Does Localization Inform Editing? Surprising Differences in Causality-Based Localization vs. Knowledge Editing in Language Models

Language models learn a great quantity of factual information during pretraining, and recent work localizes this information to specific model weights like mid-layer MLP weights. In this paper, we find that we can change how a fact is stored in a model by editing weights that are in a different location than where existing methods suggest that the fact is stored. This is surprising because we would expect that localizing facts to specific model parameters would tell us where to manipulate knowledge in models, and this assumption has motivated past work on model editing methods. Specifically, we show that localization conclusions from representation denoising (also known as Causal Tracing) do not provide any insight into which model MLP layer would be best to edit in order to override an existing stored fact with a new one. This finding raises questions about how past work relies on Causal Tracing to select which model layers to edit. Next, we consider several variants of the editing problem, including erasing and amplifying facts. For one of our editing problems, editing performance does relate to localization results from representation denoising, but we find that which layer we edit is a far better predictor of performance. Our results suggest, counterintuitively, that better mechanistic understanding of how pretrained language models work may not always translate to insights about how to best change their behavior. Our code is available at https://github.com/google/belief-localization

Paper

Similar papers

Peer review

Reviewer YdCx8/10 · confidence 4/52023-06-30

Summary

The paper provides an experimental study on localization claims of causal tracing and editing methods such as ROME. Existing literature uses localization results to decide which layer to edit for knowledge editing in large language models. The paper shows that, surprisingly, localization and edit success are nearly uncorrelated. In fact, one can edit a large range of MLP layers to achieve editing success. Yet, localization suggests that only a few early-mid-layers store the factual information. The paper then proposes new tasks and datasets to reconcile this issue.

Strengths

This is a very solid piece of empirical study. I find the conclusion convincing, as it is well supported by the experimental evidence. The conceptual contribution is clear and novel. In particular, the uncorrelatedness of localization and edit success, as far as I know, has not been noticed by the literature. In fact, this is a somewhat surprising observation given the line of prior work such as ROME. The paper is well-written.

Weaknesses

I do not find major weakness of the paper. I have some minor comments and suggestions. The Appendix contains additional experiments on zero ablation. Another technique one can try is mean ablation and resampling ablation, over a known datatset such as CounterFact. This may be more principled than zero ablation. See the [Causal Scrubbing](https://www.lesswrong.com/posts/JvZhhzycHu2Yd57RN/causal-scrubbing-a-method-for-rigorously-testing#4_Why_ablate_by_resampling_) work for some discussion. I believe this wouldn't make much difference here, but it's worth trying. Figure 1: Is this produced by aggregating over the CounterFact datasets? Would it be possible to provide a fine-grained look into this? What are the facts that appear to be stored at later layers? Line 145: Here, the noise level is chosen to be 0.094 universally. However, in ROME, this is chosen to be three times the observed standard deviation of embedding norms as sampled over a body of text. Note that this quantity depends on both the data and the specific (trained) model. Should we also follow the convention of ROME here, to make a fair comparison? In my experience, causal tracing is somewhat sensitive to the level of noise. (I doubt it would change the conclusion, but it is worth looking into.) Line 353: It might be worth a quick footnote here that not all examples have their tracing effect concentrated at the last subject token index. See Figure 11 of the original ROME paper for some natural instances. Figure 10 in Appendix: y-axis should be Tracing Effect rather than Denoising Effect. You used representation zero-ing in Appendix but representation zeroing in the main paper

Questions

From a mechanistic interpretability point of view, could the author(s) provide more thoughts on the effectiveness of causal tracing (in locating which layer *mechanistically* stores the facts)? In particular, would you argue that the model computation is really not as concentrated as causal tracing suggests? Is it rather a diffuse and incremental process, with lots of redundancy? It seems from the window size = 1 experiment (Figure 10 in Appendix), the picture clearly shows that no single layer gets much tracing effect. Thus, would you agree that the sliding window patching is not principled and could yield a false illusion that the information is sharply concentrated?

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

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

Limitations

The author(s) addressed several limitations already. I believe it's adequate.

Reviewer FzCS7/10 · confidence 3/52023-07-02

Summary

Focusing on recently popular model editing methods, the authors study whether localizing methods based on causal mediation analysis play a decisive role. Based on the sample-wise analysis of the tracing effect and rewriting score, they find some surprising results that model editing performances (rewriting score) are no correlation with the tracing effect and are more dependent on the choice of the layer. They also design other model editing performance metrics to verify their findings.

Strengths

This paper talks about an important and interesting problem, the stability or robustness of the interpretability metric. This is why I scored it 6. 1. The metrics about localization (tracing effect) and model editing (rewriting score, generalization score, etc) are inconsistent: I had this question that why the authors choose different metrics for tracing effect and model editing performance when I read ROME. Recovering the correct object is different from modifying the original relation between the subject and object. The authors consider this point and also design more metrics to measure model editing performance from different perspectives. 2. The authors conduct comprehensive experiments to support their findings. 3. As mentioned in Summary part, these findings are surprising and will inspire further exploration of localizing and model editing methods.

Weaknesses

1. The main conflict between this paper with ROME is the viewpoint to switching layers, which is also very important to understanding causal localizing for us: 1.1 ROME says that "informed by the Zhao et al. (2021) finding that transformer layer order can be exchanged with minimal change in behavior, ......, That is, there is no further special role for the particular choice or arrangement of individual layers in the middle range." This is why ROME simply chooses one layer in early layers. 1.2 This paper says that "If this hypothesis were true, it would be surprising because one cannot arbitrarily swap layers in a Transformer model without greatly damaging model performance [39]." 1.3 However, the paper [39] just considers BERT model rather than GPT models and conducts experiments on MNLI dataset which is also different from the factual reasoning task. Therefore, I think the authors should conduct more analysis of GPT models and current task to support their viewpoints. 1.4 I expect more analysis of why there are inconsistencies between causal localizing and rewriting after the authors help reveal them, or some guesses at least. That is why I only give 6.

Questions

1. The first question is also 1.4 in the Weaknesses part. 2. As shown in Figure 5 in ROME, the authors also show model editing performance to support their causal tracing method. The metric used in y-axis is the portion cases for P(o^*) > P(o) (Page 7) rather than the sample-wise difference used by the authors. Does the author believe that inconsistency is a result of metric selection? 3. What is the author's opinion on current localization methods like the denoising method? When should one choose such methods? What inherent knowledge in the model did it exactly point out? 4. What is your opinion on the importance of the early layers and what leads it?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

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

Limitations

Please see Weakness part.

Reviewer UfPL8/10 · confidence 5/52023-07-06

Summary

This paper investigates whether causal localization, which "locates" factual associations within LLMs, provides locations in models which when modified have the greatest effect on those particular factual associations. Counter to intuitions and previous claims in the literature. the authors find that causality-based localization is **much** less predictive (~3 percent vs ~95 percent) than layer-choice for predicting the efficacy of editing factual associations (at MLPs at/across layers). They confirm this behaviour across model classes, localization techniques, layer-depth-windows, datasets and models. This finding is of significant import to the mechanistic interpretability community, as it lends credence to the possibility that models can be pushed toward certain associations in-spite of the presence of existing associations _for those same prompts_ which causal tracing would have identified. Indeed, in some cases using ROME (and other approaches) to inject associations at locations identified as relevant by causal tracing is _less performant_ than injecting them elsewhere.

Strengths

This paper carefully scrutinizes a strongly held belief from previous works about "localization" and the storage of factual information in LLMs. Through precise problem formulation and many experimental validations, the authors are able to provide compelling evidence that previous assumptions were incorrect, and that choosing where to inject (/modify) association, versus finding where a model places assocations for use in a forward pass, are _meaningfully distinct_ questions. Presentation, methodology and supplemental results are well-presented and placed within a coherent narrative. Any questions that arose throughout reading were addressed in the appendices or final discussion.

Weaknesses

No major or minor weaknesses were identified. ## Nitpicks * Lines 222 and 237 are duplicated * Line 206 does not appear to relate to the _neighborhood score_, which measures the change in prediction for similar subjects; rather, it seems to be referring to a metric in which $s=s'$ with $s'\cup s* = \emptyset$

Questions

None

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

4 excellent

Presentation

4 excellent

Contribution

4 excellent

Limitations

The primary limitations would be concerns about generality of these findings. The authors carry out extensive experimental investigations along all likely-to-be-relevant choices of configuration, and address outstanding concerns about the relevance of such findings to interpretability more widely.

Reviewer q6TL7/10 · confidence 4/52023-07-08

Summary

This paper investigates the assumption that we can edit the located component to modify the facutal knowledge stored in the language model. The assumption might be false due to the empirical findings of this paper: (1) many knowledge facts stored outside the located parts of the model, and (2) the near-zero correlation between causal tracing edition and edit success. This paper further proposes a variant Fact Forcing to improve the correlation between location and editing, while finds that editting the early and middle layers still performs better.

Strengths

1. This paper challenges the conventional belief that knowledge localization and editing are strongly correlated. It provides a new and interesting perspective that suggests the relationship between knowledge localization and editing may not be as strong as previously thought. 2. The paper provides a comprehensive overview of the locate-then-edit approach for knowledge editing. 3. The experiments conducted are detailed and thorough. The paper explores different localization methods, editing targets, and variations of editing methods. This comprehensive analysis provides valuable insights into the effectiveness and limitations of various approaches.

Weaknesses

1. The paper mentions a difference between the metrics used and Rome but lacks a specific explanation. Can you explain why your metrics differ from Rome? 2. The paper discusses causal tracing methods in relation to fact forcing for editing goals but lacks tracing methods related to other editing goals. Can you consider incorporating tracing methods related to editing problem variants?

Questions

See 1 and 2 in weakness. 3. How do you view the storage of factual knowledge in GPT? According to the paper, should it be stored among a series of interconnected neurons? How should we design editing methods to address this characteristic? In other words, what insights does your conclusion provide for editing methods?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

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

4 excellent

Contribution

3 good

Limitations

N/A

Reviewer YdCx2023-08-10

Thank you for the detailed response! I maintain my rating and recommend accept. For figure 1, I would suggest at least specify what is the subsample on which you produced the plot, and how many prompts there are. In practice, I find that small datasets may be affected by outliers. For example, certain prompts can be a bit resistant to Gaussian noise corruption of their embeddings. That is, the model still retains decent performance after Gaussian noise patching (with noise level chosen as in ROME).

Reviewer FzCS2023-08-11

Thanks for your response.

I have read the author's response and am particularly satisfied with the answers to the last two questions. I also suggest that the authors should consider talking about The 1st question ("The main conflict between this paper with ROME is the viewpoint to switching layers…") in the revised version. I have raised my score to 7. The reviewer.

Reviewer UfPL2023-08-14

Dear Authors, Thank you for the additional clarification! I retain the previous recommendation for strong acceptance of this paper.

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC