
[16]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Minseok Choi, Daniel Rim, Dohyun Lee, and Jaegul Choo. 2024. Snap: Unlearning selective knowledge in large language models with negative instructions. arXiv preprint arXiv:2406.12329

arXiv 2024

[2] [2]

Ronen Eldan and Mark Russinovich. 2023. Who's harry potter? approximate unlearning in llms. arXiv preprint arXiv:2310.02238

arXiv 2023

[3] [3]

Dirk Groeneveld, Iz Beltagy, Pete Walsh, Akshita Bhagia, Rodney Kinney, Oyvind Tafjord, Ananya Harsh Jha, Hamish Ivison, Ian Magnusson, Yizhong Wang, et al. 2024. Olmo: Accelerating the science of language models. arXiv preprint arXiv:2402.00838

arXiv 2024

[4] [4]

Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. 2022. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3

2022

[5] [5]

Joel Jang, Dongkeun Yoon, Sohee Yang, Sungmin Cha, Moontae Lee, Lajanugen Logeswaran, and Minjoon Seo. 2022. Knowledge unlearning for mitigating privacy risks in language models. arXiv preprint arXiv:2210.01504

arXiv 2022

[6] [6]

Na Li, Chunyi Zhou, Yansong Gao, Hui Chen, Zhi Zhang, Boyu Kuang, and Anmin Fu. 2025. Machine unlearning: Taxonomy, metrics, applications, challenges, and prospects. IEEE Transactions on Neural Networks and Learning Systems

work page 2025

[7] [7]

Lipton, and J

Pratyush Maini, Zhili Feng, Avi Schwarzschild, Zachary C. Lipton, and J. Zico Kolter. 2024. Tofu: A task of fictitious unlearning for llms

work page 2024

[8] [8]

Anmol Mekala, Vineeth Dorna, Shreya Dubey, Abhishek Lalwani, David Koleczek, Mukund Rungta, Sadid Hasan, and Elita Lobo. 2024. Alternate preference optimization for unlearning factual knowledge in large language models. arXiv preprint arXiv:2409.13474

arXiv 2024