Massive Editing for Large Language Models via Meta Learning

Summary

This paper addresses the challenge of correcting and updating the knowledge of large language models (LLMs) that have been pre-trained on extensive text data. It introduces a novel approach called the MAssive Language Model Editing Network (MALMEN). MALMEN formulates the parameter shift aggregation as a least square problem, seeking the most effective parameter shifts for all facts to be injected. This approach improves the statistical significance of the editing process, mitigating the issues of gradient accumulation and the cancellation effect. Furthermore, this paper separates the computation between the hyper-network and the language model, enabling the use of arbitrary batch sizes for both neural networks. Exceptional performance on multiple knowledge-intensive tasks is a testament to MALMEN's effectiveness.

Strengths

1.This is overall a well-written paper: it tackles a very important problem, formulating the parameter shift aggregation as the least square problem. This approach differs from traditional fine-tuning and overcomes the challenges of scalability and memory usage in existing hyper-network methods. 2.This paper focuses on scalability. MALMEN is designed to edit multiple facts simultaneously, making it more practical for mass editing in real-world scenarios. This is a crucial aspect given the need to update the knowledge of large language models comprehensively. 3.Despite being somewhat math-heavy, the paper is written in a very clear and didactic way. I found it easy to follow and an enjoyable read overall. 4.Comprehensive (although not entirely convincing, see below) experiments on various knowledge-intensive NLP tasks and across different LLM architectures. This demonstrates the effectiveness and versatility of the proposed method.

Weaknesses

1.Baselines are limited. Why not compare with T-Patcher (Huang et al., 2023), which I believe is more suitable for sequential knowledge editing? 2.The criteria for successful edits are, in my opinion, insufficient, in that they do not consider the portability of the edit. Previous work such as Yao et al., 2023, introducing an additional assessment metric, portability, finding that the model-editing methods lack robustness when applied to related one-hop fact or synonyms. 3.Which layers to apply MALMEN? All layers or some picked layers? Section 5.2 claims that “Edit first FC in FFN” achieves inferior performance. How to select the layer in practical application? 4.The experiments are lacking in qualitative examples, it would be helpful to analyze some success and failure cases to see where the proposed method begins to fail (e.g., with respect to generalization). 5.MALMEN is essentially a combination of meta-learning based (e.g. MEND, Mitchell 2022) and parametric (e.g. ROME, Meng 2022 or MEMIT, Meng 2023) editing ideas that shows some promise. The method is not particularly technically novel (minor point).

Questions

In Figure 5 and Figure 6, the starting x coordinates are inconsistent, please provide a more detailed description. Is it fair to assume that MALMEN is less effective with fewer edits?

Rating

6: marginally above the acceptance threshold

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Summary

This paper proposes MALMEN, a massive editing method for LLM, which employs the least square method to aggregate parameter shifts inspired from MEMIT, and then applies the parameter updating method by taking the least squared solution as increment of the parameter metric, for minimizing the meta loss. To efficiently design the back propagation for massive editing, the paper separates the backprop on LM and hyper-network such that the back props are proceeded in a cascaded way, maintaining a set of cache values. Experiment results on FEVER and zsRE dataset show that the proposed MALMEN improves MEND on BERT-based and GPT-2, and often improves MEMIT on GPT-J, under some types of edits.

Strengths

- The proposed combination of the least square method and the loss-based updating for massive editing is quite interesting and novel. - The truncated backprop algorithm is solidly designed to improve the efficiency, which is also quite interesting. - The experiment results show that the proposed method improves MEND or MEMIT under various settings.,

Weaknesses

- Instead of the least squared solution, the simple sum-based aggregation is not compared. To prove the effect of the proposed method, this simplified aggregation needs to be compared. - The description of Section 4.2 is largely dense, too hard to capture the details. In particular, Figure 2 provides the overall backprop flow, but why the training algorithm using the truncated backprop is not explicitly and clearly provided? - In GPT-J (6B), the proposed method doesn’t improve MEMIT, in terms of LS metric. This result needs to be properly discussed.

Questions

In Section 4.2, some derivations are not very clear. 1) how the following is derived? Delta_D L_meta = Delta_W L_Meta * (U_l U_l^T + lambda_L I)^-1 U_l Other remaining formulas need more explanation on how they are derived. 2) What does mean the method of “Cache all tokens”?

Rating

6: marginally above the acceptance threshold

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Summary

This paper proposes an improvement to MEND for large-scale fact editing. Similar to MEND, MALMEN uses a hypernetwork that takes in gradients (with respect to some input/output tuples) and hidden states and outputs a parameter update. The general training objective is similar to MEND, and the primary improvement proposed is a better method for combining multiple "fact" updates as opposed to naively summing/accumulating over single updates. They evaluate on standard memory editing tasks (based on FEVER), on BERT-base, GPT-2 and GPT-J

Strengths

- The paper provides plenty of technical details, and is fairly clear (though somewhat dense) - The method is straightforward and intuitive. I am unclear about the broader applicability of memory editing, but the technical details and performance are sufficiently convincing to me that this is a meaningful contribution.

Weaknesses

- The paper requires quite a bit of background on MEND. This is not inherently a bad thing since the paper is basically a direct modification of MEND, and the paper already spends a good deal of space building the background, but I think providing higher-level intuition in the exposition could help. - Section 4.2 wasn't very clear to me (in particular "truncating the back-propagation at the end of linear layers"). Figure 2 was significantly clearer, and I wonder if the authors could revisit the section and tweak it for ease of understanding the somewhat complicated procedure for training. - The results on scaling to GPT-J seem a little unstable

Questions

- Can you clarify "truncating the back-propagation at the end of linear layers"? - The line "Edit first FC in FFN” turns to edit the first linear layer in the FFN of the last 6 Transformer blocks" is unclear to me. How does the non-ablated MALMEN differ?

Rating

10: strong accept, should be highlighted at the conference

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

4 excellent

Summary

This paper considers a problem of knowledge editing, which involves altering the parametric knowledge of LMs without retraining them from scratch. This work specifically focuses on the scalability of hypernetwork-based approaches, which are generally considered less effective for multiple concurrent edits. The authors claim that there are two major challenges: 1) the parameter shifts could be contradictory between the set of modified facts, and 2) accommodating a large number of edits in a hypernetwork is memory demanding. This work presents an approach that addresses these challenges. Concretely, this work extends MEND (Mitchell et al., 2022) by introducing additional parameter updates specifically for linear layers in the FFNs. Assuming that the linear layers are key-value memories, the motivation behind this is to find a better single parameter shift matrix S for _m_ updates. This additional step adjust the hypernetwork output (i.e., gradients) which is not a simple sum of gradients for different inputs. When scaling up to a large number of edits, backpropagating from the meta loss to the input is costly (e.g., computing pre and post-edit losses for each edit end to end). The proposed approach decomposes the optimization process by caching pre-edit computation (after finetuning), reducing the memory usage substantially. The experimental setup focuses on scalability (i.e., editing thousands of facts at once), and the proposed approach is applied to different model families such as encoder-only (e.g., BERT) and decoder-only (e.g., GPT-2 and GPT-J 6B). In addition to FT and MEND baselines, GPT-J with MEMIT is included as a baseline. For evaluation, FEVER is used for BERT, and zsRE is used for GPT models, largely following prior work. For evaluation metrics, edit success (ES – how often new facts get higher probability after editing), generalization success (GS – performance on related facts), and locality success (LS – performance of unrelated facts). In summary, the experimental results show that the proposed approach consistently outperforms FT and MEND with BERT and GPT-2, and it has better scalability compared to the original MEND. When it comes with GPT-J, which is a much larger LM, it is always better than MEND but underperforms MEMIT and FT on LS, indicating that the post-edit model forgets unrelated facts.

Strengths

- This work is tackling a well-motivated problem, scaling up knowledge editing approaches. - The motivation behind the proposed approach (adjusting FFN weights, decomposing the optimization process) is clearly explained, and the solutions presented are reasonable.

Weaknesses

- The scope of the problem (scalability of MEND) could be narrow, and the proposed approach is only applicable for a specific knowledge editing approach. - Based on the experimental results, it is difficult to assert that this approach is significantly better than all other knowledge editing approaches in terms of scalability (not only MEND). - The poor LS score with GPT-J (6B) shows that this approach still edits unrelated facts. - Qualitative analysis is not provided. It’s hard to see when/why this approach is beneficial without seeing error cases.

Questions

- Section 4.1: The clarity of the notations could be improved, especially the parameter shift matrix S and the different matrix D. It’s unclear which parameters are trainable/frozen from the notations. And, it’s hard to see how those operations are applied to _m_ edits. - “in the case of parameter shifts generated by the hyper-network, summing them lacks statistical significance”: This sounds intuitive, but is there any theoretical or empirical research that substantiates this? Yeh et al., (2022) is mainly talking about the cancellation effect in the last layer of a transformer if I understand it correctly. - Did you use the original implementation of MEND? If not, it would be nice to show that the results match with your implementation. - It would be nice to explain data statistics briefly.

Rating

5: marginally below the acceptance threshold

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair