MoCa: Measuring Human-Language Model Alignment on Causal and Moral Judgment Tasks

Human commonsense understanding of the physical and social world is organized around intuitive theories. These theories support making causal and moral judgments. When something bad happens, we naturally ask: who did what, and why? A rich literature in cognitive science has studied people's causal and moral intuitions. This work has revealed a number of factors that systematically influence people's judgments, such as the violation of norms and whether the harm is avoidable or inevitable. We collected a dataset of stories from 24 cognitive science papers and developed a system to annotate each story with the factors they investigated. Using this dataset, we test whether large language models (LLMs) make causal and moral judgments about text-based scenarios that align with those of human participants. On the aggregate level, alignment has improved with more recent LLMs. However, using statistical analyses, we find that LLMs weigh the different factors quite differently from human participants. These results show how curated, challenge datasets combined with insights from cognitive science can help us go beyond comparisons based merely on aggregate metrics: we uncover LLMs implicit tendencies and show to what extent these align with human intuitions.

Paper

References (100)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer HXMa7/10 · confidence 4/52023-06-23

Summary

The paper examines the causal and moral judgments made by large language models (LLMs) and their alignment with human intuitions. To do this, the researchers created a dataset of stories from 24 cognitive science papers, annotating each story with factors that influence people's judgments, such as norm violations and the avoidability or inevitability of harm. The authors find that, on an aggregate level, the alignment between LLMs and human intuition has improved with newer models. However, statistical analyses reveal that LLMs and humans weigh these factors differently when making judgments.

Strengths

**Originality**: The paper presents an interesting approach to evaluating large language models (LLMs) by testing their ability to handle tasks related to causal judgments and moral permissibility. The authors have transcribed stories from various papers and used them to test the LLMs, focusing on several factors that influence people's causal judgments and moral dilemmas. This approach is original and provides a new perspective on the capabilities of LLMs. **Quality**: The authors have meticulously transcribed stories from a number of papers, ensuring a wide range of scenarios for testing the LLMs. They have also collected responses for each story from a crowd-sourcing platform, ensuring a diverse set of responses for analysis. **Clarity**: The paper is written in a clear and understandable manner. The authors have explained their methodology and the factors they focused on in a detailed and comprehensible way. **Significance**: The paper contributes to understanding how LLMs handle complex tasks related to causality and morality. This is an important area of research, given the increasing use of LLMs in various applications. The insights gained from this study could be useful for improving these models in the future. The paper also opens up new avenues for research in this area.

Weaknesses

There are a few areas where it could be improved: 1. **Evaluating with more models**: The paper is a bit skewed towards OpenAI models (GPT3 and beyond) for the evaluation. Including more diverse models could provide a more comprehensive understanding of how different LLMs perform on the tasks. This could also help identify whether the observed behaviors are specific to these models or are more generally applicable to LLMs. 2. **Comparing with human performance**: While the paper does a good job of comparing the performance of different LLMs, it does not provide a clear comparison with human performance. This makes it difficult to assess how close the models are to human-level performance on these tasks. Including a human baseline could provide a more meaningful context for the results. 3. **Analyzing incorrect predictions**: The paper could benefit from a more detailed analysis of the models' incorrect predictions. This could help identify common patterns or biases in the models' errors, which could provide insights for improving the models. 4. **Generalizability of findings**: The paper's findings are based on a specific set of stories and tasks. It's unclear how generalizable these findings are to other tasks or domains. The authors could address this by testing the models on a wider range of tasks or by discussing the limitations of their approach in more detail.

Questions

1. **Evaluating with more models**: The paper primarily focuses on GPT-type models and its variants. Could the authors elaborate on why they chose to focus on these models? Would the inclusion of other language models provide different insights? 2. **Comparing with human performance**: The paper lacks a clear comparison with human performance. Could the authors provide a human baseline for these tasks? This could help in understanding how close the models are to achieving human-level performance. 3. **Analyzing incorrect predictions**: The paper could benefit from a more detailed analysis of the models' incorrect predictions. Could the authors provide more insights into the common patterns or biases in the models' errors? This could potentially help in improving the models. 4. **Generalizability of findings**: The findings of the paper are based on a specific set of stories and tasks. Could the authors discuss how generalizable these findings are to other tasks or domains? 5. **Interpretability and transparency**: The paper presents an analysis of how LLMs reason about causality and morality. However, it's not clear how these insights can be used to improve the interpretability and transparency of these models. Could the authors provide some thoughts on this?

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

2 fair

Contribution

3 good

Limitations

The paper addresses the limitations of the work and potential negative societal impact, although not in a dedicated section. The authors acknowledge that their focus is narrow and only on certain aspects of alignment with humans. They caution that their work should not be used to make sweeping and general statements about AI-human alignment. They also note that their moral permissibility task is not a certification task and should not be used as a flat benchmark to beat. The authors also discuss the ethical considerations of their work, emphasizing the importance of assessing implicit intuitions underlying commonsense reasoning abilities in large language models (LLMs), especially in cases related to morality. They acknowledge that even if a model is not explicitly given the responsibility to make moral judgments, these judgments can appear across many forms of freely generated text. They also recognize the potential for replicating human biases in LLMs and state that this is something they would want to avoid. In terms of potential negative societal impact, the authors don't explicitly discuss this. However, they do acknowledge the potential for misuse of LLMs and the ethical considerations that come with their use. They also discuss the importance of transparency and consent in their data collection process.

Reviewer 7W7v7/10 · confidence 4/52023-07-06

Summary

This paper investigates to what extent LLMs can align with human intuitions when making causal and moral judgments. To do this, they collected a dataset of stories from 24 cognitive science papers and created a causal and moral judgment challenge set. They evaluate different LLMs about their alignment with humans and reveal that the implicit preferences can be different even for LLMs trained with the same technique. They find that increasing model sizes actually impact those models’ aggregate-level alignment differently.

Strengths

With the wide spread of LLMs, to understand the alignment between humans and models is an important topic. In this paper: - They have provided a dataset to understand the human-model alignment, especially in the causal and moral judgment perspective. - The resources are from cognitive studies which makes it more reliable than the normal text resources.

Weaknesses

The paper wants to analyze the alignment between humans and models, however it lacks some description of how they conducted the human study.

Questions

- For the factors in Table 2, are they from existing literature reviews or summarized by the authors? - For the dataset, are those factor labels annotated by the authors or by original cognitive scientists? - For the human participants, do you have any criteria to select who can participate in the survey? - Did you educate the participants about different factors before you conduct the survey? - It seems only part of Fig.2 is visible on my side. Please check that.

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

3 good

Contribution

3 good

Limitations

NA

Reviewer duEY7/10 · confidence 2/52023-07-07

Summary

This model presents a new challenge set of hard edge cases intended to test models understanding of the nuances of the directness of causation and moral culpability, by collecting them from a set of cognitive science papers. This has the clever effect of not only getting challenging stories, but those which would vary along specific features important to humans. They test LLMs on those outputs to measure agreement with human intuitions; and annotate those cases for a set of features so that one could draw insight from those disagreements.

Strengths

It is a well-written and well-considered paper which both presents a new useful challenge set, and utilizes it to provide interesting analysis of LLM tendencies in causal culpability and moral judgments. It could clearly lead to further uses both in the evaluation of new models and in further analysis. The literature review is, as far as I could tell, comprehensive. The work seem rigorous throughout - I appreciate the thorough explorations with personas and automatic prompt engineering, which alleviate worries about the normal fickleness of prompt choice.

Weaknesses

- The size of the challenge set (around 200 stories I believe) is somewhat limited; I don't think that that's too much of a worry for such a challenge set, so I wouldn't view it as a major weakness. - quibble: A seemingly left-over note on line 232: " This is very very interesing, make the flow better."

Questions

- I'd be very curious about which personas and prompts would lead to worst-case performance for various models, since that might give insight into how the models go awry. - The improvements in alignment with human judgements from adopting a utilitarian/consequentialist framing is fascinating. However, that doesn't mean that all humans have a utiliarian framing. Is there any concern that measuring against the average of human judgements might ignore variance between different humans on such judgement 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

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

3 good

Contribution

4 excellent

Limitations

The ethical considerations section seems thoughtful, and I see no unaddressed limitations.

Reviewer DYrH6/10 · confidence 4/52023-07-11

Summary

This paper focused on large language models' causal and moral intuitions and investigated the alignment between LLMs and humans' causal and moral judgments. For this purpose, the authors collected story datasets from the field of cognitive science and manually annotated each story with human judgments and underlying latent factors. Based on this dataset, a diverse range of LLMs with different model scales and training methods are evaluated. The authors then statistically revealed that LLMs weigh factors differently than humans, indicating divergent implicit preferences and emphasizing the importance of curated datasets and cognitive science insights in understanding model preferences and alignment.

Strengths

* This paper is well-motivated by philosophy and cognitive science and focused on an exciting topic, LLMs' causal and moral intuitions. Such an interdisciplinary insight would benefit the better understanding of LLMs' behaviours. * The authors summarized a systematic framework of the underly latent factors of casual and moral judgements based on cognitive science, which might help improve the interpretability of LLMs. * The constructed judgment dataset is high-quality, with a well-designed annotation protocol and high inter-rater agreement (>0.8). * The authors benchmarked the alignment level between humans and a wide range of LLMs. They also conducted comprehensive analyses and made inspiring conclusions like those in Sec. 4.2.2, e.g., differences in Benefits.

Weaknesses

* The constructed dataset is too small, and the coverage is limited. Two hundred six instances are highly insufficient to investigate LLMs' properties which might make the conclusion biased. This can be observed in Table 3 (a). The relatively high bootstrapped confidence interval indicates a high variance and unreliable results. This is my biggest concern of this work. * Some essential results need more in-depth analysis and explanation. (1) The unnatural results in Table 3(a) need more analysis. Why did the aligned and larger Alpaca-7B get lower Agg than GPT3-curie-v1 on Causal Judgement? Why did davinci-002 outperform the well-aligned davinci-003 on moral judgement? (2) The authors should provide some (even initial) analysis of the differences introduced in Sec. 4.2.2 though they are attractive.

Questions

* How do you explain the unnatural results in Table 3(a): the aligned and larger Alpaca-7B got lower Agg than GPT3-curie-v1 on Causal Judgement; GPT-4 performed even worse than davinci-003 on Causal Judgement; davinci-002 outperformed the well-aligned davinci-003 on moral judgement. * Would you release your Judgement dataset?

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

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

The authors have discussed the ethical considerations in Appendix A. However, the authors should also include more discussions of limitations, like the small dataset and variance of the results, as stated above.

Reviewer sFvB7/10 · confidence 3/52023-08-01

Summary

The motivation of this paper is that people constantly make lots of causal and moral judgments to reason about why did what things and why. This paper contributes a dataset of stories compiled from cog sci papers, with detailed annotation of the factors that contributes to the human judgment. Then, the paper looks at how LLMs make judgments, and check the alignment with humans.

Strengths

- The paper addresses an important topic to check the causal and moral reasoning and the alignment of LLMs with humans - The proposed dataset looks solid and well-annotated - The analysis provides insights to the community to develop safer and more aligned LLMs.

Weaknesses

- The size of the dataset is a bit limited, 144 causal stories and 62 moral stories, making the insights drawn upon them be not extensive enough - The yes/no binary answer is reasonable, but analyzing LLMs behavior using a binary classification task might have a little signal-noise ratio. There needs to be lots of human annotation to evaluate the reasoning quality of LLMs, and whether any misalignment or unsafe reasoning was provided apart from the binary answer.

Questions

1. Are there domain experts in moral psychology / philosophy involved in the design process of this paper? How do you make sure the factors in 2a and 2b are comprehensive and can explain for all the judgment decisions? I saw the appendix A.1, but I would like to see one dedicated paragraph for each of Table 2a and 2b, describing the rationale behind each factor and how they correlate with human intuitions in the main text in the next version of the paper. 2. Can the authors let LLMs to output its reasoning, and then annotate what type of tendencies LLMs show in its reasoning (maybe doing it on a subset, e.g., 50 samples)? [I have read the rebuttal, and acknowledge the author's effort into it. I'm supportive of the acceptance of this paper.]

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

4 excellent

Contribution

3 good

Limitations

N/A

Reviewer sFvB2023-08-11

Thank you to the authors for the rebuttal and additional evaluation. I'd support acceptance for this paper.

Reviewer 7W7v2023-08-13

Thanks! I have read the authors' rebuttal and do not have further questions.

Reviewer duEY2023-08-13

Hi authors! Thank you for the rebuttal; the answers to my own questions seem persuasive, and I feel that the new analysis (in the rebuttal to other reviews) is useful . I continue to support the acceptance of this paper.

Reviewer HXMa2023-08-15

Thank you for the rebuttal! The answers to my own questions seem persuasive, and I feel that the new analysis is useful. I continue to support the acceptance of this paper.

Reviewer DYrH2023-08-21

Thanks for response

Thanks for the detailed response. I think most of my concerns have been addressed, and I have raised my score to support the acceptance. Please include your additional experiments (more sampled examples would be better) in your final version, which would make this work more convincing.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC