Delving into the Reversal Curse: How Far Can Large Language Models Generalize?

While large language models (LLMs) showcase unprecedented capabilities, they also exhibit certain inherent limitations when facing seemingly trivial tasks. A prime example is the recently debated "reversal curse", which surfaces when models, having been trained on the fact "A is B", struggle to generalize this knowledge to infer that "B is A". In this paper, we examine the manifestation of the reversal curse across various tasks and delve into both the generalization abilities and the problem-solving mechanisms of LLMs. This investigation leads to a series of significant insights: (1) LLMs are able to generalize to "B is A" when both A and B are presented in the context as in the case of a multiple-choice question. (2) This generalization ability is highly correlated to the structure of the fact "A is B" in the training documents. For example, this generalization only applies to biographies structured in "[Name] is [Description]" but not to "[Description] is [Name]". (3) We propose and verify the hypothesis that LLMs possess an inherent bias in fact recalling during knowledge application, which explains and underscores the importance of the document structure to successful learning. (4) The negative impact of this bias on the downstream performance of LLMs can hardly be mitigated through training alone. These findings offer a novel perspective on interpreting LLMs' generalization through their intrinsic mechanisms and provide insights for developing more effective learning methods. Our code and data are available at https://github.com/alibaba/thinking_bias.git.

Paper

Similar papers

Peer review

Reviewer e5Vg7/10 · confidence 4/52024-06-26

Summary

The paper investigates the reversal curse a phenomena where LLM trained on documents on the format “A is B” are unable to improve their likelihood on statements of the format “B is A”. The authors extend the work to consider documents of the format “Name is Description” and “Description is Name”, whilst also considering two tangible downstream tasks - answer completion and multiple choice answering. They find: 1) In multiple-choice settings, models are able to reverse information from “Name is Description” documents to solve questions w.r.t Description (successful mitigation of the reversal curse) 2) however in the multiple-choice settings, models trained on “Description is Name” (“B is A”) documents fail to solve multiple-choice questions better than guessing (which they dub the “thinking bias”). In this setting, both the reversal curse and the thinking bias stifle model behaviour. 3) By inspecting both chain-of-thought reasoning and saliency maps over attention heads - models regularly attend to the name within questions much more than they attend to the description - providing an explanation of why models are so bad at answering questions which begin with a description. 4) Authors demonstrate that further training and synthetic augmentation do not address these issues

Strengths

This paper provides conceptual clarity and an explanation of the original reversal curse and finds a further bias within regressive language models. Figure 1 is exceptionally useful for explaining bais and experiments are completed over a variety of open-weight models (llama, vicuna, mistral) The book-story dataset results in the appendix were really compelling. I recommend that the authors move/highlight these earlier in the paper.

Weaknesses

For these results, my fundamental question is if this is just a symptom of the model sizes used and if, with a larger capacity model - these biases would be mitigated. The original reversal curse work is done on GPT-3.5-Turbo models, and a similar class of models is used here. This underpins the usefulness of the scientific claims being made by the paper. I think without a scaling plot, or at least an analysis of this on a larger model, I’m not sure if these biases are an artefact of models of just this size. Note this does not need to be interpretability or COT work (which are costly), which explains why these biases exist within models, but proof that they exist.

Questions

Could the above weaknesses be addressed?

Rating

7

Confidence

4

Soundness

3

Presentation

4

Contribution

4

Limitations

N/a

Reviewer p2dN6/10 · confidence 3/52024-07-05

Summary

The paper investigates the reversal curse, which is the finding that LLMs fail to generalize from seeing “A=B” to “B=A”, in a broader range of settings. They reproduce the reversal curse findings from Berglund et al (2023). They also investigate a new setting, multiple choice Q&A (rather than free form question answering). In this context, they find that instead of a reversal curse, models can generally answer questions well after seeing demonstrations of the form "Name is Descriptions", regardless whether the question is about Name or Description. However, they fail to answer in general after being trained on “Description is Name”. They investigate this phenomenon further by looking at models’ chain of thought and by applying saliency methods to attention layers and attribute the failure to a “Thinking Bias” in LLMs. They also show that “training on the test” does not work to improve models’ abilities here, like in the reversal curse, indicating that it’s a fundamental bias of LLMs.

Strengths

- Originality: As far as I can tell, their results are novel. They reproduce the reversal curse, and then look at a related phenomenon in multiple choice QA which hasn’t been investigated before. They also apply various different analysis methods which haven’t been applied to this problem before. - Significance: I found this interesting and important to better understand and get a more detailed picture of the generalization abilities of LLMs. They look at several open source models and do relevant additional analysis which shows that this result is significant and not just a fluke. - Quality: Overall the experiments seem well thought out and rigorous. Though I haven’t inspected the data myself. - Clarity: I like a lot of the plots and examples in the paper. I also think I was able to understand the paper well overall and their exploration of LLMs' thinking bias made sense to me.

Weaknesses

My main criticism is about the presentation of the results. I think this work isn’t really about the reversal curse. Rather, I think the work is about a related, but orthogonal issue. They say “this work remains orthogonal” in Line 297 when discussing related work, but I think the rest of the paper isn’t written like this. I think the setting where you put the answer options in the context is just a different setting from the one which is defined in the paper on the reversal curse, so one should study this separately and give it a different name. It’s good to highlight the relation to reversal curse, and I really like the replication of the reversal curse too. But again, the core contribution of the paper seems to be about multiple choice QA and “thinking bias” in LLMs, so I wouldn’t frame the whole paper as “delving into the reversal curse”. Of course, I might have misunderstood something here, so I would be happy to be corrected and would like to hear how the authors see their main contribution.

Questions

- I think you should state explicitly from the beginning (in abstract and introduction) that, when testing multiple choice QA, you found that training on “Name is Description” works, but “Description is Name” consistently fails. I feel like you kind of buried the lead. (As an aside, the abstract reads a bit like it was written by ChatGPT to me) - Line 31: what does it mean to have a capability but to fail to elicit it? Does it mean that a different prompt would elicit the capability? But note that using a different prompt might turn it into a different problem - I wonder why the authors look at saliencies instead of just directly the magnitude of the attention matrices. Is there a reason to do this? Does looking at attention scores not yield the same results? Note I’m not an expert on interpretability, so someone else might be in a better position to comment on this. - Regarding Figure 2: do you have a hypothesis why there is more flow from description in early layers when evaluating name to description? Is this because the last thing in the context is the description in this case? - I think you don’t need to introduce the reversal curse again in related work since you already discussed it in depth in section 2.1 (and you should discuss the relation of your work to the reversal curse throughout the paper anyway) - Line 283: Can you say more specifically how this outcome diverges? What did [52] find? - I think you should discuss Limitations and Future Work in the main text, not in the appendix

Rating

6

Confidence

3

Soundness

4

Presentation

2

Contribution

3

Limitations

I think they address the limitations adequately (except that this section should be in the main text, not appendix). Maybe one thing to mention is that as far as I can tell, they only study one specific dataset. One could of course study all of this in many other settings as well. E.g., a dataset about things other than celebrities, or a dataset where descriptions are not in natural language, etc.

Reviewer p2dN2024-08-11

Thanks for your detailed responses and clarifications (including the new results in the other rebuttals)! I have read them and have found them interesting and convincing. I will increase my score by one point. One comment regarding presentation: I understand that it's useful to refer to the reversal curse in this paper (including in title, abstract, introduction). I still think that your main contributions (thinking bias in multiple-choice Q&A) are separate from the reversal curse and study a different (but very related) phenomenon.

Authorsrebuttal2024-08-12

Appreciation to Reviewer's Response

Dear Reviewer p2dN, Thank you for dedicating your time to read our rebuttals and for raising your score! We sincerely appreciate your constructive comments for enhancing the presentation of our paper. These comments will be carefully considered when introducing the contributions of our work in the final version. We hope that all readers of our work will find their interest and inspiration in the experimental phenomena and our interpretations of today's LLMs' abilities. --- Best regards, Authors

Reviewer PVmx7/10 · confidence 4/52024-07-10

Summary

Building off prior work that studies the “reversal curse” in LLMs, the present paper provides additional analysis on 1) characterizing the limitations of LLMs on the reversal curse through more detailed experimentation (e.g., limitations with chain of thought prompting or providing multiple choice questions), and 2) interpreting the reasons as to why LLMs are biased towards correctly answering A is B when A is a name/proper noun. The main findings report that LLMs can improve generalization on B is A when 1) the prompt includes a multiple choice question; 2) LLMs are biased towards A (proper noun) is B (description), since this is how facts are typically represented in the training corpus; 3) This existing negative bias in LLMs cannot be mitigated by training/finetuning alone.

Strengths

1) The paper addresses a timely question 2) The paper was, for the most part, straightforward and easy to understand. 3) The experiments are interesting, and the effects appear strong

Weaknesses

One of the main claims, that LLMs can disproportionately perform “NameIsDescription” correctly, is due to the fact pretraining datasets are biased towards having text in the form of NameIsDescription (i.e., A is B), but not the reverse. Despite the claim being mentioned several times, and though despite the claim being intuitive, the paper does not empirically quantify or demonstrate this that I could find. Are statistics reported of how often “A is B” is exhibited in the training documents relative to “B is A”? And is performance of the LLMs proportional to the ratio found in training documents? There seems to be some reference to Berglund et al., but this is a result of a prior paper, not the present paper. It would be helpful to quantify how biased LLMs are for A is B vs. B is A relative to the proportion they are exhibited in training data. I found figure 3 to be confusing. What are the different colors supposed to indicate? (There’s no associated color bar.) Also, the incorrect answer of D makes it appear as if the incorrect selection was due to the tokenization of the name Graham Redwood. There is also an issue of novelty – many of the reported results do not seem to particularly ‘novel’, perhaps because the results seem almost obvious. I think it would significantly help if the authors were to more clearly delineate their work from prior work in the Introduction, and to “signpost” exactly what the specific contributions of this work are (relative to prior work).

Questions

1. Is the finding that names are easier to trigger recall potentially due the fact that names typically have fewer number of tokens than descriptions? Prior work has shown that the token-wise MLP layer of the transformers act as key-value memories (Geva et al. 2021). If there are fewer tokens associated with a name, wouldn’t it be easier to coordinate retrieval of memories (i.e., facts) across fewer token streams, rather than to coordinate memory retrieval across the many streams that comprise of “Description” tokens? 2. Related to a weakness mentioned above: What are the statistics/proportions that show a bias in the pretraining corpus of “A is B” over “B is A”? And how does that proportion match with actual LLM performance? 3. In multiple choice question prompting, are LLMs biased towards any particular response (e.g., ‘A’, ‘B’, ‘C’, or ‘D’)? I’m curious to know if the attention weights in a decoder-only model could potentially bias the model to retrieve more facts associated with ‘D’, since it is later in the prompt. 4. I’m skeptical of the interpretation (or over-interpretation) of “information flow” by computing the average attention weights to a given token. This concern is compounded by the fact that the models they used (Llama) are decoder-only models, which, by construction, have greater attention weights towards tokens presented later in the prompt. Might this metric be confounded by this (results in Fig. 2 & 3)? Geva, Mor, Roei Schuster, Jonathan Berant, and Omer Levy. “Transformer Feed-Forward Layers Are Key-Value Memories.” In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, 5484–95. Online and Punta Cana, Dominican Republic: Association for Computational Linguistics, 2021. https://doi.org/10.18653/v1/2021.emnlp-main.446.

Rating

7

Confidence

4

Soundness

2

Presentation

3

Contribution

2

Limitations

The authors claim that the curse of reversal can be somewhat mitigated if multiple choice questions are used. However, this appears to be a major limitation, while also being a strange suggestion – incorporating multiple choice questions assumes that the prompter knows the correct answer. Thus, in what scenario would this be helpful, aside from evaluating and adjudicating performances of multiple models?

Reviewer Mp9z5/10 · confidence 3/52024-07-14

Summary

The authors extend the original reversal curse dataset to two tasks: open ended QA and MCQ. The authors analyze the generalization capabilities of LLMs on reversal tasks and provide several experiments towards their claims. They show that LLMs can generalize from A is B training to B to A, when both B and A are present in the question. Some of the results are expected based on existing results. For me the most significant contribution was the result in the relative intensities across layers. I may be willing to increase my score based on responses to certain questions.

Strengths

- Strong applied work. The experiments are well posed (if only for small models), and do support the papers claims. - In a welcome change from literature, the authors also provide negative results of their experiments in Section 4. While this is great, it is not clear from the paper how to mitigate the reversal curse on the said datasets. - I liked the saliency score used in Section 3.2. Specifically, using saliency as an importance score for the l'th layer was insightful. I was wondering if this could be written as a dot product instead of a Hadamard product?

Weaknesses

- Is the paper analyzing reversal curse? Was that not already done by the original reversal curse paper? - The authors claim that generalization occurs when both A and B are present in the question. Is this possibly due to the structure of attention? - Figure 5 shows that the models have a decreased accuracy when CoT is used. Why so? Should CoT not improve the results (as per earlier descriptions in the paper)? Any discussion on this?

Questions

- Line 115 Was the model overfit to the fine-tuning data? - What is the authors explanation for MCQs on NameIsDescription exhibiting generalization whereas DescriptionIsName not exhibiting the same? - Does thinking step by step (CoT) put the name is description into context? Since in-context does not have the reversal curse problem, does CoT fix reversal curse on the given datasets? - Is there an equivalent Table 4 identifying whether the Description is present in the CoT? Perhaps both Name as well as Description can exist in the CoT? - Figure 5 goes against existing literature [52] as noted by the authors themselves. "We attribute this divergence to the structure of the training documents since their training samples mainly use names or personal pronouns as subjects, which generally mirrors the structure of the NameIsDescription subset." The description provided was not clear to me. - Why not use the saliency scores of particular tokens, like the names, instead of a particular position? - Snt and Sdt are great observations. Can these be used to actually modify the attention? i.e. since we have white-box access to the model, can one actually increase the attention manually to check if this (a) solves the RC problem, or (b) improves generalization in terms of accuracy? - Can the authors clearly outline why their contribution is novel, and what are the major take-aways from their work? ### Suggestions: - Based on the introduction, I could not understand the paper contributions. Can the paper provide a short section or paragraph on that in Section 1? - The legend in Figure 5 is unclear. Does it mean that OOD are the smaller bars which are having the masks?

Rating

5

Confidence

3

Soundness

3

Presentation

3

Contribution

2

Limitations

- No major limitations. I feel some of these results are expected based on Berglund et al. - My concern is on novelty. I am hoping the authors can address this question in the rebuttal.

Authorsrebuttal2024-08-07

Rebuttal-Part 2

Dear Reviewer Mp9z: We are sorry that we were unable to respond to all your questions in a single rebuttal section due to the character limit. We have included our responses to the remaining questions here for your reference, as well as for the benefit of other reviewers. We apologize for any inconvenience this may have caused. # Part 2 > Q8: Can the authors clearly outline why their contribution is novel, and what are the major take-aways from their work? **A**: We hope our response in Global Rebuttal will effectively address your concerns. --- > Suggestion: In Figure 5, does OOD are the smaller bars which are having the masks? **A**: Yes. And we will improve the clarity of this legend in the revised version of the paper. --- > L1 & 2: I feel some of these results are expected based on Berglund et al. My concern is on novelty. **A**: We respectfully disagree. We believe that the only experimental results that can be expected based on RC are the model's performance on Open-QA. The generalization difference between the NameIsDescription and DescriptionIsName groups, particularly the phenomenon that the models cannot identify the correct answer even though they can directly answer the original question without options, is **completely unexpected and even counter-intuitive**. It goes against the common belief that identifying the correct answer would be much easier than producing it from scratch. We hope that the novelty listed in the Global Rebuttal could adequately address your concerns. --- > Additional Question: Can dot product replace Hadamard product in saliency score computation? **Ans**: The incorporation of Hadamard product in Equation 1 is inspired by one of the first works [1] that introduced saliency analysis into language models. Suppose a new mask variable $Z\in \mathbb{R}^{L\times L}, Z\in [0, 1]$ is applied to the attention matrix $A \in \mathbb{R}^{L\times L}$ to control the interaction between token $i$ and token $j$, which gives us $A'=A \odot Z$. The sensitivity of the model to the mask variable $Z$, $I_{ij} = |\frac{\partial \mathcal{L}}{\partial Z_{ij}}|$ can be seen as the importance score of the interaction between token $i$ and $j$. After applying the chain rules, the final expression for $I$ as well as our definition of the saliency score can be written as $I_{ij} = |A \odot \frac{\partial \mathcal{L}}{\partial A}|_{ij}$. We are not sure whether dot product could serve the same purpose as well. But we remain open to any further discussion.

Reviewer Mp9z2024-08-13

Thanks to the authors. It addressed some questions I had during my initial reading of the work. I hope, if the paper gets accepted, the authors can move some of these points into the main body of the work for camera ready. I will be increasing my score, as the rebuttal addresses some of my earlier concerns. I have no objections to the paper getting accepted. There are no technical concerns, or issues related to the soundness of the paper.

Authorsrebuttal2024-08-13

Appreciation to Reviewer's Response

Dear Reviewer Mp9z, Thank you for your feedback! We are deeply grateful for your time in reading our rebuttals and your willingness to reassess our work. In response, we will further polish our final paper based on the valuable points raised by you and other reviewers. --- Best regards, Authors

Reviewer PVmx2024-08-09

Reviewer response to author rebuttal

I thank the authors for thoroughly engaging with my questions, and am impressed with the amount of work they have been able to perform during the rebuttal period. I think this paper would be a great contribution to NeurIPS, and commend the authors on their study. I will increase my score to 7.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC