pith. sign in

arxiv: 2606.07688 · v1 · pith:ELBFFVVOnew · submitted 2026-06-05 · 💻 cs.IR · cs.AI· cs.CL· cs.LG

TRACER: Token ReAssignment for Concept ERasure in Generative Recommendation

Ziheng Chen , Jiali Cheng , Zezhong Fan , Hadi Amiri , Diyuan Wu , Gabriele Tolomei , Yang Zhang This is my paper

Pith reviewed 2026-06-27 21:02 UTC · model grok-4.3

classification 💻 cs.IR cs.AIcs.CLcs.LG
keywords generative recommendationconcept unlearningtoken reassignmentsemantic IDsautoregressive generationprivacy in recommender systemsforget-retain separation
0
0 comments X

The pith

TRACER erases target concepts in generative recommenders by reassigning shared semantic IDs to alternative tokens.

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

The paper aims to show that reassigning semantic IDs for concept-related items allows effective removal of sensitive concepts while avoiding the utility loss that occurs when shared IDs are suppressed directly. Generative recommendation systems treat item sequences as token sequences for autoregressive prediction, so shared abstract identifiers create direct conflicts between forgetting and retaining performance. If the approach works, it enables practical concept unlearning without the side effects seen in standard LLM unlearning techniques applied to recommendation data. Experiments on real-world datasets are presented to demonstrate that the reassignment plus coherence regularization removes the target concepts more cleanly than prior baselines.

Core claim

TRACER is an end-to-end unlearning framework that reassigns items associated with a target concept to alternative semantic ID tokens chosen to reduce overlap with retain items, then applies a coherence regularizer during fine-tuning to keep semantic consistency among the retained items; this produces models that no longer generate the forget-set concepts at rates seen before unlearning while recommendation metrics on retain sets remain higher than those achieved by direct suppression baselines.

What carries the argument

Token reassignment of semantic IDs, which moves concept-related items to new identifiers that minimize shared tokens between forget and retain sets.

If this is right

  • Target concepts can be removed from the model's generative distribution without direct suppression of shared tokens.
  • Recommendation utility on items outside the target concept stays closer to the original model than with existing unlearning methods.
  • A coherence regularizer keeps semantic relationships among retained items stable during the reassignment process.
  • The framework operates end-to-end on the autoregressive generation objective used in semantic-ID recommenders.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same reassignment idea might reduce interference when multiple concepts must be forgotten sequentially.
  • If semantic IDs are learned rather than fixed, the choice of which alternative tokens to assign could itself be optimized as part of the unlearning objective.
  • Systems that already use discrete semantic IDs for other generative tasks could adopt the same separation technique to isolate sensitive content.

Load-bearing premise

Reassigning shared semantic IDs to alternative tokens separates the forget set from the retain set enough to allow forgetting without creating new conflicts that degrade generation for retained items.

What would settle it

Run the generative model after TRACER training on a held-out test set containing both forget and retain items and measure whether the rate of generating forget-set items remains near the pre-unlearning level or whether NDCG and recall on retain items fall below the levels reported for the strongest baseline.

Figures

Figures reproduced from arXiv: 2606.07688 by Diyuan Wu, Gabriele Tolomei, Hadi Amiri, Jiali Cheng, Yang Zhang, Zezhong Fan, Ziheng Chen.

Figure 1
Figure 1. Figure 1: Retraining vs. Original GenRec model. (a): Original has highly overlapped semantic tokens, while Retraining mitigates overlap score significantly. (b): Original produces highly clustered embeddings, while Retraining produces dispersed embeddings. The overlap score is computed as the product of token frequencies in the forget and retain sets. autoregressively predicts the next item from the SID sequence of … view at source ↗
Figure 2
Figure 2. Figure 2: Overview of TRACER. a) Semantic token overlap in GenRec unlearning: removing conceptual knowledge from GenRec models is challenging, since many non-interpretable semantic tokens are shared by both forget and retain sets. b) Token reassignment: TRACER reassigns the semantic token of unlearning-affected items to the closest item of the same category, which preserves semantic coherence to the maximum extent. … view at source ↗
Figure 3
Figure 3. Figure 3: Probing Results on MOR. HR, NDCG, and MRR. For end-to-end backbones such as LETTER and ETEGRec, TRACER also consistently achieves the best performance in both concept erasing and utility preservation. TRACER is Efficient Moreover, TRACER runs 33.5% and 36.7% faster than PISCES and REP￾Noise, respectively. TRACER also requires the least time to achieve effective unlearning compared with existing methods. Ad… view at source ↗
Figure 4
Figure 4. Figure 4: Effect of Reassignment. References [1] Lucas Bourtoule, Varun Chandrasekaran, Christopher A Choquette-Choo, Hengrui Jia, Adelin Travers, Baiwu Zhang, David Lie, and Nicolas Papernot. 2021. Machine unlearning. In IEEE Symposium on Security and Privacy (SP). [2] Pengfei Cao, Chenhao Wang, Zhitao He, Hongbang Yuan, Jiachun Li, Yubo Chen, Kang Liu, Jun Zhao, et al. 2024. Rwku: Benchmarking real-world knowledge… view at source ↗
Figure 5
Figure 5. Figure 5: Ablation studies on K (left) and τ (right). Left: K=5 achieves the best trade-off between semantic coherence and unlearning effectiveness. Right: τ=0.005 yields the strongest performance; larger τ gradually degrades all metrics as the token-reassignment distribution becomes too diffuse. 28 [PITH_FULL_IMAGE:figures/full_fig_p028_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Ablation study on loss weights. (a) λ1 controls the forget loss LF: increasing λ1 improves forgetting but degrades retain performance when too large. (b) λ2 controls the coherence loss LCoh: higher values improve semantic similarity with diminishing returns. (c) λ3 controls the regularization loss Lreg: stronger regularization preserves retain quality but weakens forgetting [PITH_FULL_IMAGE:figures/full_f… view at source ↗
read the original abstract

Generative recommendation formulates next-item prediction as autoregressive generation over semantic ID (SID) sequences derived from users' historical interactions, making modern recommender systems structurally similar to large language models (LLMs). As privacy and safety concerns grow, these systems increasingly require concept unlearning to remove sensitive or harmful concepts associated with items. However, existing LLM unlearning methods cannot be directly applied to generative recommendation. Unlike word tokens with explicit semantics, SIDs are abstract identifiers that are often shared by both forget and retain items, leading to severe conflicts between concept removal and recommendation utility preservation. To address this challenge, we propose TRACER, an end-to-end concept unlearning framework based on token reassignment. Rather than directly suppressing shared SIDs, TRACER reassigns concept-related items to alternative tokens that better facilitate forgetting while minimizing side effects on retained items. We further introduce a coherence regularizer to preserve semantic consistency among retain items during unlearning. Experiments on real-world recommendation datasets demonstrate that TRACER effectively removes target concepts while substantially better preserving recommendation utility than existing unlearning baselines.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The manuscript proposes TRACER, an end-to-end concept unlearning framework for generative recommendation systems that model next-item prediction as autoregressive generation over semantic ID (SID) sequences. It identifies the core problem that SIDs are often shared between forget-set and retain-set items, creating conflicts that prevent direct application of LLM unlearning techniques. TRACER instead reassigns concept-related items to alternative tokens chosen to facilitate forgetting while minimizing side effects, and adds a coherence regularizer to preserve semantic consistency among retain items. The central claim is that this token-reassignment approach removes target concepts while substantially better preserving recommendation utility than existing unlearning baselines, as demonstrated on real-world recommendation datasets.

Significance. If the empirical results hold, the work would be significant for privacy and safety in generative recommenders, a setting that structurally resembles LLMs but cannot use standard unlearning methods because of the abstract, shared nature of SIDs. The targeted reassignment strategy plus coherence regularizer constitutes a concrete, domain-specific solution rather than a generic adaptation, and the paper correctly frames the shared-SID conflict as the load-bearing obstacle. Credit is due for identifying this incompatibility and for grounding the method in the generative-recommendation pipeline.

major comments (2)
  1. [Abstract] Abstract: the central claim of 'substantially better preserving recommendation utility than existing unlearning baselines' is stated without any quantitative metrics, dataset names, ablation results, or implementation details. Because the soundness of the contribution rests entirely on these unreported experiments, the claim cannot be evaluated from the provided text.
  2. [Abstract] The weakest assumption identified in the stress-test note—that reassigning shared SIDs to alternative tokens can separate forget and retain sets without introducing new conflicts or degrading the generative process for retained items—is treated as solved by construction in the abstract, but no derivation, algorithm, or preliminary analysis is supplied to show that the reassignment procedure is guaranteed to avoid such conflicts.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on the abstract. We agree that it can be strengthened with more concrete details on results and the method. We will revise the abstract in the next version. Point-by-point responses follow.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim of 'substantially better preserving recommendation utility than existing unlearning baselines' is stated without any quantitative metrics, dataset names, ablation results, or implementation details. Because the soundness of the contribution rests entirely on these unreported experiments, the claim cannot be evaluated from the provided text.

    Authors: We agree the abstract lacks quantitative support. The full manuscript reports these results in Sections 4 and 5 on real-world datasets (including metrics such as NDCG and HR improvements over baselines, plus ablations). We will revise the abstract to include key quantitative findings, dataset names, and references to the experimental setup so the central claim can be evaluated from the abstract. revision: yes

  2. Referee: [Abstract] The weakest assumption identified in the stress-test note—that reassigning shared SIDs to alternative tokens can separate forget and retain sets without introducing new conflicts or degrading the generative process for retained items—is treated as solved by construction in the abstract, but no derivation, algorithm, or preliminary analysis is supplied to show that the reassignment procedure is guaranteed to avoid such conflicts.

    Authors: The abstract is a high-level summary. Section 3 details the token reassignment algorithm, including the selection criterion for alternative tokens that minimizes overlap with retain-set items and the coherence regularizer that prevents degradation. Section 5 provides empirical analysis showing no new conflicts are introduced. We will revise the abstract to reference these design elements explicitly rather than implying the issue is solved by construction. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper presents an empirical method (TRACER) for concept unlearning via token reassignment and a coherence regularizer in generative recommendation. The abstract and description contain no equations, fitted parameters called predictions, self-definitional loops, or load-bearing self-citations that reduce any claimed result to its inputs by construction. The central claims rest on experiments on real-world datasets, which constitute external validation rather than internal reduction. This is a standard applied ML contribution without the derivation patterns that trigger circularity flags.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Only abstract available; no explicit free parameters, axioms, or invented entities can be extracted beyond the domain assumption that SIDs are shared across forget and retain sets.

axioms (1)
  • domain assumption Semantic IDs are abstract identifiers often shared by both forget-set and retain-set items
    Stated directly in the abstract as the source of conflict with existing unlearning methods.

pith-pipeline@v0.9.1-grok · 5749 in / 1071 out tokens · 37152 ms · 2026-06-27T21:02:23.666380+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

62 extracted references · 3 canonical work pages

  1. [1]

    Lucas Bourtoule, Varun Chandrasekaran, Christopher A Choquette-Choo, Hengrui Jia, Adelin Travers, Baiwu Zhang, David Lie, and Nicolas Papernot. 2021. Machine unlearning. InIEEE Symposium on Security and Privacy (SP)

    2021

  2. [2]

    Pengfei Cao, Chenhao Wang, Zhitao He, Hongbang Yuan, Jiachun Li, Yubo Chen, Kang Liu, Jun Zhao, et al. 2024. Rwku: Benchmarking real-world knowledge unlearning for large language models.Advances in Neural Information Processing Systems37 (2024), 98213–98263

    2024

  3. [3]

    Yinzhi Cao and Junfeng Yang. 2015. Towards making systems forget with machine unlearning. InProceedings of the IEEE Symposium on Security and Privacy

    2015

  4. [4]

    Chong Chen, Fei Sun, Min Zhang, and Bolin Ding. 2022. Recommendation unlearning. In Proceedings of the ACM web conference 2022. 2768–2777

    2022

  5. [5]

    Ziheng Chen, Jiali Cheng, Hadi Amiri, Kaushiki Nag, Lu Lin, Sijia Liu, Gabriele Tolomei, and Xiangguo Sun. 2025. FROG: Fair Removal on Graph. InProceedings of the 34th ACM International Conference on Information and Knowledge Management. 415–424

    2025

  6. [6]

    Ziheng Chen, Jiali Cheng, Zezhong Fan, Hadi Amiri, Yunzhi Yao, Xiangguo Sun, and Yang Zhang. 2026. CURE: Circuit-Aware Unlearning for LLM-based Recommendation.arXiv preprint arXiv:2604.04982(2026)

    Pith/arXiv arXiv 2026

  7. [7]

    Ziheng Chen, Jin Huang, Jiali Cheng, Yuchan Guo, Mengjie Wang, Lalitesh Morishetti, Kaushiki Nag, and Hadi Amiri. 2025. FUTURE: Flexible Unlearning for Tree Ensemble. InProceedings of the 34th ACM International Conference on Information and Knowledge Management. 4680– 4684

    2025

  8. [8]

    Jiali Cheng and Hadi Amiri. 2024. MU-bench: A multitask multimodal benchmark for machine unlearning.arXiv preprint arXiv:2406.14796(2024)

    arXiv 2024

  9. [9]

    Jiali Cheng and Hadi Amiri. 2025. MultiDelete for Multimodal Machine Unlearning. In Computer Vision – ECCV 2024, Aleš Leonardis, Elisa Ricci, Stefan Roth, Olga Russakovsky, Torsten Sattler, and Gül Varol (Eds.). Springer Nature Switzerland, Cham, 165–184

    2025

  10. [10]

    Jiali Cheng and Hadi Amiri. 2025. Speech Unlearning. InInterspeech 2025. 3209–3213. doi:10.21437/Interspeech.2025-2412

    work page doi:10.21437/interspeech.2025-2412 2025

  11. [11]

    Jiali Cheng and Hadi Amiri. 2025. Tool Unlearning for Tool-Augmented LLMs. InForty-second International Conference on Machine Learning

    2025

  12. [12]

    2026.Machine Unlearning Across Tasks and Modalities

    Jiali Cheng and Hadi Amiri. 2026.Machine Unlearning Across Tasks and Modalities. Springer Nature Switzerland, Cham, 227–240. doi:10.1007/978-3-032-17282-2_15

    work page doi:10.1007/978-3-032-17282-2_15 2026

  13. [13]

    Jiali Cheng, Ziheng Chen, Chirag Agarwal, and Hadi Amiri. 2026. Toward Understanding Unlearning Difficulty: A Mechanistic Perspective and Circuit-Guided Difficulty Metric.arXiv preprint arXiv:2601.09624(2026). 10

    arXiv 2026

  14. [14]

    Jiali Cheng, George Dasoulas, Huan He, Chirag Agarwal, and Marinka Zitnik. 2023. GN- NDelete: A General Strategy for Unlearning in Graph Neural Networks. InThe Eleventh International Conference on Learning Representations. https://openreview.net/forum? id=X9yCkmT5Qrl

    2023

  15. [15]

    Jiali Cheng, Rui Pan, and Hadi Amiri. 2026. Investigating Tool-Memory Conflicts in Tool- Augmented LLMs.arXiv preprint arXiv:2601.09760(2026)

    arXiv 2026

  16. [16]

    Jiaxin Deng, Shiyao Wang, Kuo Cai, Lejian Ren, Qigen Hu, Weifeng Ding, Qiang Luo, and Guorui Zhou. 2025. Onerec: Unifying retrieve and rank with generative recommender and iterative preference alignment.arXiv preprint arXiv:2502.18965(2025)

    Pith/arXiv arXiv 2025

  17. [17]

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

    arXiv 2023

  18. [18]

    Chongyu Fan, Jinghan Jia, Yihua Zhang, Anil Ramakrishna, Mingyi Hong, and Sijia Liu

  19. [19]

    Towards llm unlearning resilient to relearning attacks: A sharpness-aware minimization perspective and beyond.arXiv preprint arXiv:2502.05374(2025)

    arXiv 2025

  20. [20]

    Chongyu Fan, Jiancheng Liu, Yihua Zhang, Eric Wong, Dennis Wei, and Sijia Liu. 2024. SalUn: Empowering Machine Unlearning via Gradient-based Weight Saliency in Both Image Classifi- cation and Generation. InThe Twelfth International Conference on Learning Representations. https://openreview.net/forum?id=gn0mIhQGNM

    2024

  21. [21]

    Rohit Gandikota, Sheridan Feucht, Samuel Marks, and David Bau. 2024. Erasing conceptual knowledge from language models.arXiv preprint arXiv:2410.02760(2024)

    arXiv 2024

  22. [22]

    Rohit Gandikota, Joanna Materzy´nska, Jaden Fiotto-Kaufman, and David Bau. 2023. Erasing Concepts from Diffusion Models. InProceedings of the 2023 IEEE International Conference on Computer Vision

    2023

  23. [23]

    Aditya Golatkar, Alessandro Achille, and Stefano Soatto. 2020. Eternal sunshine of the spotless net: Selective forgetting in deep networks. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 9304–9312

    2020

  24. [24]

    Chuan Guo, Tom Goldstein, Awni Hannun, and Laurens Van Der Maaten. 2019. Certified data removal from machine learning models.arXiv preprint arXiv:1911.03030(2019)

    arXiv 2019

  25. [25]

    Gupta, Vittorio Orlandi, Chia-Rui Chang, Tianyu Wang, Marco Morucci, Pritam Dey, Thomas J

    Neha R. Gupta, Vittorio Orlandi, Chia-Rui Chang, Tianyu Wang, Marco Morucci, Pritam Dey, Thomas J. Howell, Xian Sun, Angikar Ghosal, Sudeepa Roy, Cynthia Rudin, and Alexander V olfovsky. 2025. dame-flame: A Python Package Providing Fast Interpretable Matching for Causal Inference.Journal of Statistical Software113, 2 (2025), 1–26. doi: 10.18637/jss. v113.i02

    work page doi:10.18637/jss 2025

  26. [26]

    Yoav Gur-Arieh, Clara Haya Suslik, Yihuai Hong, Fazl Barez, and Mor Geva. 2025. Precise in-parameter concept erasure in large language models. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 18997–19017

    2025

  27. [27]

    Ruidong Han, Bin Yin, Shangyu Chen, He Jiang, Fei Jiang, Xiang Li, Chi Ma, Mincong Huang, Xiaoguang Li, Chunzhen Jing, et al. 2025. Mtgr: Industrial-scale generative recommendation framework in meituan. InProceedings of the 34th ACM International Conference on Information and Knowledge Management. 5731–5738

    2025

  28. [28]

    Joel Jang, Dongkeun Yoon, Sohee Yang, Sungmin Cha, Moontae Lee, Lajanugen Logeswaran, and Minjoon Seo. 2023. Knowledge unlearning for mitigating privacy risks in language models. InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 14389–14408

    2023

  29. [29]

    Jinghan Jia, Yihua Zhang, Yimeng Zhang, Jiancheng Liu, Bharat Runwal, James Diffender- fer, Bhavya Kailkhura, and Sijia Liu. 2024. SOUL: Unlocking the Power of Second-Order Optimization for LLM Unlearning.arXiv preprint arXiv:2404.18239(2024). 11

    arXiv 2024

  30. [30]

    Xiaoyu Kong, Leheng Sheng, Junfei Tan, Yuxin Chen, Jiancan Wu, An Zhang, Xiang Wang, and Xiangnan He. 2025. Minionerec: An open-source framework for scaling generative recommendation.arXiv preprint arXiv:2510.24431(2025)

    arXiv 2025

  31. [31]

    Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. 2022. Autoregres- sive image generation using residual quantization. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 11523–11532

    2022

  32. [32]

    Nathaniel Li, Alexander Pan, Anjali Gopal, Summer Yue, Daniel Berrios, Alice Gatti, Justin D Li, Ann-Kathrin Dombrowski, Shashwat Goel, Long Phan, et al. 2024. The wmdp benchmark: Measuring and reducing malicious use with unlearning.arXiv preprint arXiv:2403.03218 (2024)

    Pith/arXiv arXiv 2024

  33. [33]

    Luyun Lin, Lixing Lin, Zhen Zhang, Moxuan Zheng, and Yiqing Wang. 2026. A V olume-Price- Adjusted MACD Trading Strategy with Sensitivity Calibration for US Equity Indices.arXiv preprint arXiv:2604.26063(2026)

    Pith/arXiv arXiv 2026

  34. [34]

    Luyun Lin and Yiqing Wang. 2025. Shap stability in credit risk management: A case study in credit card default model.Risks13, 12 (2025), 238

    2025

  35. [35]

    Lixing Lin, Juli You, Yue Li, Luyun Lin, Yiqing Wang, Zhen Zhang, and Moxuan Zheng

  36. [36]

    Reflect-Guard: Enhancing LLM Safeguards against Adversarial Prompts via Logical Self-Reflection.arXiv preprint arXiv:2605.24834(2026)

    Pith/arXiv arXiv 2026

  37. [37]

    Enze Liu, Bowen Zheng, Cheng Ling, Lantao Hu, Han Li, and Wayne Xin Zhao. 2025. Gen- erative recommender with end-to-end learnable item tokenization. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. 729–739

    2025

  38. [38]

    Sijia Liu, Yuanshun Yao, Jinghan Jia, Stephen Casper, Nathalie Baracaldo, Peter Hase, Yuguang Yao, Chris Yuhao Liu, Xiaojun Xu, Hang Li, et al. 2024. Rethinking machine unlearning for large language models.arXiv preprint arXiv:2402.08787(2024)

    arXiv 2024

  39. [39]

    Yue Liu, Zhiyuan Cheng, and Longying Lai. 2026. Improving the Completeness and Compara- bility of Segment Disclosures: A Large Language Model Approach. arXiv:2605.23924 [cs.CL] https://arxiv.org/abs/2605.23924

    Pith/arXiv arXiv 2026

  40. [40]

    Zhanyu Liu, Shiyao Wang, Xingmei Wang, Rongzhou Zhang, Jiaxin Deng, Honghui Bao, Jinghao Zhang, Wuchao Li, Pengfei Zheng, Xiangyu Wu, et al. 2025. Onerec-think: In-text reasoning for generative recommendation.arXiv preprint arXiv:2510.11639(2025)

    arXiv 2025

  41. [41]

    Ana Lucic, Harrie Oosterhuis, Hinda Haned, and Maarten De Rijke. 2022. FOCUS: Flexi- ble optimizable counterfactual explanations for tree ensembles. InProceedings of the AAAI conference on artificial intelligence, V ol. 36. 5313–5322

    2022

  42. [42]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer.Journal of machine learning research21, 140 (2020), 1–67

    2020

  43. [43]

    Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Hulikal Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Tran, Jonah Samost, et al . 2023. Recommender systems with generative retrieval.Advances in Neural Information Processing Systems36 (2023), 10299–10315

    2023

  44. [44]

    Hadi Reisizadeh, Jinghan Jia, Zhiqi Bu, Bhanukiran Vinzamuri, Anil Ramakrishna, Kai-Wei Chang, V olkan Cevher, Sijia Liu, and Mingyi Hong. 2026. Blur: A bi-level optimization approach for llm unlearning. InProceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers). 7043–7058

    2026

  45. [45]

    Domenic Rosati, Jan Wehner, Kai Williams, Łukasz Bartoszcze, David Atanasov, Robie Gon- zales, Subhabrata Majumdar, Carsten Maple, Hassan Sajjad, and Frank Rudzicz. 2024. Rep- resentation noising: A defence mechanism against harmful finetuning.Advances in Neural Information Processing Systems37 (2024), 12636–12676. 12

    2024

  46. [46]

    Smith, and Chiyuan Zhang

    Weijia Shi, Jaechan Lee, Yangsibo Huang, Sadhika Malladi, Jieyu Zhao, Ari Holtzman, Daogao Liu, Luke Zettlemoyer, Noah A. Smith, and Chiyuan Zhang. 2025. MUSE: Machine Unlearning Six-Way Evaluation for Language Models. InThe Thirteenth International Conference on Learning Representations.https://openreview.net/forum?id=TArmA033BU

    2025

  47. [47]

    Anvith Thudi, Gabriel Deza, Varun Chandrasekaran, and Nicolas Papernot. 2022. Unrolling sgd: Understanding factors influencing machine unlearning. In2022 IEEE 7th European Symposium on Security and Privacy (EuroS&P). IEEE, 303–319

    2022

  48. [48]

    Wenjie Wang, Honghui Bao, Xinyu Lin, Jizhi Zhang, Yongqi Li, Fuli Feng, See-Kiong Ng, and Tat-Seng Chua. 2024. Learnable item tokenization for generative recommendation. InProceed- ings of the 33rd ACM International Conference on Information and Knowledge Management. 2400–2409

    2024

  49. [49]

    Yaxuan Wang, Jiaheng Wei, Chris Yuhao Liu, Jinlong Pang, Quan Liu, Ankit Shah, Yujia Bao, Yang Liu, and Wei Wei. 2024. Llm unlearning via loss adjustment with only forget data. InThe Thirteenth International Conference on Learning Representations

    2024

  50. [50]

    Yaxuan Wang, Jiaheng Wei, Chris Yuhao Liu, Jinlong Pang, Quan Liu, Ankit Parag Shah, Yujia Bao, Yang Liu, and Wei Wei. [n. d.]. Llm unlearning via loss adjustment with only forget data, 2024.URL https://arxiv. org/abs/2410.11143([n. d.])

    arXiv 2024

  51. [51]

    Rongwu Xu, Zehan Qi, Zhijiang Guo, Cunxiang Wang, Hongru Wang, Yue Zhang, and Wei Xu

  52. [52]

    Knowledge conflicts for llms: A survey.arXiv preprint arXiv:2403.08319(2024)

    arXiv 2024

  53. [53]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. 2025. Qwen3 technical report.arXiv preprint arXiv:2505.09388(2025)

    Pith/arXiv arXiv 2025

  54. [54]

    Shuai Yuan, Xian Sun, Hannah Kim, Shuzhi Yu, and Carlo Tomasi. 2022. Optical Flow Training under Limited Label Budget via Active Learning. InEuropean Conference on Computer Vision (ECCV). Springer, 410–427

    2022

  55. [55]

    Zhang, Gaoyuan Du, Qianyi Sun, Shiqi Wang, Jiaxuan Li, and Xian Sun

    Johnny R. Zhang, Gaoyuan Du, Qianyi Sun, Shiqi Wang, Jiaxuan Li, and Xian Sun. 2026. Carbon-Aware Compute–Power Scheduling for AI Data Centers with Microgrid Prosumer Operations. arXiv:2605.03751 [cs.CE]https://arxiv.org/abs/2605.03751

    Pith/arXiv arXiv 2026

  56. [56]

    Yunqi Zhang and Shaileshh Bojja Venkatakrishnan. 2023. Kadabra: Adapting kademlia for the decentralized web. InInternational Conference on Financial Cryptography and Data Security. Springer, 327–345

    2023

  57. [57]

    Yunqi Zhang and Shaileshh Bojja Venkatakrishnan. 2025. Honeybee: Byzantine Tolerant Decentralized Peer Sampling with Verifiable Random Walks. InProceedings of the Twenty-sixth International Symposium on Theory, Algorithmic Foundations, and Protocol Design for Mobile Networks and Mobile Computing. 321–330

    2025

  58. [58]

    Yunqi Zhang and Shaileshh Bojja Venkatakrishnan. 2023. Rethinking incentive in payment channel networks. In2023 IEEE 43rd International Conference on Distributed Computing Systems Workshops (ICDCSW). IEEE, 61–66

    2023

  59. [59]

    Guorui Zhou, Jiaxin Deng, Jinghao Zhang, Kuo Cai, Lejian Ren, Qiang Luo, Qianqian Wang, Qigen Hu, Rui Huang, Shiyao Wang, et al. 2025. OneRec Technical Report.arXiv preprint arXiv:2506.13695(2025)

    arXiv 2025

  60. [60]

    Chunzheng Zhu, Jianxin Lin, Guanghua Tan, Ningbo Zhu, Kenli Li, Chunlian Wang, and Shengli Li. 2024. Advancing Ultrasound Medical Continuous Learning with Task-Specific Generalization and Adaptability. In2024 IEEE International Conference on Bioinformatics and Biomedicine (BIBM). IEEE, 3019–3025

    2024

  61. [61]

    Chunzheng Zhu, Yangfang Lin, Shen Chen, Yijun Wang, and Jianxin Lin. 2026. Medeyes: Learning dynamic visual focus for medical progressive diagnosis. InProceedings of the AAAI Conference on Artificial Intelligence, V ol. 40. 13916–13924. 13

    2026

  62. [62]

    Forget Leaked

    Chunzheng Zhu, Yangfang Lin, Jialin Shao, Jianxin Lin, and Yijun Wang. 2025. Pathology- Aware Prototype Evolution via LLM-Driven Semantic Disambiguation for Multicenter Diabetic Retinopathy Diagnosis. InProceedings of the 33rd ACM International Conference on Multime- dia. 9196–9205. 14 A Broader Impact This work addresses the growing need formachine unlea...

    arXiv 2025