Pith. sign in

REVIEW 4 major objections 4 minor 4 cited by

R-TOFU: Unlearning in Large Reasoning Models

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Erasing only a reasoning model's final answers leaves the forgotten knowledge intact inside its chain-of-thought; unlearning must target the reasoning trace itself, and evaluation must probe multiple decoding settings.

desk verdict Solid first benchmark for reasoning-model unlearning; the main qualitative finding holds up, but the method rankings rest on per-method early stopping and need a fixed-epoch check. read the letter →

arxiv 2505.15214 v2 pith:AOXK7XWT submitted 2025-05-21 cs.CL

Sangyeon Yoon , Wonje Jeung , Albert No This is my paper
classification cs.CL
keywords machineunlearninglargereasoningmodelschain-of-thoughtR-TOFUbenchmarkpreferenceoptimizationgradientascentstep-wiseevaluationdecodingstrategies
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

Large Reasoning Models (LRMs) do not just output answers; they produce a multi-step chain-of-thought, and private or copyrighted facts can live in those intermediate steps even when the final answer is suppressed. This paper argues that unlearning in such models therefore has to erase the reasoning trace itself, not just the final answer, and that evaluation must look inside the trace step by step. To make that case it builds R-TOFU, a benchmark that attaches realistic chain-of-thought annotations to the existing TOFU fictitious-author tasks, along with step-wise metrics and an LLM-judge score for residual knowledge. On this benchmark, answer-only objectives (gradient ascent on the answer, or replacing the answer with 'I don't know') leave substantial forget traces in the reasoning, whereas targeting the chain-of-thought—CoT-only unlearning for gradient methods, and a new 'Reasoned IDK' refusal style for preference optimization—forgets more effectively while preserving utility. The paper also shows a failure mode: decoding settings that suppress reasoning (ZeroThink, LessThink) can still reveal the supposedly forgotten content, so single-setting evaluation overestimates unlearning success.

What carries the argument

The central object is the chain-of-thought trace treated as a separate locus of knowledge: the paper defines an output as a triple (question, reasoning trace, answer) and makes the reasoning trace an explicit unlearning target. The machinery has three parts. First, R-TOFU, a benchmark that augments TOFU's 200 fictitious authors (4,000 question-answer pairs) with synthetic CoT traces generated by GPT-4o and styled on DeepSeek-R1 traces of the corresponding real-author questions. Second, step-wise evaluation, which aligns each ground-truth reasoning step to the most similar generated step before scoring (step-wise ROUGE, step-wise cosine similarity) and adds an LLM-as-judge score (0.0–1.0) for whether the trace retains, reconstructs, or hints at the forgotten answer. Third, the unlearning strategies themselves: gradient-based methods (GA, GD, KL) are run in CoT+Answer, Answer-only, and CoT-only variants, and preference optimization (PO) is run in Answer IDK, Direct IDK, and the proposed Reasoned IDK variant, which replaces the forget response with coherent but inconclusive reasoning.

What would settle it

Take a reasoning model fine-tuned on naturally occurring (non-synthetic) chain-of-thought traces for a forget set, apply Answer-only unlearning until final-answer ROUGE on the forget set drops to near zero, then read the model's reasoning under DefaultThink with step-wise metrics. If the step-wise retention scores are also near zero—the reasoning trace no longer reconstructs the forgotten facts—the paper's central claim that answer-level objectives leave residual CoT knowledge is falsified. A second direct check: take a model that appears fully unlearned under DefaultThink and query it with ZeroThink or LessThink prompting; if the forget answer cannot be elicited, the claimed decoding-dependent failure mode does not occur.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that knowledge in an LRM is distributed across the whole reasoning trajectory, and unlearning that touches only the final answer leaves residual, detectable knowledge inside the chain-of-thought. The paper establishes this with R-TOFU, built by rewriting TOFU's fictitious author questions for real authors, collecting DeepSeek-R1 reasoning traces on those real questions, and then having GPT-4o generate fictitious-author CoTs conditioned on the real-author reasoning style. Its step-wise metrics (step-aligned ROUGE, step-aligned cosine similarity, and an LLM-as-judge retention score) show that Answer-only and Answer IDK strategies score high on answer-level forgetting but low on reasoning-level forgetting, while CoT-only gradient ascent and Reasoned IDK achieve the best balance. A further finding is that decoding strategies ZeroThink and LessThink, which suppress or shorten the reasoning trace, can still reconstruct the forget answer even after DefaultThink evaluation looks fully unlearned.

Load-bearing premise

The benchmark's conclusions rest on the assumption that GPT-4o-generated chain-of-thought traces, styled on real-author reasoning, faithfully represent how actual large reasoning models embed and reason about forget knowledge; the paper itself flags this synthetic-trace limitation, along with testing on a single model architecture (DeepSeek-R1-Distill-Llama-8B).

Editorial extensions

If this is right

  • Unlearning pipelines for LRMs must include the chain-of-thought in the forget objective; answer-only gradients or answer replacement leave the reasoning trace intact.
  • Among gradient-based methods, CoT-only unlearning gives the strongest utility-forgetting balance, suggesting that disrupting the reasoning path is sufficient and that degrading answer tokens too is unnecessary.
  • Preference-optimization unlearning works best when the refusal itself keeps a coherent reasoning shape (Reasoned IDK), rather than a flat 'I don't know' that removes all reasoning.
  • Evaluation of LRM unlearning should report reasoning-level (step-wise) forget efficacy, because full-sequence metrics misread reordered or partially overlapping traces.
  • Unlearning success must be checked under multiple decoding settings (DefaultThink, ZeroThink, LessThink), since constrained decoding can reveal forgotten answers that standard decoding hides.

Reading between the lines

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

  • One could design unlearning objectives that explicitly minimize worst-case leakage across decoding strategies, treating ZeroThink and LessThink as adversarial probes rather than as evaluation afterthoughts—an extension the paper does not pursue.
  • The step-wise alignment idea applies beyond text: the same 'align before score' logic could detect residual knowledge in latents, attention maps, or activations, where no surface chain-of-thought is generated at all.
  • For regulators or auditors, this implies that answer-level compliance checks are insufficient for reasoning models; auditing queries must also inspect internal reasoning traces under varied reasoning budgets.
  • How much the synthetic CoT distribution matters is testable: re-run the R-TOFU pipeline with human-annotated or model-native reasoning traces and compare step-wise forget efficacy; if results shift materially, benchmark conclusions would need qualifying.
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.

Referee Report

4 major / 4 minor

Summary. The paper introduces R-TOFU, a benchmark for machine unlearning in large reasoning models (LRMs). It augments the TOFU fictitious-author task with synthetic chain-of-thought (CoT) traces generated by GPT-4o and conditioned on DeepSeek-R1 traces for real authors, then fine-tunes DeepSeek-R1-Distill-Llama-8B on the augmented data. The evaluation uses model utility (MU), answer-level forget efficacy (AFE), and a new step-wise CoT forget efficacy (CFE) that combines step-wise ROUGE, step-wise cosine similarity, and GPT-4o-as-judge scores. Experiments compare gradient-based (GA, GD, KL) and preference-optimization (PO) strategies over forget sets of 1%, 5%, and 10%. The central claims are that answer-only unlearning leaves residual CoT traces, that CoT-only gradient ascent gives the best utility--forget trade-off, that Reasoned IDK outperforms direct refusal in PO, and that decoding strategies such as ZeroThink and LessThink can reveal residual forgotten answers even when default decoding appears to have unlearned them.

Significance. If the empirical findings are robust, the paper makes a timely contribution: it provides the first benchmark explicitly targeting CoT-level unlearning in LRMs, introduces step-wise metrics that improve on full-sequence ROUGE and cosine similarity, identifies a concrete failure mode involving constrained decoding, and proposes a practical PO baseline (Reasoned IDK). The release of data, code, and full prompt templates is a strength for reproducibility. However, the paper's main quantitative comparisons rest on a per-method early-stopping protocol and single-run point estimates, so the specific rankings in Table 2 and the strength of Finding 4 are not yet fully supported. The synthetic construction of the CoT traces and the use of GPT-4o as both trace generator and judge also add an evaluation risk that needs to be addressed.

major comments (4)
  1. [Appendix C.1 and Table 3] The evaluation protocol confounds training duration with method efficacy. The paper trains each method for up to five epochs and reports results at the last epoch where model utility remains at or above 0.6, and Table 3 shows that the chosen epoch differs sharply across methods: Reasoned IDK is evaluated at epoch 2 in forget01 and epoch 1 in forget10, while Answer-only and Answer IDK methods are evaluated at epochs 4 or 5. Because AFE and CFE evolve at different rates (see Figure 3), a method stopped early can show lower AFE and higher CFE simply because it received less training, and an answer-only method evaluated later can show high AFE and low CFE partly because the extra epochs corrupt the reasoning trace. The paper never reports a fixed-epoch comparison, so the rankings in Table 2 that underlie Findings 1-3 do not yet separate method efficacy from stopping point. Please provide results for all methods at matched epochs (for example, each of the five epochs) and show that the qualitative ranking is invariant to the stopping rule, or otherwise justify that the chosen epochs are fair.
  2. [Section 7 and Appendix C.1] All reported numbers are point estimates from a single run, with no error bars, standard deviations, or number of seeds, and the learning rate is selected per scenario based on empirical performance (1e-5 for forget01 and 2e-6 for forget05 and forget10). The central quantitative gaps, such as the CFE difference between Answer-only (0.1838) and CoT-only (0.4608) in the forget01 GA row of Table 2, or the separation between DefaultThink and ZeroThink/LessThink curves in Figure 3, could be affected by run-to-run variance and by the choice of learning rate. Please report at least three seeds with means and standard deviations (or another variance estimate), and either report results under all tested learning rates or provide a principled justification for the per-scenario selection.
  3. [Section 4 and Appendix A.3] The CFE metric includes a GPT-4o-as-judge score, and the benchmark's CoT traces were also generated by GPT-4o. This is a correctness-risk concern rather than an automatic disqualifier, but it means the judge may reward stylistic similarity to the synthetic traces instead of detecting actual residual knowledge. The paper does not validate the judge against human annotations or compare it with an alternative judge model. Please provide such validation, or at a minimum show that the step-wise ROUGE and cosine similarity components alone reproduce the Finding 1 ranking, and report the agreement of the judge with those automatic metrics.
  4. [Section 8] Finding 4 is currently supported only by plots for the forget01 scenario, using ROUGE in Figure 3 and cosine similarity in Figure 13, with no numerical values at the reported evaluation epochs and no results for forget05 or forget10. The claim that Direct IDK appears to achieve complete forgetting by epoch 4 when judged with DefaultThink alone, yet still yields high scores under ZeroThink and LessThink, should be backed by explicit scores at the chosen epoch. Please quantify the decoding gap numerically and verify whether the failure mode persists across the larger forget sets.
minor comments (4)
  1. [References and Appendix figures] There are several typographical issues: 'V oigt and V on dem Bussche' in the references has stray spaces, 'Tremblay v. OpenAI, Inc.,,' has a double comma, Figure 8 contains 'FRrench', and the prompt templates in Appendix A use 'Y ou' with an inserted space.
  2. [Table 1] The up arrows in Table 1 are unexplained; please state what they indicate (presumably improvement over the pretrained model) and clarify the intended comparisons.
  3. [Section 5.2] The step-wise evaluation aligns each ground-truth step with 'its most similar generated step,' but the exact matching procedure is not specified: which similarity function is used for alignment, how ties are broken, and how steps with different cardinality are handled.
  4. [Appendix E] The normalization of the underlying metrics to [0,1] is mentioned but not defined; please specify the clipping or normalization procedure for ROUGE, cosine similarity, token entropy, and entailment score so that the harmonic mean aggregation is reproducible.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; the benchmark findings are empirical measurements, with only mild self-referentiality from GPT-4o generating and judging CoT traces.

full rationale

R-TOFU is an externally grounded benchmark: the forget/retain structure is inherited from TOFU, the unlearning baselines optimize standard losses (gradient ascent, KL divergence, preference optimization) whose objectives do not include the evaluation metrics, and the headline claims (answer-only unlearning leaves CoT traces; CoT-only gives the best gradient-ascent trade-off; Reasoned IDK balances utility and forgetting; ZeroThink/LessThink reveal residual knowledge) are reported empirical results rather than derivations. No equation in the paper reduces a target result to a fitted parameter or to a self-citation. The only self-referential element is that GPT-4o generated the synthetic CoT traces and also serves as the LLM-as-judge for CFE; the paper itself acknowledges this limitation: 'the R-TOFU benchmark is constructed using synthetic CoT traces generated by GPT-4o, which may not fully capture real-world reasoning behaviors.' This is a validity concern, not a constructional equivalence, and it does not force any particular method ranking. The early-stopping rule in Appendix C.1 is a methodological threat to the method ranking, but it is not circularity. Accordingly the circularity score is low, reflecting only mild self-referentiality rather than any derivation that reduces to its inputs.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

No new physical or ontological entities are postulated; R-TOFU is a dataset and Reasoned IDK is a training strategy. The main load-bearing assumptions are the realism of the synthetic reasoning traces, the validity of the step-wise and judge-based metrics, and the transferability from one distilled model to other LRMs.

free parameters (2)
  • Early-stopping MU threshold = 0.6
    Appendix C.1 selects the final epoch per method where model utility stays at or above 0.6; this hand-set threshold changes which results get reported.
  • Per-scenario learning rate = 1e-5 (forget01), 2e-6 (forget05 and forget10)
    Appendix C.1 says learning rates were selected based on empirical performance, a hand-tuned choice that can favor certain methods in the comparison.
assumptions (3)
  • domain assumption Synthetic GPT-4o CoT traces conditioned on real-author examples are realistic proxies for LRM reasoning.
    Used in Section 4 dataset construction; the benchmark's validity rests on this. The Limitations section admits the traces 'may not fully capture real-world reasoning behaviors.'
  • domain assumption Step-wise metric alignment and GPT-4o-as-judge scores faithfully measure residual reasoning knowledge.
    Section 5.2 defines CFE via these measures; if the judge or alignment is biased, the forgetting conclusions shift.
  • domain assumption Behavior observed on the fine-tuned DeepSeek-R1-Distill-Llama-8B transfers to LRMs generally.
    All experiments use one model (Section 7.1); the Limitations section flags this as an open question.

how reviews work

0 comments
Cite this review

Pith. "Pith review of R-TOFU: Unlearning in Large Reasoning Models." pith.science (2026). https://pith.science/paper/AOXK7XWT

@misc{pith2026250515214,
  author       = {Pith},
  title        = {Pith review of: R-TOFU: Unlearning in Large Reasoning Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AOXK7XWT}},
  note         = {Machine review of arXiv:2505.15214}
}
read the original abstract

Large Reasoning Models (LRMs) embed private or copyrighted information not only in their final answers but also throughout multi-step chain-of-thought (CoT) traces, making reliable unlearning far more demanding than in standard LLMs. We introduce Reasoning-TOFU (R-TOFU), the first benchmark tailored to this setting. R-TOFU augments existing unlearning tasks with realistic CoT annotations and provides step-wise metrics that expose residual knowledge invisible to answer-level checks. Using R-TOFU, we carry out a comprehensive comparison of gradient-based and preference-optimization baselines and show that conventional answer-only objectives leave substantial forget traces in reasoning. We further propose Reasoned IDK, a preference-optimization variant that preserves coherent yet inconclusive reasoning, achieving a stronger balance between forgetting efficacy and model utility than earlier refusal styles. Finally, we identify a failure mode: decoding variants such as ZeroThink and LessThink can still reveal forgotten content despite seemingly successful unlearning, emphasizing the need to evaluate models under diverse decoding settings. Together, the benchmark, analysis, and new baseline establish a systematic foundation for studying and improving unlearning in LRMs while preserving their reasoning capabilities.

Figures

Figures reproduced from arXiv: 2505.15214 by the authors.

Figure 1
Figure 1. Overview of LRMs Unlearning. (a) Concept. Unlike standard LLMs, LRMs require unlearning both the final answer and the associated reasoning trace. (b) Unlearning Strategies. The top row illustrates gradient ascent-based strategies, while the bottom row presents preference optimization-based strategies. Red indicates forget information that the model is trained to suppress, while blue indicates non-forget responses (e… view at source ↗
Figure 2
Figure 2. Detailed Analysis of CFE Results in forget01. Step-wise ROUGE-L scores, step-wise Cosine Similarity, and LLM-as-Judge evaluations across four unlearning methods, showing reasoning trace unlearning efficacy. 7.2 Main Results Finding 1: Unlearning only the final answer is insufficient to remove forget information embedded in the reasoning process. Strategies that target only the final answer, such as Answer-only unlea… view at source ↗
Figure 3
Figure 3. ROUGE scores of forget answers under different decoding strategies in the forget01 scenario. We plot ROUGE across unlearning epochs under DefaultThink, ZeroThink, and LessThink. formulated within the CoT critically influences un￾learning performance. As shown in [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Prompt used to rewrite fictitious author questions into real-author questions while preserving the original [PITH_FULL_IMAGE:figures/full_fig_p014_4.png]
Figure 5
Figure 5. Figure 5: Prompt used to generate new chain-of-thought (CoT) traces for the original TOFU fictitious question [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: Prompts used for Reasoned IDK generation. [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: Prompts used for LLM-as-Judge evaluation. [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 9
Figure 9. Figure 9: Example of a generated CoT with the un￾learning target reasoning successfully erased. Com￾parison of full-sequence and step-wise evaluation based on cosine similarity. (a) shows the full-sequence set￾ting where all reasoning steps are concatenated. (b) shows the step-w…
Figure 10
Figure 10. Figure 10: Example of a reasoning-augmented (CoT + Answer) data point in the R-TOFU. [PITH_FULL_IMAGE:figures/full_fig_p017_10.png]
Figure 11
Figure 11. Figure 11: Example of a forget set response under the Reasoned IDK strategy (CoT + Answer). [PITH_FULL_IMAGE:figures/full_fig_p018_11.png]
Figure 12
Figure 12. Figure 12: Example of a forget set response under the Direct IDK strategy (CoT + Answer). [PITH_FULL_IMAGE:figures/full_fig_p018_12.png]
Figure 13
Figure 13. Figure 13: Cosine Similarity scores of forget answers under different decoding strategies in the forget01 scenario. We plot CS across unlearning epochs under DefaultThink, ZeroThink, and LessThink. Method Strategy forget01 forget05 forget10 GA CoT+Answer epoch 3 epoch 5 epoch 3 …

Discussion (0). Sign in to comment.

Forward citations

Cited by 4 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Selective Forgetting for Large Reasoning Models

    cs.AI 2026-04 conditional novelty 6.0 of 10

    FRUL uses multi-LLM RAG to isolate and replace sensitive CoT segments, then applies a suppress-and-reinforce loss so LRMs forget targeted knowledge while retaining structured reasoning.

  2. R1-ACT: Efficient Reasoning Model Safety Alignment by Activating Safety Knowledge

    cs.AI 2025-08 conditional novelty 5.0 of 10

    Adding an explicit 'is this harmful?' step to the reasoning chain, trained on just 1,000 examples, substantially reduces harmful responses from reasoning models while roughly preserving benchmark reasoning performance.

  3. SoK: Machine Unlearning for Large Language Models

    cs.LG 2025-06 conditional novelty 5.0 of 10

    A new taxonomy for LLM unlearning distinguishes removal-intended from suppression-intended methods, and argues that gradient ascent methods functionally behave like suppression.

  4. A Comprehensive Survey on Trustworthiness in Reasoning with Large Language Models

    cs.CL 2025-09 conditional novelty 4.0 of 10

    A structured literature survey concluding that reasoning capabilities do not automatically make LLMs more trustworthy and can introduce new vulnerabilities in safety, robustness, and privacy.

Reference graph

Works this paper leans on

47 extracted references · 8 canonical work pages · cited by 4 Pith papers

  1. [1]

    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 INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    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 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, and 1 others. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

    arXiv 2023

  4. [4]

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, and 1 others. 2023. Qwen technical report. arXiv preprint arXiv:2309.16609

    arXiv 2023

  5. [5]

    Yinzhi Cao and Junfeng Yang. 2015. Towards making systems forget with machine unlearning. In 2015 IEEE symposium on security and privacy, pages 463--480. IEEE

    2015

  6. [6]

    Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-Voss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, and 1 others. 2021. Extracting training data from large language models. In 30th USENIX security symposium (USENIX Security 21), pages 2633--2650

    2021

  7. [7]

    Daniel Cer, Mona Diab, Eneko Agirre, Inigo Lopez-Gazpio, and Lucia Specia. 2017. Semeval-2017 task 1: Semantic textual similarity-multilingual and cross-lingual focused evaluation. arXiv preprint arXiv:1708.00055

    arXiv 2017

  8. [8]

    Jiaao Chen and Diyi Yang. 2023. Unlearn what you want to forget: Efficient unlearning for llms. In EMNLP

    work page 2023

Show all 47 references
  1. [10]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

    2024 arXiv

  2. [11]

    OpenAI, Inc.,

    Tremblay v. OpenAI, Inc.,. 2023. 23-cv-03416-AMO, (N.D. Cal.)

    2023

  3. [12]

    Aditya Golatkar, Alessandro Achille, and Stefano Soatto. 2020. Eternal sunshine of the spotless net: Selective forgetting in deep networks. In CVPR

    2020

  4. [13]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, and 1 others. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948

    2025 arXiv

  5. [14]

    Shengyuan Hu, Yiwei Fu, Steven Wu, and Virginia Smith. 2025. Unlearning or obfuscating? jogging the memory of unlearned llms via benign relearning. In The Thirteenth International Conference on Learning Representations

    2025

  6. [15]

    Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. 2022. Editing models with task arithmetic. arXiv preprint arXiv:2212.04089

    2022 arXiv

  7. [16]

    Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, and 1 others. 2024. Openai o1 system card. arXiv preprint arXiv:2412.16720

    2024 arXiv

  8. [17]

    Wonje Jeung, Sangyeon Yoon, Hyesoo Hong, Soeun Kim, Seungju Han, Youngjae Yu, and Albert No. 2025 a . Dusk: Do not unlearn shared knowledge. arXiv preprint arXiv:2505.15209

    2025 arXiv

  9. [18]

    Wonje Jeung, Sangyeon Yoon, Minsuk Kahng, and Albert No. 2025 b . Safepath: Preventing harmful reasoning in chain-of-thought via early alignment. arXiv preprint arXiv:2505.14667

    2025

  10. [19]

    Jinghan Jia, Yihua Zhang, Yimeng Zhang, Jiancheng Liu, Bharat Runwal, James Diffenderfer, Bhavya Kailkhura, and Sijia Liu. 2024. Soul: Unlocking the power of second-order optimization for llm unlearning. In EMNLP

    2024

  11. [20]

    Fengqing Jiang, Zhangchen Xu, Yuetai Li, Luyao Niu, Zhen Xiang, Bo Li, Bill Yuchen Lin, and Radha Poovendran. 2025. Safechain: Safety of language models with long chain-of-thought reasoning capabilities. arXiv preprint arXiv:2502.12025

    2025 arXiv

  12. [21]

    Zhuoran Jin, Pengfei Cao, Chenhao Wang, Zhitao He, Hongbang Yuan, Jiachun Li, Yubo Chen, Kang Liu, and Jun Zhao. 2024. Rwku: Benchmarking real-world knowledge unlearning for large language models. arXiv preprint arXiv:2406.10890

    2024 arXiv

  13. [22]

    Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74--81

    2004

  14. [23]

    Bo Liu, Qiang Liu, and Peter Stone. 2022. Continual learning and private unlearning. In Conference on Lifelong Learning Agents, pages 243--254. PMLR

    2022

  15. [24]

    Zhenhua Liu, Tong Zhu, Chuanyuan Tan, and Wenliang Chen. 2024. Learning to refuse: Towards mitigating privacy risks in llms. arXiv preprint arXiv:2407.10058

    2024 arXiv

  16. [25]

    Ximing Lu, Seungju Han, David Acuna, Hyunwoo Kim, Jaehun Jung, Shrimai Prabhumoye, Niklas Muennighoff, Mostofa Patwary, Mohammad Shoeybi, Bryan Catanzaro, and 1 others. 2025. Retro-search: Exploring untaken paths for deeper and efficient reasoning. arXiv preprint arXiv:2504.04383

    2025 arXiv

  17. [26]

    Yingwei Ma, Yue Liu, Yue Yu, Yuanliang Zhang, Yu Jiang, Changjian Wang, and Shanshan Li. 2023. At which training stage does code data help llms reasoning? arXiv preprint arXiv:2309.16298

    2023 arXiv

  18. [27]

    Yingzi Ma, Jiongxiao Wang, Fei Wang, Siyuan Ma, Jiazhao Li, Jinsheng Pan, Xiujun Li, Furong Huang, Lichao Sun, Bo Li, and 1 others. 2024. Benchmarking vision language model unlearning via fictitious facial identity dataset. arXiv preprint arXiv:2411.03554

    2024 arXiv

  19. [28]

    Pratyush Maini, Zhili Feng, Avi Schwarzschild, Zachary C Lipton, and J Zico Kolter. 2024. Tofu: A task of fictitious unlearning for llms. arXiv preprint arXiv:2401.06121

    2024 arXiv

  20. [29]

    Martin Pawelczyk, Seth Neel, and Himabindu Lakkaraju. 2023. In-context unlearning: Language models as few shot unlearners. arXiv preprint arXiv:2310.07579

    2023 arXiv

  21. [30]

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model. In NeurIPS

    2023

  22. [31]

    Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084

    2019 arXiv

  23. [32]

    Matthew Renze and Erhan Guven. 2024. Self-reflection in llm agents: Effects on problem-solving performance. arXiv preprint arXiv:2405.06682

    2024 arXiv

  24. [33]

    Mark Russinovich and Ahmed Salem. 2025. Obliviate: Efficient unmemorization for protecting intellectual property in large language models. arXiv preprint arXiv:2502.15010

    2025 arXiv

  25. [34]

    Damien Sileo. 2023. tasksource: A dataset harmonization framework for streamlined nlp multi-task learning and evaluation. arXiv preprint arXiv:2301.05948

    2023 arXiv

  26. [35]

    Paul Voigt and Axel Von dem Bussche. 2017. The eu general data protection regulation (gdpr). A practical guide, 1st ed., Cham: Springer International Publishing, 10(3152676):10--5555

    2017

  27. [36]

    Jiaan Wang, Yunlong Liang, Fandong Meng, Zengkui Sun, Haoxiang Shi, Zhixu Li, Jinan Xu, Jianfeng Qu, and Jie Zhou. 2023. Is chatgpt a good nlg evaluator? a preliminary study. arXiv preprint arXiv:2303.04048

    2023 arXiv

  28. [37]

    Boyi Wei, Weijia Shi, Yangsibo Huang, Noah A Smith, Chiyuan Zhang, Luke Zettlemoyer, Kai Li, and Peter Henderson. 2024. Evaluating copyright takedown methods for language models. arXiv preprint arXiv:2406.18664

    2024 arXiv

  29. [38]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, and 1 others. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824--24837

    2022

  30. [39]

    Jin Yao, Eli Chien, Minxin Du, Xinyao Niu, Tianhao Wang, Zezhou Cheng, and Xiang Yue. 2024 a . Machine unlearning of pre-trained large language models. arXiv preprint arXiv:2402.15159

    2024 arXiv

  31. [40]

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. 2023 a . Tree of thoughts: Deliberate problem solving with large language models. Advances in neural information processing systems, 36:11809--11822

    2023

  32. [41]

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023 b . React: Synergizing reasoning and acting in language models. In International Conference on Learning Representations (ICLR)

    2023

  33. [42]

    Yuanshun Yao, Xiaojun Xu, and Yang Liu. 2024 b . Large language model unlearning. Advances in Neural Information Processing Systems, 37:105425--105475

    2024

  34. [43]

    Xiaojian Yuan, Tianyu Pang, Chao Du, Kejiang Chen, Weiming Zhang, and Min Lin. 2025. A closer look at machine unlearning for large language models. In ICLR

    2025

  35. [44]

    Qingbin Zeng, Qinglong Yang, Shunan Dong, Heming Du, Liang Zheng, Fengli Xu, and Yong Li. 2024. Perceive, reflect, and plan: Designing llm agent for goal-directed city navigation without instructions. arXiv preprint arXiv:2408.04168

    2024 arXiv

  36. [45]

    Dan Zhang, Sining Zhoubian, Ziniu Hu, Yisong Yue, Yuxiao Dong, and Jie Tang. 2024 a . Rest-mcts*: Llm self-training via process reward guided tree search. Advances in Neural Information Processing Systems, 37:64735--64772

    2024

  37. [46]

    Ruiqi Zhang, Licong Lin, Yu Bai, and Song Mei. 2024 b . Negative preference optimization: From catastrophic collapse to effective unlearning. In COLM

    2024

  38. [47]

    Yizhe Zhang, Michel Galley, Jianfeng Gao, Zhe Gan, Xiujun Li, Chris Brockett, and Bill Dolan. 2018. Generating informative and diverse conversational responses via adversarial information maximization. Advances in Neural Information Processing Systems, 31

    2018

  39. [48]

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, and 1 others. 2023. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595--46623

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.