Uncovering Overfitting in Large Language Model Editing

Thank you for your thoughtful review and for acknowledging the contributions of our work. We hope the following points will clarify and resolve your concerns.

(W1) I believe Section 5 may not be essential to the overall paper. While it contributes some useful experimental insights with basic mitigation techniques, it does not appear closely aligned with the paper's primary contributions. Instead, it occupies space that could be better devoted to the LTI section, which I consider more significant.

Thank you for your valuable suggestions! The analysis in Section 5 serves as a deeper exploration of phenomena observed in our primary experiment, and the limitation of general overfitting mitigation techniques against Editing Overfit also underscores the challenges of this issue. While we believe demonstrating the impact of current mitigation strategies on Editing Overfit remains a key contribution, we will consider refining the paper by adjusting the length of Section 5 and subsequent sections, and enhancing the LTI section to better align with the paper's core focus!

(W2) I understand that this work focuses on the overfitting issue in model editing. However, to my knowledge, model editing itself performs poorly on multi-hop editing tasks, likely due in large part to the overfitting phenomenon discussed. I suggest that the authors consider addressing this perspective in the paper.

Thank you for pointing out this perspective! We have further emphasized that the suboptimal performance of edited models in multi-hop editing tasks may be largely due to the phenomenon of Editing Overfit in the revised paper.

(W3) I strongly recommend that the authors include a description of in-context editing approaches in relevant sections, such as Related Work or the Baseline. Compared to model editing, these methods are potentially better suited for handling multi-hop editing tasks. Additionally, relevant citations are missing, such as [1][2][3].

Thank you for your suggestion! Our work specifically investigates the overfitting issues inherent in parameter-modifying methods and proposes LTI as a targeted mitigation strategy for these approaches. As such, our baselines primarily focus on parameter-modifying methods. In the revised version, we have added a Detailed Related Work section (Appendix B) to provide an overview of in-context editing approaches and their applicability to multi-hop tasks.

[1] Bi B, Liu S, Mei L, et al. Decoding by Contrasting Knowledge: Enhancing LLMs' Confidence on Edited Facts[J]. arXiv preprint arXiv:2405.11613, 2024.

[2] Zhong Z, Wu Z, Manning C D, et al. Mquake: Assessing knowledge editing in language models via multi-hop questions[J]. arXiv preprint arXiv:2305.14795, 2023.

[3] Bi B, Liu S, Wang Y, et al. Struedit: Structured outputs enable the fast and accurate knowledge editing for large language models[J]. arXiv preprint arXiv:2409.10132, 2024.

We highly appreciate your valuable insights and acknowledgment of our contributions. We hope the following comments could address your concerns.

(W1): It seems that LTI is only applicable in cases where the edited knowledge can be explicitly represented as knowledge triples $(s, r, o, o^{*})$. However, more complex or nuanced knowledge editing tasks may not fit into this structured format, potentially limiting LTI's applicability in real-world scenarios.

Thank you for raising this insightful point! We acknowledge that our approach implemented via Multi-stage Constraint, is best suited for edits that can be explicitly represented as knowledge triples (s, r, o, o*), which also the key focus in current research. We agree that enabling models to "Learn To Inference" in more general knowledge contexts is an intriguing topic for future research, and we've added a dedicated Limitations section in Appendix A of the revised paper to explicitly discuss this limitation.

(W2 and Q1): The phenomenon of Editing Overfit and the proposed LTI approach are only tested on small-scale LLMs such as GPT-J and GPT-2. Extending the analysis to larger models, such as LLaMA-2/3 or other larger models, would provide a more comprehensive understanding of LTI’s effectiveness at scale.

Thank you for your valuable comment! To address the concern, we have extended our experiments to evaluate the performance of baseline methods (FT, FT-L, ROME and MEMIT) and LTI on the Llama-2-7B using our EVOKE benchmark (Details results are included in Appendix G), and some results are as follows:

From the results, we observe the following conclusions, which are consistent with our findings on GPT-J and GPT 2-XL models:

• The baseline methods still demonstrate a pronounced Editing Overfit phenomenon on our EVOKE benchmark. Nearly all successfully edited models exhibit significantly higher Direct Probability (DP) scores compared to the unedited model.

• Both ROME-LTI and MEMIT-LTI demonstrate significant improvements in overfitting metrics (DP and EOS) compared to ROME and MEMIT, indicating effective mitigation of editing overfit.

(Q2): Are there specific types of facts or domains where the problem of Editing Overfit is more severe? Is there any specific types of knowledge where LTI may be less effective?

We appreciate the reviewer’s insightful comments! The central point of our paper is that the Editing Overfit phenomenon likely stems from existing knowledge editing paradigms, which emphasize the direct correspondence between the input prompt $p(s,r)$ and the output $o^*$ for each edit sample $(s,r,o,o^{*})$. Given that most existing editing methods do not differentiate or focus on knowledge types or domains, and primarily establish input-output mappings, we believe that the Editing Overfit phenomenon likely appears independent of specific knowledge types, while we acknowledge that the extent of overfit could be influenced by knowledge types. We also believe that exploring the performance of various editing models across different types of knowledge domains represents a promising topic for future research.

We sincerely appreciate your recognition of our work and thank you for your valuable feedback. We hope the following responses will effectively address your concerns.

(W1): The proposed LTI method relies on the model's in-context learning capability. This approach requires the unedited model to generate correct answers based solely on the context of the new knowledge, thereby providing an accurate representation and output distribution constraint for the editing process. This implies that for smaller models less proficient in in-context learning, or for models particularly rigid regarding certain facts, the constraints provided by the LTI method may negatively impact editing performance.

Thank you for your insightful comment regarding the reliance of the proposed LTI method on the model’s in-context learning capability. We agree that this dependence could pose limitations for smaller models less proficient in in-context learning or for models rigid in their handling of certain facts. While acknowledging this dependency, we believe the continued advancements in LLMs and their growing capabilities make approaches like ours, which leverage these inherent strengths, increasingly promising. We have added a dedicated Limitations section in Appendix A of the revised version. This section thoroughly discusses these limitations and their implications.

(Q1) I am curious whether the representation and output distribution of the unedited model are computed only once. If so, I wonder whether applying Paraphrase Augmentation to this new knowledge would improve performance. The paper mentions that Paraphrase Augmentation can actually exacerbate overfitting in knowledge editing methods. Therefore, would applying Paraphrase Augmentation to the new knowledge enhance the quality of the constraint, or would it ultimately reduce the effectiveness of the edit?

Yes, the representation and output distribution of unedited model are computed only once. To further answer your question, we apply Paraphrase Augmentation to our LTI strategy (MEMIT-LTI w/ Para). The results are as follows:

From the results, we can observe that MEMIT-LTI w/ Para demonstrates a notable reduction in overfitting metrics compared to MEMIT-LTI, indicating that Paraphrase Augmentation can limit LTI's effective in addressing the Editing Overfit. Nevertheless, the overall overfitting metrics remain superior to those of MEMTI w/ Para, further validating the effectiveness of our LTI in mitigating overfiting. We also acknowledge that while Paraphrase Augmentation exacerbates overfitting to some extent, it enhances the correctness of paraphrase-form responses to edited prompts, presenting a trade-off in practical applications. Its usage should therefore be evaluated on a case-by-case basis.

Thank you for your recognition and interest in our work!

We agree that the dataset factors you mentioned could indeed be a possible reason for the slight decrease in paraphrase performance observed in Figure 7. The paraphrase task setting in EVOKE, similar to most existing datasets, follows the COUNTERFACT dataset, where the test questions primarily differ from the original Edit Prompt in terms of sentence structure (expression of relations), while maintaining consistency in the subject form with the original Edit Prompt.

We also believe that a crucial factor is that the strong paraphrase metrics demonstrated by many baseline models (including ROME and MEMIT) might be partly attributed to overfitting. For instance, COUNTERFACT's paraphrase tests primarily modify the form of relations, while keeping the subject form unchanged. Since the subject form remains unchanged, models that overfit to the s → o^* pattern might benefit from this setting: they can easily identify the subject pattern in the question and output $o^{*}$ accordingly. This perfectly satisfies the requirements of the paraphrase task. Therefore, when suppressing model overfitting, we may simultaneously reduce the contribution of overfitting to paraphrase metrics. This could be a significant reason for the results shown in Figure 7, where we observe a slight decrease in paraphrase metrics when applying data augmentation and LTI.

Furthermore, we agree with your perspective that current paraphrase evaluations might not be comprehensive enough. Although this may slightly exceed the scope of this paper, some previous works [1] have introduced evaluation tasks that replace the subject $s$ with its aliases, and they found that most editing methods do not perform well in such scenarios. This approach of transforming $s$ contrasts with the relation-focused paraphrasing in most existing works, which we believe indicates that real-world paraphrase tasks might have more complex manifestations, just as how the overfitting studied in this paper might manifest in various complex situations. This could be an interesting topic for future research.

Thank you again for recognizing our work and we welcome further discussions!

[1] Yunzhi Yao, Peng Wang, Bozhong Tian, Siyuan Cheng, Zhoubo Li, Shumin Deng, Huajun Chen, and Ningyu Zhang. 2023. Editing Large Language Models: Problems, Methods, and Opportunities. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 10222–10240, Singapore. Association for Computational Linguistics.