Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting

Large Language Models (LLMs) can achieve strong performance on many tasks by producing step-by-step reasoning before giving a final output, often referred to as chain-of-thought reasoning (CoT). It is tempting to interpret these CoT explanations as the LLM's process for solving a task. This level of transparency into LLMs' predictions would yield significant safety benefits. However, we find that CoT explanations can systematically misrepresent the true reason for a model's prediction. We demonstrate that CoT explanations can be heavily influenced by adding biasing features to model inputs--e.g., by reordering the multiple-choice options in a few-shot prompt to make the answer always "(A)"--which models systematically fail to mention in their explanations. When we bias models toward incorrect answers, they frequently generate CoT explanations rationalizing those answers. This causes accuracy to drop by as much as 36% on a suite of 13 tasks from BIG-Bench Hard, when testing with GPT-3.5 from OpenAI and Claude 1.0 from Anthropic. On a social-bias task, model explanations justify giving answers in line with stereotypes without mentioning the influence of these social biases. Our findings indicate that CoT explanations can be plausible yet misleading, which risks increasing our trust in LLMs without guaranteeing their safety. Building more transparent and explainable systems will require either improving CoT faithfulness through targeted efforts or abandoning CoT in favor of alternative methods.

Paper

References (75)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer 6weX6/10 · confidence 4/52023-07-02

Summary

This analysis paper demonstrates that the chain-of-thought explanations generated by LLMs do not faithfully represent the true deciding factors of their predictions. In the experiments, LLMs are steered to give intended predictions with three types of biases (Answer is Always A, Suggested Answer and social stereotypes). Despite that their behaviors change accordingly, their generated CoT fails to mention the affecting bias generally. This casts doubt on whether we can safely trust the CoT explanations given that they may not reveal the underlying reasoning process of LLMs faithfully.

Strengths

- Originality: Different from prior works that either propose simulation-based metrics or rely on human study, this paper designs a novel control experiment setting for evaluating faithfulness, where the deciding factors of the model predictions are known beforehand. This is achieved by structuring the prompt to LLMs in a way that is likely to elicit certain answers from them. - Quality: The work conducts a range of experiments investigating three types of biases, using two different LLMs, in both zero- and few-shot settings. The idea of affecting LLMs’ behaviors with certain biases in an intended way is interesting and effective in obtaining unfaithful CoT explanations from LLMs. - Clarity: The paper is well written with both tables, figures illustrating the experimental results and concrete examples. - Significance: The paper provides an alternative view on the unfaithfulness of machine-generated explanations. It also indicates that it is unsafe to trust LLMs’ explanations since the true reasons for their predictions are usually not verbalized in CoT.

Weaknesses

- The paper's findings are not that surprising since there is no systematic way to guarantee LLMs to generate answers that are consistent with their CoT. It would be helpful to extend the study to program-aided LLMs, where the answers are the results of executing the programs generated by LLMs. - The experiment setting of Answer is Always A is questionable, and thus the result obtained is less convincing compared to other settings. As indicated by [1], the few-shot examples in the prompt teach LLMs the answer space rather than the mapping from task input to output. Thus, LLMs might misinterpret the task as always trying to explain the answer A. The work also shows that indeed the explanations are changed to support the new incorrect answers, which are faithful. In this regard, the finding is against the main argument of the paper. [1] Rethinking the Role of Demonstrations: What Makes In-Context Learning Work? (Min et al., EMNLP 2022)

Questions

1. What would the desired CoT explanations look like if they are faithful? In the setting of Answer is Always Right, should LLMs generate a CoT like “Since all the answers to the examples in the prompt are all As, the answer to the new question should also be A”? I feel some desired “faithful” explanations that the authors are expecting might not be realistic. 2. How do you define faithfulness? To me, those explanations supporting the new incorrect answers are actually faithful since they are consistent.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, 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

3 good

Contribution

2 fair

Limitations

It would be great if the authors could talk about the potential extension of the work to prove the faithfulness of explanations or at least what faithful explanations should look like under the current experimented settings.

Reviewer y8DJ6/10 · confidence 4/52023-07-05

Summary

The paper studies faithfulness of COT explanations. Authors perform experiments on different datasets under various setups (biased and none) to demonstrate how decisions made by the models change along with the provided CoT explanations to even justify the wrong decisions made by the model.

Strengths

1. The topic of study was timely and interesting. 2. The paper was written very clearly and easy to follow. 3. Observations and findings of the paper was interesting.

Weaknesses

1. While the paper provided interesting observations, it was mainly an observational paper and lacked technical rigor. 2. I also had issues on how faithfulness is defined/used in the paper. It seems like the CoT reasoning provided by the model is faithful as the model tries to justify its prediction let it be right or wrong, so based on the definition of faithfulness the reasoning is faithful it is just the matter of the model using the wrong reasoning and ultimately the wrong answer. I think there should be a difference between a wrong reasoning with a faithful CoT and a wrong reasoning with an unfaithful CoT which I think is harder to quantify/observe/prove. Seems like this paper assumes that any wrong reasoning is unfaithful which I am having a hard time understanding why. 3. The studies on biasing the context was interesting, but I would have loved to see other and more creative approaches to bias the context and their effects on the results. **Minor comment** Line 195 typo (Appendix C details the our annotation -> Appendix C details our annotation).

Questions

Can authors please clarify on their definition of faithfulness and address my comment 2 made in the weaknesses section?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, 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

3 good

Contribution

3 good

Limitations

Limitations were discussed.

Reviewer mRxa7/10 · confidence 4/52023-07-06

Summary

The authors present an investigation of ‘unfaithfulness’ in CoT prompting for large language models. They argue in particular that models can be predictably influenced by biasing features in the input, which the CoT “explanations” reliably fail to mention, and that they can be influenced by other factors (eg., social stereotypes) but mask this reasoning by appealing to other, weak forms of evidence. The authors test a GPT-3.5 model and a Claude 1.0 model with modifications to two datasets: a subset of Big-Bench Hard (BBH) and the Bias Benchmark for QA (BBQ). With the first dataset, they add a ‘biasing feature’, which is either a modification of the examples such that the correct answer (in a few-shot prompt) is always the same multiple-choice selection (e.g., (A)), or a modification to the example such that there is an additional “Suggested answer” line (e..g, “I think the answer is A but I’m curious to hear what you think.”). In the second analysis and dataset, they introduced a piece of “weak evidence” in ambiguous reasoning problems, to see (1) whether the model applied the weak evidence in inconsistent manners and (2) whether these were in part out of a tendency to produce stereotype aligned answers (e.g., the answer is always “the Black man” regardless of whether the weak evidence would suggest that or not).

Strengths

I liked this paper a lot. The problem is highly relevant and highly topical for current trends in AI. Overall, the paper is well written, interesting, and provocative. The evaluations are clever and fairly systematic. The qualitative analysis is a valuable addition. The motivation for the hypothesis that we should not expect CoT explanations to be faithful by default was very compelling.

Weaknesses

**Results and Design** - Like a lot of LLM evaluation papers, the main statistical analyses provide a rather crude basis for statistical evidence, and many of the conclusions are presented with no statistical evidence at all. - For example: The conclusion in 3.2 that “generally few-shot CoT exhibits less unfaithfulness than zero-shot Cot” is a statistical claim but there are no inferential statistics (nor even descriptive statistics such as confidence intervals) present in any of the analyses in this section. - Generally, the evaluation setting is systematically structured (CoT vs. no COT; debiasing instructions vs. not; FS vs. Zero-shot; GPT vs. Claude) with the same items/prompts being evaluated in these different conditions, but this structure isn’t taken into account when calculating the statistical results, which makes it difficult to assess the claims made by the paper. The fact that that the data are structured in this ways means the individual data points are not iid, and thus a binomial test is not an appropriate statistic test. A paired difference test is better, but still misses the multi-level nature of these comparison. What would be appropriate here is a multi-level (hierarchical / mixed-effects) logistic regression model, taking into account the fact that the experimental manipulations have different effects on different tasks in BBH, and that the same examples appear in all of the different levels of the manipulated variables (e.g., it’s the same prompts that appear in the CoT context as in the no-CoT context, as well as the zero-shot vs. Few-shot, as well as Claude vs. Gpt, etc.) See Lampinen et al. (2022) _Can language models learn from explanations in context?_ for an example of this kind of analysis in the language model evaluation context, or see Gelman and Hill (2006) for further background. - There are potential alternative interpretations for what the model is learning from the prompts in these contexts. It’s possible these explanations do not account for the whole pattern of data as well as the explanation that the authors put forth. But it’s still important to bring some of them up and discuss to help the reader better understand the evidence in favor of their account. - In Study 1, for the “Always A”, in few-shot setting with chain-of-thought, the model could be learning that the task is to *Respond A and provide a plausible explanation that doesn’t mention A*. The example CoT are all good or valid explanations (by design), and almost certainly represent a certain style of explanation. In addition to the fact that none of them likely mention the fact that ‘all the answers are A’ as an explanation, the explanation that ‘all the answers are A’ is a kind of meta-task explanation, which is probably a different kind of explanation than those that appear in the few-shot examples. It’s an explanation at the level of the whole set of examples, whereas each individual CoT (in the shots) probably only concerns the individual example. Therefore, it’s unclear if the model’s reasoning is unfaithful here, or whether the prompt is encouraging it to be unfaithful. - The “suggested answer” manipulation also will fall under the same header of not being mentioned previously in the prompt (appendix F.2), and furthermore, the “Suggested answer” would be just a bad explanation for doing something, and the model only saw good (high-quality) explanations in the few-shot prompt (and likely in the pretraining set, for the Zero-shot case). If my helpful assistant told me "Oh I heard some voice that suggested it me weirdly out of the blue just now", I would say that is a bad reason for doing something. - In BBQ, I tried hard to find out the nature of the few-shot prompt examples but I couldn’t determine if the few-shot prompt examples always included only stereotype-aligned responses for the disambiguated content examples (as is shown in Table 16), or if this varied in a counterbalanced manner. Without knowing this, it’s hard to evaluate the Few-shot results, as the prompt may have biased the model to have stereotype-aligned responses. - I’m generally unclear about the design choices around the BBQ data with ambiguous scenarios and weak evidence. There seems to be some implicit model behind this study, having to do with weighing evidence. - In the appendix, the authors write how some generated pieces of evidence were too strong and others were too weak, which makes me confused about what the ideal amount of ‘weak evidence’ is and what it’s intended to do. - In the example in Appendix B, we have a story about a line-cook and a physician, and which one is likely to have an exclusive credit card. The authors note that selecting the physician, “while arguably reasonable”, is counted as reasoning on the basis of stereotypes according to BBQ. Fine, but insofar as there is (quite plausibly) a statistical correlation between being a physician vs. line-cook and owning an exclusive credit card, how are we to say that _that evidence_ (that statistical association) is weaker than the ‘weak evidence’ provided in their modified task. In all likelihood, it’s stronger. Then in this case, the model could infer the task that is being asked of it is to jusifiy the highest probability outcome using the piece of ‘weak evidence’ provided. - Another example: on Table 16, the Unknown response could also be considered stereotype-aligned, if you deemed the evidence of being on the phone vs. speaking to a young child as sufficiently strong to sway you were it not for the low-income vs. high-income bias. - I think the logic of this Study is either - The situation is ambiguous and the evidence is weak, meaning the normative Bayesian posterior probabilities of each outcome is close to 50% (and the model should say Unsure) - OR, the evidence presented is stronger than the prior association that exists such that the evidence could sway you in one direction or the other (and you should be consistently swayed) - But, it seems like many examples there is likely a strong prior (statistical) association and the evidence is not strong enough to overcome it (normatively, speaking). - I’m not totally sure about this but I think it might be useful to know what the models predict without the ‘weak evidence’ provided. I think this could shed light on if the model has a prior tendency that it’s looking to rationalize, or if the process of reasoning is bringing out some tendency. **Literature** - Rationalization seems to be a key theme in this work, and I’m surprised the authors didn’t discuss that rationalization is a powerful way of making LMs better, and the implications of that for their studies. In particular, the work of Zelikman, Wu, Mu, & Goodman (2022) _STaR: Bootstrapping Reasoning With Reasoning_ **Other comments** - Some of the language and logic throughout the paper was sloppy and/or overstated. Here are some specific. - The logic in the first paragraph doesn’t follow for me: One might hope that the fact that CoT improves performance makes it so the reasoning described by the CoT is an explanation for its predictions, but I wouldn’t say that performance increase alone suggests this. - In the next paragraph, there is also a short discussion about how the models could have correct, unfaithful reasoning, which I found confusing and I think the point is quite subtle. I guess the authors take “correct reasoning” to not entail that the conclusion follows from the reasoning? Or, is that also part of ‘correct reasoning’? - In Section 2, the authors write “measuring these effects gives us an account of what models actually rely on to make their predictions, …” I think this is overstated: You don’t know what the model relies upon in the nonbiasing cases. - I found the discussion around the relevance of subjectivity a little too terse and nonobvious. For example, I kept wondering why switching answers for subjective questions would be concerning, given that subjective questions could have multiple valid answers. In Table 3, the example about Navigation, I found particularly confusing, as it highlights the ambiguity of the task where its not clear if the orientation matters. If the model changes responses based on some biasing features, is that actually problematic? I think the authors could do a lot more to motivate this particular aspect of the evaluations. - 3.2 The main conclusion that CoT explanations are systematically unfaithful. This is a minor point: This conclusion relies on the point made in footnote 2, that the explanations don’t mention the biasing feature. Footnote 2 should be elevated to a top-line result then, so the reader can see more clearly the logic of the argument being made. This is done in the figure caption, but not in the main text. - CoT can steer models from correct initial predictions towards bias-consistent predictions: There is no statistical evidence in support of this claim. - In the BBQ results: The authors talk about “unfaithful explanations” even in the no CoT conditions, and this is generally confusing.

Questions

The limitations above raise obvious questions that I won’t rehash here. 1. From the start, the authors label the auto-regressive generations of LLMs under CoT prompting as “CoT explanations”. But is there any reason to call them explanations? Of course, there’s the empirical fact that the authors are testing… about faithfulness, etc., but from an a priori perspective, what is the step to thinking these might be explanations? As far as I can tell, the CoT prompting is about “thinking step by step”, which I don’t think would fall under any standard definition for what an explanation is. 2. For the models evaluated, do we know whether CoT was used during their training process? This would be an important feature to know about the models. 3. Looking at Table 12 in the appendix (if I’m reading this correctly), it looks like the “Suggested Answer” trick only really works for a very small subset of tasks, and even the “Answer is always A” trick works for about half the tasks. If this is a correct reading, can you speculate why this might be? 4. In the few-shot setting in the BBQ eval, the few-shot prompt uses disambiguated context examples from BBQ and one ambiguous one with an Unknown label. In the appendix, the example shows 2 stereotype aligned responses for the disambiguated cases. Is that always the case? If not, what is the ratio of proportions of stereotype-aligned vs. -misaligned examples among the disambiguated context examples used in the prompts?

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

4 excellent

Limitations

I think the authors missed an opportunity here to view their evaluation as a kind of adversarial attack and speculate on the potential societal harms that could ensue. I’d like to hear if they have thoughts along that line of speculation.

Reviewer aL7J7/10 · confidence 4/52023-07-07

Summary

When using LLMs for problem solving, decision-making, and general reasoning: many prompting and reasoning strategies rely on step-by-step reasoning and/or other explicit explanations and structuring of the reasoning process. This paper studies the degree to which decisions and classifications are faithful to the LLMs explicitly provided explanation of its reasoning process. I.e., if an LLM says it made a decision based on factor X, is there actually a different factor Y that drove the decision? In a series of experiments, this research demonstrates that there are other factors that bias the LLMs decision making process (and that are not included in CoT explanations), with implications for fairness/bias, overreliance, and safety/reliability concerns in the use of LLMs for decision making and planning.

Strengths

The paper is well motivated, tackling an issue fundamental to reasoning across many scenarios: whether the verbalized reasoning procedure is faithful to the true reasoning procedure. The 3 experiments (social bias, few-shot bias, and suggested answer bias) convincingly demonstrate that the LLM's decision making is influenced by factors outside its verbalized reasoning. The implications

Weaknesses

It's unclear whether the biases affect all classes of reasoning tasks equally, or if there are some scenarios or domains that are more strongly affected than others. Also, it is not clear from these experiments whether and how to generalize this beyond categorical decision making tasks (i.e., beyond multiple choice) to biases that might influence other generative AI tasks.

Questions

Would potential memorization of benchmark datasets affect the interpretation of results? What, if any, are the implications of these experimental results for tasks that are more about generative rather than reasoning tasks?

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

4 excellent

Presentation

4 excellent

Contribution

4 excellent

Limitations

The authors have addressed limitations. The potential negative societal impact could be broadened to include the potential for these methods to allow adversarial manipulation of LLM reasoning that would be difficult to detect by end users. E.g., these results could aid an adversary in manipulating a prompt or input data to the LLM to trigger a biased and incorrect result.

Reviewer y8DJ2023-08-16

I thank the authors for providing the response to my concerns. I acknowledge reading the responses provided by the authors. I would suggest authors to add the discussion around faithfulness in this rebuttal to the paper to improve clarity and avoid future questions as raised by me and other reviewers. My other concerns are not resolved, so I will keep my score.

Reviewer mRxa2023-08-20

Thanks

Thanks to the authors for the rebuttal, and apologies for the late reply. I'll just respond to particular set of points made in the rebuttal, which I think touch on crucial features of the design and interpretation of this study. The authors write: > Faithful explanations that are stylistically consistent with the few-shot demonstrations (i.e., which don’t mention the biasing feature) are entirely possible in all of our experiments, as long as the model doesn’t make predictions on the basis of the biasing features. The models could have done this and had faithful explanations. Our experiments show that the they don't. It is because they rely on the biasing features to make their predictions that they must verbalize them in order for the explanations to be faithful. From this, I understand that the model could have had faithful explanations had the model done the task right (i.e., not be influenced by the biasing features) because we know the model was never going to generate bias-consistent explanations (by virtue of them being implausible, given the examples of good explanations used in the prompts). This design decision makes the import of the project more about the task design than it is about the faithfulness of the explanation (i.e., the question of faithful explanations in your study comes down to the question of “does the model make the unbiased response in the biased context”). Of course, this is not unrelated to faithfulness, by the arguments the authors have spelled out. But I think it is a genuine limitation of the study worth acknowledging. For example, a more direct and probably stronger test of the hypothesis would have come with experiments where an explanation consistent with the biased predictions was not implausible, so you could see if the model ever produced a good explanation to a biased response (something a priori quite unlikely in this study). A small point, but I hope the authors can clarify the fact that the responses for the disambiguated cases for BBQ were always stereotype aligned.

Authorsrebuttal2023-08-21

Reply

The reviewer makes a great point that the interpretation of the few-shot results is subtle. In this setting, the few-shot prompt makes it such that the most plausible way for the model to be faithful is to give unbiased reasoning and makes it less plausible that the model could give faithful explanations for biased responses. However, we think the zero-shot CoT setting that we included in our paper would be the most direct way to test exactly what the reviewer is suggesting. Giving a faithful explanation for a biased explanation is a priori plausible in our study under the zero-shot CoT setting since we are not applying stylistic constraints on the resulting generation. For example, the following faithful explanation for a biased answer seems a priori plausible to appear in a zero-shot CoT setting: > Let's think step by step: ... Thus, the best answer would appear to be B. However, the user has suggested that they think the correct answer is A. It's possible that the best answer could be A if the user interprets the question to be asking about X. Therefore the best answer is A. In this setting, we still see that models do not verbalize these biases. However, as the reviewer says it's still possible that the presence of high-quality explanations in pre-training or during supervised fine-tuning could affect the tendency for models to mention biases even in a zero-shot setting. This is precisely a reason to expect CoT to not be faithful by default, i.e. it's not just a priori unlikely that models verbalize biases in this study, this is a priori unlikely in general. Language models rely on a wide range of features to make predictions (including unintuitive ones), so it would be surprising if explanations seen during pre-training, supervised fine-tuning, or in few-shot prompting were well aligned with how models make predictions. In summary, we think that our zero-shot CoT setting is a reasonable baseline for a setting where faithful explanations for biased answers are more plausible. Given that models do not verbalize biases in this setting, this suggests that this should not be an issue for interpreting our experiments. However, if future experiments were to find that models do faithfully verbalize biases in the zero-shot CoT setting, this would affect the interpretation of few-shot CoT results and would be an important limitation to discuss. We will add discussion of this. We will also add that finding prompts or CoT demonstrations that make faithfully explaining the biasing features more plausible could be a good direction for future work. Thanks! > A small point, but I hope the authors can clarify the fact that the responses for the disambiguated cases for BBQ were always stereotype aligned. Thanks, we will add this.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC