Locate-then-edit for Multi-hop Factual Recall under Knowledge Editing

Thank you very much for your recognition of our work. Here is our response.

Response to Q1.a in Questions For Authors

In the past exploration process, we also used patchscope to perform the experimental effects shown in figure 2, and judged whether the corresponding information was contained by counting whether the subsequent output of patchscope contained the answer. During the experiment, the results were surprisingly consistent with logitlens, so considering the universality of the logitlens method and cost-effectiveness, we still use it as our interpretability tool. We believe that the results obtained by logitlens are already trustworthy, but in the future, we can continue to explore further through methods such as tuned-lens.

Response to Q1.b in Questions For Authors

First,regarding the optimization of s2 related information, we did not use SGD optimization, but replaced the logits of the token in s2 with the minimum value on the whole vocabulary. Here it should be expressed as $s_l^*[j]$ = min($s_l$) in equations 1, where $s_l$ represents the all tokens in vocabulary, and then a simple mathematical calculation is performed to get the corresponding hidden state with a combination of least squares and minimum-norm methods. And, we are sorry that we omitted explicit description of the Equation 2 in the paper. Your understanding is completely correct. We did not take all j∈o2 tokens into account. In practice, we only consider the first token in o2. As you said, this is enough for the purpose of this localization experiment.

Response to Q1.c in Questions For Authors

Corresponding to the example you gave, it is a two-hop fact recall chain. It includes (the louvre, located country, france) + (France, capital, Pairs). Instead of using a counterfactual example to perform a conventional corruption, we refer to [1] and reduce the amount of information about France in the last token position by minimizing its corresponding logits value. Then, we observe the causal effect of France, the implicit intermediate subject, by observing the decrease in the probability of Paris. And the "invent" is a spelling error, and we will correct it in the camera-ready version.

Regarding the issue you mentioned about the significance of causal effects, we have provided a detailed explanation in the section "Response to Q1.2 & Q1.3 & Q1.4 in Claims And Evidence about the causal intervention experiments" in our reply to Reviewer JTzm. Please refer to that section for more details.

[1] Understanding and Patching Compositional Reasoning in LLMs

Response to Q2 in Questions For Authors

We construct a multi-hop prompt to put the knowledge to be modified into the implicit reasoning step. While mquake only provides a single edit template, we give a practical example here.

We use all the relationships that have appeared in MQuAKE to construct possible multi-hop edit prompts and then delete those that have the same explicit subject as the test multi-hop question, keeping the best one that the model can successfully answer.

In Table 3, the other methods we reported only used single-hop editing prompt, and only IFMET used our framework. For the application of multi-hop prompt in other methods, please refer to the section in the reviewer 1ooJ's response. We reported in detail the improvement achieved by applying our framework to MEMIT, which illustrates the generalization of our method for editing baselines.

Response to Q3 in Questions For Authors about the CoT.

We are sorry that our examples are confusing to you. We may not have explained them clearly enough. As you said, our cot setting is to cut the prompt after Thoughts:, so that the LM has to generate the thought itself (pressured by the prompt and ICL examples). For the same multi-hop problem, we generate cot and its corresponding answers three times to ensure that possible errors are avoided. We will correct this in the new version PDF.

Thank you very much for your recognition of our work and for your suggestion to extend the framework to other editing methods (such as MEMIT) to demonstrate the generalization of the method itself. We conducted detailed experiments to verify this point. Due to time constraints, we tested it on the ablation set (a subset of 500 items consistent with the ablation experiment). The experimental results on MEMIT are as follows

It can be seen that using MEMIT as the baseline and using PMET as the baseline (Table 7) have similar trends in the four metrics of Multi-hop Acc, Efficacy, Specificity, and Paraphrase, which proves that our framework is generalizable on different editing baselines.

Response to Q1.1 in Claims And Evidence.

In our actual experiments, we conducted a simple exploration, and found that the probability of articles, e.g. the, a will increase significantly in the last few layers, which will squeeze the absolute probability of the answer to a certain extent. But the answer token is still in the position with the highest probability, so it will be output in the last layer.

Response to Q1.2 & Q1.3 & Q1.4 in Claims And Evidence about the causal intervention experiments.

First of all, we need to state that we reported the absolute probability change in the paper. This was our oversight. Based on your and reviewer ixiZ’s suggestions, we revised the causal intervention method in the paper. We re-conducted a detailed intervention experiment with a window size = 5 to explore the the role of layer hidden state and components like MLP and ATTN at different positions of prompt with the decrease of probability in the last prediction layer as the causal effect IE. The experimental results are as follows:

1. Prompt position: We compared last subject token position (which is also generally considered important), the last token of first hop and the last token position. Any intervention effect at the last subject token position (<3%) and last token of fist hop (<2.5%) was weak at the layer level or at the attention head and mlp levels. The response at the last token position was significantly stronger with largest response over 12% in deep layers.
2. Layer intervention: We reached a consistent conclusion that the highest decay occurred in the 16-21th layer, whose largest effect about 12% on prob. We compared the strength of the average causal effect mentioned in figure 2 and 3 of the paper[1]. The intervention effect ranged from about 2.5% to 15%. We believe that such strength on prob and logit is acceptable.
3. Component: We compared the effect of intervening the input of the attention head and the MLP on answer prediction. The effect of the MLP is significantly better than that of the ATTN head, generally reaching more than 12% in the deep 18th-24th layers, while the ATTN is less than 3% over all layers.
4. In general, the results of the intervention experiment are consistent with the phenomenon observed in Figure 2. The implicit subject information plays a causal role in the generation of the final answer through the deep MLP.

[1] Locating and Editing Factual Associations in GPT

Response to Q2 in Questions For Authors about the performance of pre-type tasks.

Sorry for that our definition of pre and post type in the Table 4 is ambiguous. In the paper, we have added the cases of editing the first n consecutive hops into the pre category, such as editing the first and second hops in 3 hops. But as you said, the important insight is to divide according to whether only the first hop is edited, so we reclassified the results into pre and post type, corresponding to the modification of only the first hop (i.e. explicit hop) or including n>1 hops (implicit hop) in correspondence with our conclusions in the exploration section and esperiment results in Table 2. After the new classification, there are 426 pre-type cases and 2574 post-type cases, the results in Table 4 under the new classification are as follows:

It can be seen that the previous method itself achieved significant results on pre type, as we expected, and our method is basically on par with it (this is predictable because the knowledge in the early MLP is fully updated in both cases). On post type, our method shows obvious superiority. This further confirms our findings in the paper. We will modify this in the paper, which will greatly improve the consistency and clarity of our article.

Response to Q1 in Questions For Authors & Q1.5 in Claims And Evidence about n-hop

The reason why we temporarily conducted experiments on 2-hop is partly because we followed the basic settings of previous work; on the other hand, the performance of the model for three-hop and four-hop in the case of cloze-prompt (being able to directly answer the final answer in the form of the highest probability token) is not ideal. Due to the limitations of data volume, existing models, and computing resources, we hope to explore this in future work.

Thank you very much for your recognition of our work. Here is our response.

Response to Q1:How is the dimension of K0 determined in Equation 3?

K0 represents the knowledge we try to preserve when modifying specific fact. Throughout the calculation process, we do not calculate K0 separately, but follow the practice in Rome[1], treating $C = KK^T$ in Equation 3 as a constant that we pre-cache by estimating the uncentered covariance of k from a sample of Wikipedia text. Our second moment statistics C ∝ E[$kk^T$] are computed using 100,000 samples of hidden states k computed from tokens sampled from all Wikipedia text in-context.

Response to Q2: Is the performance of the methods in Table 3 related to the number of edits? As shown in the table, the performance of 3-edit is worse than both 2-edit and 4-edit.

We believe that the performance of answering multi-hop questions after the model implements knowledge editing is usually inversely proportional to the number of edits. However, in the actual implementation of the KE method, it is also related to the quality of the prompt used for editing. We checked the prompts used in the 3-edit part and found that the confidence of the model when answering based on them was relatively low, which means that the model is more likely to be biased when recalling, which means that the multi-hop prompt in 3-edit cases may not be optimal, but we believe that constructing a better multi-hop prompt can be left for future exploration. This article only represents the lower limit performance of the proposed framework.

[1] Locating and Editing Factual Associations in GPT