Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Mitigating Heterogeneous Token Overfitting in LLM Knowledge Editing

T0 review · 4 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read LLM knowledge edits overfit tokens at different rates; a smoothed per-token loss called OVERTONE restores the model's ability to reason with edited facts.

desk verdict A practical and well-validated loss tweak for knowledge editing, with theory that overreaches; the empirical contribution stands and deserves reviewer time. read the letter →

arxiv 2502.00602 v2 pith:QMV6DO47 submitted 2025-02-02 cs.CL cs.LG

Tianci Liu , Ruirui Li , Zihan Dong , Hui Liu , Xianfeng Tang , Qingyu Yin , Linjun Zhang , Haoyu Wang
show 1 more author
Jing Gao
This is my paper · ORCID
classification cs.CLcs.LG MSC 68T5068T07
keywords heterogeneoustokenoverfittingknowledgeeditinglargelanguagemodelsportabilitylocalitytoken-levelsmoothingdirectpreferenceoptimizationlossclipping
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

Knowledge editing updates a fact inside a large language model, but edited models typically lose the ability to reason from the new fact, a failure measured as portability loss. This paper identifies a cause: heterogeneous token overfitting (HTO), where the standard cross-entropy objective fits the tokens of the edited text at different rates, over-training the easy tokens while the hard tokens are still being learned. The paper proposes OVERTONE, which replaces each one-hot token target with an adaptive mixture of the correct token and the model's own noise-filtered prediction, and clips each token's loss once it is sufficiently fit. Across four editing methods (FT-M, LoRA, MELO, WISE) and two model families (LLaMA-2 and LLaMA-3), the authors report that this single loss swap improves portability and locality without hurting reliability, with theoretical support that the update direction is closer to ideal retraining at negligible added cost.

What carries the argument

The load-bearing object is the OVERTONE loss and its constructed target distribution. For each token position $i$ with context $c_i$, the target is $\pi_{\mathrm{tar}} = \lambda\delta_{y_i} + (1-\lambda)\pi_{\mathrm{flt}}$, where $\pi_{\mathrm{flt}}$ is the model's current softmax prediction over logits that survive a top-$n\sigma$ filter (all logits more than $n\sigma$ below the maximum are set to $-\infty$ and the rest renormalized); the mixture is used only if it still assigns the highest probability to the correct token $y_i$, otherwise the target falls back to the one-hot $\delta_{y_i}$. The loss is the forward KL divergence from this target to the model's prediction, clipped below at $\epsilon$ to impose per-token early stopping. Lemma A.3 decomposes the cross-entropy of the mixture into $\lambda$ times the CE to the ground-truth token plus $(1-\lambda)$ times the CE to the filtered distribution, which is the identity that lets the paper treat OVERTONE as an interpolation between standard CE training and distillation of the model's own denoised knowledge, and that underlies both the influence-function analysis of the update direction and the DPO connection.

What would settle it

Compute the gradients $a$, $b$, and $c$ defined in Eq. (9) on real LLaMA-2 and LLaMA-3 edit instances and check Assumption A.7's bound on $\cos(b, c)$; if typical edits violate it, the claimed provably-better update direction does not hold in the regime the experiments use, even though the method might still work empirically.

Watch

Extended reading notes

Core claim

The paper's central claim is that knowledge-editing overfitting is a token-level phenomenon: on a typical edit, some tokens of the new answer are already almost predictable from the model's pretrained knowledge while others carry large initial loss, so maximizing the likelihood of the whole sequence indiscriminately overfits the easy tokens and pushes the parameter update away from what retraining on the new fact would do. OVERTONE counteracts this by training with the loss $\ell_{\text{OVERTONE}}(\theta) = \sum_i \max(D_{\mathrm{KL}}[\pi_{\mathrm{tar}}(y|c_i) \| \pi_\theta(y|c_i)], \epsilon)$, where the target $\pi_{\mathrm{tar}}$ mixes the ground-truth token distribution $\delta_{y_i}$ with a top-$n\sigma$-filtered version of the model's own current prediction, and the clipping at $\epsilon$ stops a token from contributing gradient once its predicted distribution is close enough to the target. The paper argues this objective gives a gradient direction closer to the ideal retraining gradient and induces smaller perturbations on unrelated knowledge, and that it is equivalent to optimizing an unbiased estimate of a DPO objective with an added KL penalty, without any preference data. Empirically, OVERTONE raises portability and locality across four editing methods on ZsRE, WikiData recent and counterfact, WikiBio, and MQuAKE, in both single and continual editing, and turns plain LoRA into a competitive continual editor.

Load-bearing premise

The proof that OVERTONE beats plain cross-entropy rests on an unverified geometric premise: that the gradient from the model's old knowledge and the gradient from the smoothed target point in nearly the same direction, and that the smoothed gradient has the smaller norm; neither condition is checked on the actual models and edits in the experiments.

Editorial extensions

If this is right

  • With OVERTONE, all four tested editing methods beat the analytic baselines ROME and MEMIT on WikiData recent, whereas only some did with ordinary cross-entropy training.
  • Plain LoRA plus OVERTONE becomes a viable continual-editing method: on ZsRE with sequences of 10 and 100 edits it approaches specialized methods like MELO and WISE, retaining an average near 59 on LLaMA-2 where vanilla LoRA collapses to 0.62.
  • OVERTONE's gains concentrate in portability and locality while reliability stays at 100.0 in most single-edit settings, so the method resolves the previous trade-off between keeping an edit reliable and keeping the model able to reason with it.
  • Because the loss reduces to cross-entropy at $\epsilon = 0$ and $\lambda = 1$ and costs only $O(|V|)$ extra computation per token, it can be adopted by any gradient-based editing method without architectural change.

Reading between the lines

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

  • The paper only proves existence of some $\lambda$ that improves the update direction, yet all experiments fix $\lambda = 0.1$; a per-edit sweep of $\lambda$ would show whether the theoretical improvement is actually attained at the deployed value.
  • HTO should grow with edit length, since longer texts contain more tokens being fit at disparate rates; a testable extension is that OVERTONE's portability gains over cross-entropy increase on long knowledge statements.
  • The authors note that reporting averages over editing samples understates uncertainty; running multiple seeds per sample would show whether the portability gains, which reach roughly threefold on ZsRE, are stable rather than seed-dependent.
  • The paper suggests machine unlearning as a natural next target; OVERTONE's selective token fitting could forget a specific fact while preserving the surrounding text, a transfer that would test whether the mechanism generalizes beyond knowledge editing.
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 proposes OVERTONE, a token-level loss modification for knowledge editing (KE). It identifies "heterogeneous token overfitting" (HTO): when an LLM is edited on a single instance, different tokens are fitted at different rates, and the standard cross-entropy objective indiscriminately maximizes all token probabilities, causing overfitting and degrading portability. OVERTONE replaces each one-hot target with a clipped forward-KL loss against an adaptively smoothed target π_tar = λδ_yi + (1−λ)π_flt, where π_flt is a top-nσ-filtered version of the model's own prediction, and the mixture is used only if it assigns the ground-truth token the highest probability. The paper claims three theoretical merits: OVERTONE generalizes CE, is computationally cheap, provides a better parameter-update direction and smaller perturbation on unrelated data, and is connected to DPO without requiring preference pairs. Experiments apply OVERTONE to FT-M, LoRA, MELO, and WISE on LLaMA-2-7b and LLaMA-3-8b across ZsRE, WikiData recent/counterfact, WikiBio, and MQuAKE, reporting large portability and locality improvements in both single and continual editing.

Significance. If the empirical tables are taken at face value, the paper makes a useful and practical contribution: a simple, drop-in loss replacement that markedly improves portability and locality across four editing methods and two model families without sacrificing reliability, and that costs less than a forward pass. The HTO phenomenon, evidenced by token-level loss plots, is a plausible and testable explanation, and the use of externally benchmarked metrics avoids definitional circularity. The theoretical analyses in Appendix A give a concrete framework, but their assumptions are not verified on the evaluated models and the deployed hyperparameters are not covered by the theorems; the theory therefore currently functions as motivation rather than as a proof of superiority. The manuscript would also be strengthened by repeated runs or confidence intervals before the magnitude of the gains is treated as established.

major comments (4)
  1. [A.3.1, Theorem A.10 and Appendix B.1] Theorem A.10 establishes only the existence of some λ ∈ [0,1] (specifically λ = ||c||/(||b||+||c||)) for which the OVERTONE gradient has higher cosine with the ideal direction G than CE, but Algorithm 1 and every experiment use the fixed λ = 0.1 (Appendix B.1). No argument or experiment shows that 0.1 lies in the favorable set, so the 'better update direction' theorem does not cover the method as actually deployed.
  2. [A.3.2, Assumption A.13 and Theorem A.15] Assumption A.13 (κ_R = ||c||/||a|| < 1) is not a regularity condition but essentially the conclusion being proved: under the isotropy Assumption A.11, the inequality E[|c^T W|] < E[|a^T W|] used in Theorem A.15 reduces to ||c|| < ||a||. Neither assumption is checked on LLaMA-2 or LLaMA-3, so the smaller-perturbation claim is currently unsupported.
  3. [A.3.1, Eq. (9), Assumption A.7, and Algorithm 1] Assumption A.7 involves b = −∇θℓCE(z_old), which is not available during actual knowledge editing because the old training datum is unknown, and the proof of Theorem A.10 also assumes ϵ = 0, whereas Algorithm 1 uses a nonzero clipping threshold in all experiments (Appendix B.1). The paper provides no empirical diagnostics for Assumptions A.7 or A.11 on the actual models; these gaps should be filled with measured quantities or the formal-superiority claims should be explicitly reframed as heuristics.
  4. [Section 4.2, Table 1] The central empirical claim rests on single-run point estimates without standard errors, confidence intervals, or multiple seeds. For example, the LoRA portability jump on LLaMA-2 ZsRE is from 23.34 to 61.16 and the locality jump on WikiData counterfact is from 9.99 to 57.85; without variance information or a stated number of edits, it is difficult to know which of the many improvements are statistically meaningful.
minor comments (4)
  1. [Section 4.2, Section 5, Appendix A.3.2] There are several typos: 'filed' should be 'field', 'genergalizable' should be 'generalizable', and 'obatined' should be 'obtained'.
  2. [Lemma A.3] The statement contains a malformed second cross-entropy term, 'CE[πflt(y | c) | (y | c)]'; it should use the same ∥ notation as the rest of the lemma.
  3. [Theorem A.16 proof] In the displayed derivation, the term '− log πflt(y+ | ci) log πflt(y− | ci)' is missing a minus sign and should read '− log πflt(y+ | ci) − log πflt(y− | ci)'.
  4. [Appendix B.1] The paper says OVERTONE hyperparameters were tuned on a KE model base and applied to both LLMs; it would help to state which base and whether the T=1 continual results are the same runs as the single-editing results in Table 1.

Circularity Check

2 steps flagged · score 4.0 of 10

Empirical results are externally benchmarked and non-circular, but the formal locality guarantee (Theorem A.15) assumes the smaller-perturbation conclusion via κ_R<1, and Theorem A.10's existence-λ proof does not cover the deployed λ=0.1.

  1. other [Appendix A.3.2, Assumption A.13 and Theorem A.15 (proof of smaller perturbation on unrelated data)]
    "Assumption A.13. We assume that κ_R < 1. ... Theorem A.15. Let Zun = (Xun, Yun) be a random vector representing unrelated data. Under Assumptions A.11 and A.13, we have E_Zun[|π_θnew_OVERTONE (Zun) − π_θold(Zun)|] < E_Zun[|π_θnew_CE (Zun) − π_θold(Zun)|]."

    The proof of Theorem A.15 reduces the desired conclusion to E|a^T W| > E|c^T W|, then invokes Assumption A.11 (isotropy of W) to compute E|c^T W|/E|a^T W| = ||c||/||a|| = κ_R. Hence the theorem's conclusion is true exactly when κ_R < 1, which is stated as Assumption A.13. The 'regularity condition' is therefore not an independent premise; it is the smaller-perturbation result itself. The formal claim that OVERTONE 'has less influence on unrelated knowledge' thus reduces, by the paper's own equations, to assuming the norm inequality it purports to prove.

  2. other [Appendix A.3.1, Theorem A.10 vs. Algorithm 1 / Appendix B.1 (deployed λ=0.1)]
    "Take λ = ||c||_2 / (||b||_2 + ||c||_2) ... there exists some λ ∈ [0,1] such that cos(∇θℓCE(znew; θold), G) < cos(∇θℓOVERTONE (znew; θold), G). ... OVERTONE is tuned on a KE model base and applied to both LLMs: ... λ = 0.1 for mixing."

    Theorem A.10 proves only that some favorable λ (specifically ||c||/(||b||+||c||)) exists, while Algorithm 1 and all experiments use a fixed λ=0.1. The claimed theoretical advantage of the deployed update is therefore not established by the theorem; the favorable-update claim is not shown to apply to the method actually evaluated. This is a theory-to-practice mismatch rather than a definitional circularity, but it leaves the stated 'better update direction' support conditional on an unverified and unused construction.

full rationale

The paper's core empirical claim is not circular: OVERTONE is evaluated against external benchmarks (ZsRE, WikiData, WikiBio, MQuAKE) and external baselines (ROME, MEMIT, FT-M, LoRA, MELO, WISE), and the loss definition (Eq. 2) is not fitted to the portability/locality metrics. No load-bearing self-citation chain or imported uniqueness theorem is present; the top-nσ filtering is cited to Tang et al. (2024), an external source. The main circularity concern is in the formal theory: Theorem A.15's guarantee of smaller perturbation on unrelated data is proven by assuming κ_R = ||c||/||a|| < 1, which under the paper's own isotropy Assumption A.11 is exactly the claimed inequality. That is a conditional theorem whose premise is as strong as its conclusion, so the theoretical 'less influence on unrelated knowledge' claim reduces to an assumption. Theorem A.10 likewise proves existence of a favorable λ that is not the deployed λ=0.1, leaving the formal better-direction claim disconnected from the evaluated algorithm. These issues undermine the stated theoretical contributions but do not invalidate the independent empirical results, so the overall circularity score is moderate rather than severe.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The theoretical claims rest on four assumptions, two of which (A.7 and A.13) are unverified and one of which (A.13) nearly states the locality conclusion. The method also introduces three tuned hyperparameters for its loss. No new physical or architectural entities are introduced.

free parameters (3)
  • lambda (mixing weight) = 0.1 for all backbones
    Mixes the delta target with the filtered model distribution; hand-chosen and not covered by the theorem's existence proof.
  • epsilon (clipping threshold) = 0.01 for FT-M; 0.05 for LoRA, MELO, WISE
    Per-token early stopping threshold; tuned per backbone.
  • n-sigma (filtering cutoff) = 0.5 for FT-M and LoRA; 1.0 for MELO and WISE
    Top-n-sigma logit pruning threshold; tuned per backbone.
assumptions (5)
  • domain assumption Pretrained LLM is converged: gradient of pretraining loss is zero at theta_old (Assumption A.6).
    Used to define the ideal retraining gradient G; real LLMs are not exact stationary points.
  • ad hoc to paper The old-knowledge gradient b and the filtered-model gradient c point in nearly the same direction (Assumption A.7).
    Required for Theorem A.10 to prove a better update direction; not empirically verified.
  • ad hoc to paper For unrelated data, the normalized gradient direction is uniformly distributed on the unit sphere and independent of its norm (Assumption A.11).
    Used in Theorem A.15 for locality; isotropic gradient assumption is convenient but not grounded in LLM behavior.
  • ad hoc to paper The filtered gradient norm is smaller than the ground-truth gradient norm, kappa_R < 1 (Assumption A.13).
    Directly used to conclude that OVERTONE perturbs unrelated data less; this nearly restates the locality conclusion it supports.
  • domain assumption The model's own prediction, after top-n-sigma filtering, is a reliable prior for preserving pretrained knowledge.
    Invoked in Section 3.1 to justify mixing pi_flt with the delta target; relies on Tang et al. (2024) and on self-distillation intuitions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mitigating Heterogeneous Token Overfitting in LLM Knowledge Editing." pith.science (2026). https://pith.science/paper/QMV6DO47

@misc{pith2026250200602,
  author       = {Pith},
  title        = {Pith review of: Mitigating Heterogeneous Token Overfitting in LLM Knowledge Editing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QMV6DO47}},
  note         = {Machine review of arXiv:2502.00602}
}
read the original abstract

Large language models (LLMs) have achieved remarkable performance on various natural language tasks. However, they are trained on static corpora and their knowledge can become outdated quickly in the fast-changing world. This motivates the development of knowledge editing (KE) to update specific knowledge in LLMs without changing unrelated others or compromising their pre-trained capabilities. Previous efforts sought to update a small amount of parameters of a LLM and proved effective for making selective updates. Nonetheless, the edited LLM often exhibits degraded ability to reason about the new knowledge. In this work, we identify a key issue: heterogeneous token overfitting (HTO), where the LLM overfits different tokens in the provided knowledge at varying rates. To tackle this, we propose OVERTONE, a token-level smoothing method that mitigates HTO by adaptively refining the target distribution. Theoretically, OVERTONE offers better parameter updates with negligible computation overhead. It also induces an implicit DPO but does not require preference data pairs. Extensive experiments across four editing methods, two LLMs, and diverse scenarios demonstrate the effectiveness and versatility of our method.

Figures

Figures reproduced from arXiv: 2502.00602 by the authors.

Figure 2
Figure 2. Token-level initial loss and UD (negative indicates over￾fitted). Dashed lines mark the mean values. Towards a deeper understanding of this overfitting phe￾nomenon, we check the loss of each token, and find that different tokens tend to have distinct initial loss values. As depicted in Fig 2a, before editing LLaMA2, only certain tokens (e.g., the beginning) have significant loss values. On the other hand, some token… view at source ↗
Figure 1
Figure 1. Loss (average) change of ground truth answers to gener￾ality (rephrased, left) and portability (reasoning, right) questions. As a direct evidence, [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 3
Figure 3. Continual Editing performance under different sequence length T. Solid and transparent bars show performance with and without OVERTONE. Unfilled area marks the performance gap. ROME and MEMIT didn’t use OVERTONE. 5. Related Works Existing KE methods mainly fall into two classes. Internal Storage updates model parameters for the adapta￾tion. Early studies fine-tuned a LLM directly but suffered from severe forgetting … view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. ScienceMeter: Tracking Scientific Knowledge Updates in Language Models

    cs.CL 2025-05 reject novelty 6.0 of 10

    ScienceMeter evaluates language model knowledge updates across three axes, preservation of old scientific claims, acquisition of new claims, and projection to future findings, and finds all current methods fall short.

Reference graph

Works this paper leans on

64 extracted references · 11 canonical work pages · cited by 1 Pith paper

  1. [1]

    A neural probabilistic language model

    Yoshua Bengio, R \'e jean Ducharme, and Pascal Vincent. A neural probabilistic language model. Advances in neural information processing systems, 13, 2000

    2000

  2. [2]

    Pattern recognition and machine learning, volume 4

    Christopher M Bishop and Nasser M Nasrabadi. Pattern recognition and machine learning, volume 4. Springer, 2006

    2006

  3. [3]

    On the opportunities and risks of foundation models

    Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021

    arXiv 2021

  4. [4]

    Language models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33: 0 1877--1901, 2020

    1901

  5. [5]

    Sparks of artificial general intelligence: Early experiments with gpt-4

    S \'e bastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, et al. Sparks of artificial general intelligence: Early experiments with gpt-4. arXiv preprint arXiv:2303.12712, 2023

    arXiv 2023

  6. [6]

    Learning phrase representations using rnn encoder-decoder for statistical machine translation

    Kyunghyun Cho, Bart Van Merri \"e nboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078, 2014

    arXiv 2014

  7. [7]

    Evaluating the ripple effects of knowledge editing in language models

    Roi Cohen, Eden Biran, Ori Yoran, Amir Globerson, and Mor Geva. Evaluating the ripple effects of knowledge editing in language models. Transactions of the Association for Computational Linguistics, 12: 0 283--298, 2024

    work page 2024

  8. [8]

    Elements of information theory

    Thomas M Cover. Elements of information theory. John Wiley & Sons, 1999

    1999

Show all 64 references
  1. [9]

    Knowledge neurons in pretrained transformers

    Damai Dai, Li Dong, Yaru Hao, Zhifang Sui, Baobao Chang, and Furu Wei. Knowledge neurons in pretrained transformers. arXiv preprint arXiv:2104.08696, 2021

    2021 arXiv

  2. [10]

    Editing factual knowledge in language models

    Nicola De Cao, Wilker Aziz, and Ivan Titov. Editing factual knowledge in language models. arXiv preprint arXiv:2104.08164, 2021

    2021 arXiv

  3. [11]

    Calibrating factual knowledge in pretrained language models

    Qingxiu Dong, Damai Dai, Yifan Song, Jingjing Xu, Zhifang Sui, and Lei Li. Calibrating factual knowledge in pretrained language models. arXiv preprint arXiv:2210.03329, 2022 a

    2022 arXiv

  4. [12]

    A survey for in-context learning

    Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Zhiyong Wu, Baobao Chang, Xu Sun, Jingjing Xu, and Zhifang Sui. A survey for in-context learning. arXiv preprint arXiv:2301.00234, 2022 b

    2022 arXiv

  5. [13]

    The llama 3 herd of models, 2024

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models, 2024. URL https://arxiv.org/abs/2407.21783

    2024 arXiv

  6. [14]

    Toxigen: A large-scale machine-generated dataset for adversarial and implicit hate speech detection

    Thomas Hartvigsen, Saadia Gabriel, Hamid Palangi, Maarten Sap, Dipankar Ray, and Ece Kamar. Toxigen: A large-scale machine-generated dataset for adversarial and implicit hate speech detection. arXiv preprint arXiv:2203.09509, 2022

    2022 arXiv

  7. [15]

    Aging with grace: Lifelong model editing with discrete key-value adaptors

    Tom Hartvigsen, Swami Sankaranarayanan, Hamid Palangi, Yoon Kim, and Marzyeh Ghassemi. Aging with grace: Lifelong model editing with discrete key-value adaptors. Advances in Neural Information Processing Systems, 36, 2024

    2024

  8. [16]

    Truncation sampling as language model desmoothing

    John Hewitt, Christopher D Manning, and Percy Liang. Truncation sampling as language model desmoothing. arXiv preprint arXiv:2210.15191, 2022

    2022 arXiv

  9. [17]

    Long short-term memory

    S Hochreiter. Long short-term memory. Neural Computation MIT-Press, 1997

    1997

  10. [18]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models, 2021

    2021

  11. [19]

    Transformer-patcher: One mistake worth one neuron

    Zeyu Huang, Yikang Shen, Xiaofeng Zhang, Jie Zhou, Wenge Rong, and Zhang Xiong. Transformer-patcher: One mistake worth one neuron. arXiv preprint arXiv:2301.09785, 2023

    2023 arXiv

  12. [20]

    Survey of hallucination in natural language generation

    Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. Survey of hallucination in natural language generation. ACM Computing Surveys, 55 0 (12): 0 1--38, 2023

    2023

  13. [21]

    Learning to edit: Aligning llms with knowledge editing

    Yuxin Jiang, Yufei Wang, Chuhan Wu, Wanjun Zhong, Xingshan Zeng, Jiahui Gao, Liangyou Li, Xin Jiang, Lifeng Shang, Ruiming Tang, et al. Learning to edit: Aligning llms with knowledge editing. arXiv preprint arXiv:2402.11905, 2024

    2024 arXiv

  14. [22]

    Understanding black-box predictions via influence functions

    Pang Wei Koh and Percy Liang. Understanding black-box predictions via influence functions. In International conference on machine learning, pp.\ 1885--1894. PMLR, 2017

    2017

  15. [23]

    Large language models are zero-shot reasoners

    Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35: 0 22199--22213, 2022

    2022

  16. [24]

    Professor forcing: A new algorithm for training recurrent networks, 2016

    Alex Lamb, Anirudh Goyal, Ying Zhang, Saizheng Zhang, Aaron Courville, and Yoshua Bengio. Professor forcing: A new algorithm for training recurrent networks, 2016. URL https://arxiv.org/abs/1610.09038

    2016 arXiv

  17. [25]

    Adaptive label smoothing with self-knowledge in natural language generation

    Dongkyu Lee, Ka Chun Cheung, and Nevin L Zhang. Adaptive label smoothing with self-knowledge in natural language generation. arXiv preprint arXiv:2210.13459, 2022

    2022 arXiv

  18. [26]

    Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension

    Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. arXiv preprint arXiv:1910.13461, 2019

    1910 arXiv

  19. [27]

    Locating and editing factual associations in gpt

    Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Locating and editing factual associations in gpt. Advances in Neural Information Processing Systems, 35: 0 17359--17372, 2022 a

    2022

  20. [28]

    Mass-editing memory in a transformer

    Kevin Meng, Arnab Sen Sharma, Alex Andonian, Yonatan Belinkov, and David Bau. Mass-editing memory in a transformer. arXiv preprint arXiv:2210.07229, 2022 b

    2022 arXiv

  21. [29]

    Fast model editing at scale

    Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D Manning. Fast model editing at scale. arXiv preprint arXiv:2110.11309, 2021

    2021 arXiv

  22. [30]

    Memory-based model editing at scale

    Eric Mitchell, Charles Lin, Antoine Bosselut, Christopher D Manning, and Chelsea Finn. Memory-based model editing at scale. In International Conference on Machine Learning, pp.\ 15817--15831, 2022

    2022

  23. [31]

    When does label smoothing help? Advances in neural information processing systems, 32, 2019

    Rafael M \"u ller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? Advances in neural information processing systems, 32, 2019

    2019

  24. [32]

    Training language models to follow instructions with human feedback

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35: 0 2...

    2022

  25. [33]

    Language models are unsupervised multitask learners

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1 0 (8): 0 9, 2019

    2019

  26. [34]

    Direct preference optimization: Your language model is secretly a reward model

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36, 2024

    2024

  27. [35]

    Exploring the limits of transfer learning with a unified text-to-text transformer

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21 0 (1): 0 5485--5551, 2020

    2020

  28. [36]

    Knowledge editing in language models via adapted direct preference optimization

    Amit Rozner, Barak Battash, Lior Wolf, and Ofir Lindenbaum. Knowledge editing in language models via adapted direct preference optimization. arXiv preprint arXiv:2406.09920, 2024

    2024 arXiv

  29. [37]

    Do massively pretrained language models make better storytellers? arXiv preprint arXiv:1909.10705, 2019

    Abigail See, Aneesh Pappu, Rohun Saxena, Akhila Yerukola, and Christopher D Manning. Do massively pretrained language models make better storytellers? arXiv preprint arXiv:1909.10705, 2019

    1909 arXiv

  30. [38]

    Sequence to sequence learning with neural networks

    I Sutskever. Sequence to sequence learning with neural networks. arXiv preprint arXiv:1409.3215, 2014

    2014 arXiv

  31. [39]

    Rethinking the inception architecture for computer vision

    Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 2818--2826, 2016

    2016

  32. [40]

    Top- n : Not all logits are you need

    Chenxia Tang, Jianchun Liu, Hongli Xu, and Liusheng Huang. Top- n : Not all logits are you need. arXiv preprint arXiv:2411.07641, 2024

    2024 arXiv

  33. [41]

    Llama 2: Open foundation and fine-tuned chat models

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023

    2023 arXiv

  34. [42]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017

    2017

  35. [43]

    Beyond reverse KL : Generalizing direct preference optimization with diverse divergence constraints

    Chaoqi Wang, Yibo Jiang, Chenghao Yang, Han Liu, and Yuxin Chen. Beyond reverse KL : Generalizing direct preference optimization with diverse divergence constraints. In The Twelfth International Conference on Learning Representations, 2024 a

    2024

  36. [44]

    Making large language models better reasoners with alignment

    Peiyi Wang, Lei Li, Liang Chen, Feifan Song, Binghuai Lin, Yunbo Cao, Tianyu Liu, and Zhifang Sui. Making large language models better reasoners with alignment. arXiv preprint arXiv:2309.02144, 2023 a

    2023 arXiv

  37. [45]

    Wise: Rethinking the knowledge memory for lifelong model editing of large language models

    Peng Wang, Zexi Li, Ningyu Zhang, Ziwen Xu, Yunzhi Yao, Yong Jiang, Pengjun Xie, Fei Huang, and Huajun Chen. Wise: Rethinking the knowledge memory for lifelong model editing of large language models. arXiv preprint arXiv:2405.14768, 2024 b

    2024 arXiv

  38. [46]

    Easyedit: An easy-to-use knowledge editing framework for large language models, 2024 c

    Peng Wang, Ningyu Zhang, Bozhong Tian, Zekun Xi, Yunzhi Yao, Ziwen Xu, Mengru Wang, Shengyu Mao, Xiaohan Wang, Siyuan Cheng, Kangwei Liu, Yuansheng Ni, Guozhou Zheng, and Huajun Chen. Easyedit: An easy-to-use knowledge editing framework for large language models, 2024 c . URL ...

    2024 arXiv

  39. [47]

    Knowledge editing for large language models: A survey

    Song Wang, Yaochen Zhu, Haochen Liu, Zaiyi Zheng, Chen Chen, et al. Knowledge editing for large language models: A survey. arXiv preprint arXiv:2310.16218, 2023 b

    2023 arXiv

  40. [48]

    Deepedit: Knowledge editing as decoding with constraints

    Yiwei Wang, Muhao Chen, Nanyun Peng, and Kai-Wei Chang. Deepedit: Knowledge editing as decoding with constraints. arXiv preprint arXiv:2401.10471, 2024 d

    2024 arXiv

  41. [49]

    Chain-of-thought prompting elicits reasoning in large language models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35: 0 24824--24837, 2022

    2022

  42. [50]

    Stable knowledge editing in large language models

    Zihao Wei, Liang Pang, Hanxing Ding, Jingcheng Deng, Huawei Shen, and Xueqi Cheng. Stable knowledge editing in large language models. arXiv preprint arXiv:2402.13048, 2024

    2024 arXiv

  43. [51]

    Eva-kellm: A new benchmark for evaluating knowledge editing of llms

    Suhang Wu, Minlong Peng, Yue Chen, Jinsong Su, and Mingming Sun. Eva-kellm: A new benchmark for evaluating knowledge editing of llms. arXiv preprint arXiv:2308.09954, 2023

    2023 arXiv

  44. [52]

    On early stopping in gradient descent learning

    Yuan Yao, Lorenzo Rosasco, and Andrea Caponnetto. On early stopping in gradient descent learning. Constructive Approximation, 26 0 (2): 0 289--315, 2007

    2007

  45. [53]

    Editing large language models: Problems, methods, and opportunities

    Yunzhi Yao, Peng Wang, Bozhong Tian, Siyuan Cheng, Zhoubo Li, Shumin Deng, Huajun Chen, and Ningyu Zhang. Editing large language models: Problems, methods, and opportunities. arXiv preprint arXiv:2305.13172, 2023

    2023 arXiv

  46. [54]

    Melo: Enhancing model editing with neuron-indexed dynamic lora

    Lang Yu, Qin Chen, Jie Zhou, and Liang He. Melo: Enhancing model editing with neuron-indexed dynamic lora. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pp.\ 19449--19457, 2024

    2024

  47. [55]

    Uncovering overfitting in large language model editing

    Mengqi Zhang, Xiaotian Ye, Qiang Liu, Pengjie Ren, Shu Wu, and Zhumin Chen. Uncovering overfitting in large language model editing. arXiv preprint arXiv:2410.07819, 2024 a

    2024 arXiv

  48. [56]

    Instructedit: Instruction-based knowledge editing for large language models

    Ningyu Zhang, Bozhong Tian, Siyuan Cheng, Xiaozhuan Liang, Yi Hu, Kouying Xue, Yanjie Gou, Xi Chen, and Huajun Chen. Instructedit: Instruction-based knowledge editing for large language models. arXiv preprint arXiv:2402.16123, 2024 b

    2024 arXiv

  49. [57]

    A comprehensive study of knowledge editing for large language models

    Ningyu Zhang, Yunzhi Yao, Bozhong Tian, Peng Wang, Shumin Deng, Mengru Wang, Zekun Xi, Shengyu Mao, Jintian Zhang, Yuansheng Ni, et al. A comprehensive study of knowledge editing for large language models. arXiv preprint arXiv:2401.01286, 2024 c

    2024 arXiv

  50. [58]

    Negative preference optimization: From catastrophic collapse to effective unlearning

    Ruiqi Zhang, Licong Lin, Yu Bai, and Song Mei. Negative preference optimization: From catastrophic collapse to effective unlearning. arXiv preprint arXiv:2404.05868, 2024 d

    2024 arXiv

  51. [59]

    Self-distillation as instance-specific label smoothing

    Zhilu Zhang and Mert Sabuncu. Self-distillation as instance-specific label smoothing. Advances in Neural Information Processing Systems, 33: 0 2184--2195, 2020

    2020

  52. [60]

    Can we edit factual knowledge by in-context learning? arXiv preprint arXiv:2305.12740, 2023

    Ce Zheng, Lei Li, Qingxiu Dong, Yuxuan Fan, Zhiyong Wu, Jingjing Xu, and Baobao Chang. Can we edit factual knowledge by in-context learning? arXiv preprint arXiv:2305.12740, 2023

    2023 arXiv

  53. [61]

    Mquake: Assessing knowledge editing in language models via multi-hop questions

    Zexuan Zhong, Zhengxuan Wu, Christopher D Manning, Christopher Potts, and Danqi Chen. Mquake: Assessing knowledge editing in language models via multi-hop questions. arXiv preprint arXiv:2305.14795, 2023

    2023 arXiv

  54. [62]

    A comprehensive survey on pretrained foundation models: A history from bert to chatgpt

    Ce Zhou, Qian Li, Chen Li, Jun Yu, Yixin Liu, Guangjing Wang, Kai Zhang, Cheng Ji, Qiben Yan, Lifang He, et al. A comprehensive survey on pretrained foundation models: A history from bert to chatgpt. arXiv preprint arXiv:2302.09419, 2023

    2023 arXiv

  55. [63]

    Modifying memories in transformer models

    Chen Zhu, Ankit Singh Rawat, Manzil Zaheer, Srinadh Bhojanapalli, Daliang Li, Felix Yu, and Sanjiv Kumar. Modifying memories in transformer models. arXiv preprint arXiv:2012.00363, 2020

    2012 arXiv

  56. [64]

    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 9, 2026 · model on record in the stance chip above.