COLD: Causal reasOning in cLosed Daily activities

We thank you for your detailed and insightful review, pointing towards some suitable directions to make our work more impactful. We would like to address the raised concerns below:

• We want to mention that the human/expert annotations were done to construct the underlying causal graph, capturing the causal relationship between the events. Previously, works like Jin et. al have released causal queries (https://huggingface.co/datasets/causalnlp/corr2cause, https://huggingface.co/datasets/causalnlp/CLadder) generated from the underlying causal graph that does require causal inference theory understanding to be purely annotated by humans. Whereas in our work, the commonsense reasoning nature of the proposed framework provides the primary strength for bridging the gap between current causal reasoning benchmarks/datasets in NLP. Please also refer to our detailed comment on human evaluation in the response to Reviewer iBzV, where we explain the challenges in human evaluation.
• The advantage of using a human crowdsource dataset is the marginalization that we obtain when the same activity is written by different humans. We would like to highlight that the primary strength of this work providing real-world grounding is coming from the observational graphs that are coming directly from humans (capturing the commonsense knowledge acquired by humans). When creating a benchmark or evaluation criteria, it becomes imperative to consider human knowledge. Using the constructed graphs, the proposed design scheme helps sample enormous causal queries that essentially facilitate large-scale causal benchmarking.
• We thank you for pointing this out. We agree that it would be interesting to explore the fine-tuning over the created samples to observe the model’s behavior. We highlight the same in the Future directions of our work in lines 354-357, “In the future, it would be interesting to sample trajectories from the observational distribution to create a training dataset and check if causal reasoning ability can be acquired by language modeling objectives (including other variants like presented in Lampinen et al. (2023)).” Some of the prior arts like Corr2Cause (Jin et. al, 2024) have considered training/finetuning over the symbolic dataset and have observed no generalization, claiming LLMs fail to robustly acquire the causal reasoning skill in out-of-distribution settings.
• We thank you for pointing out these interesting resources. In fact, we found the strategy formulated by us for zero-shot evaluation similar to the one used in the first resource (Ma et. al), the formulation used in equation 1 of Ma et. al's work [https://doi.org/10.1609/aaai.v35i15.17593 ] is similar to our formulation, the difference being the multiple numbers of tokens being used in their setup when compared to simple option id prediction in our formulation. The remaining two resources pointed out (Wang et. al and Kim et. al) are somewhat different and consider relying on the commonsense base (Knowledge Graphs), and would require more work to evaluate in our setting of causal reasoning. We agree that another set of baselines focusing on zero-shot commonsense question answering will be a good direction to explore in the future.
• We thank you for pointing out the typos, we will fix them in the camera-ready version of the paper.

Thank you for providing your insightful comments. Please find the response to the clarifications below.

• Given the wider scope of this work, we agree that the paper might have become too dense, affecting the presentation quality. We would love to hear some more feedback from you to incorporate in the final version of our paper to improve the presentation quality, focussing on the reading sequence for better comprehension.
• We would like to respectfully disagree regarding the difference between COPA and the proposed dataset. COPA (with only 1000 samples), is a more open-ended dataset and the causal queries can be of any generic domain. However, the dataset proposed in our work has enormous causal queries (8.3M in total), covering all the aspects of a particular activity (i.e. not open-ended), that essentially facilitate a more rigorous analysis (coming close to simulating the mini-turing test) and hard to be beaten by just memorization. Thanks for pointing this out, the table shown in the paper was to help consider the similarity between the datasets, we will add a more detailed comparison stating the differences between the datasets in the camera-ready version of the paper.
• We have already highlighted in the limitation section regarding the limited set of activities, and we agree that it would be good to explore more real-world activities in the future. However, the number of causal queries that can be sampled is enormous and is not exactly the paraphrases of each other. Every scenario is different and has a different set of texts describing the events in the particular activity. The paraphrased version refers to the same action/event described by different humans with varying granularity, that are written independently by the crowdsource workers. For example, when describing an event like “preheating oven”, some people will write “preheat oven to 350 degrees”, whereas some people will write it in detail, “preheat the oven by switching on the oven and change the temperature setting to required temperature”. We agree that using the term “paraphrased” for such cases may not be appropriate, we thank you for pointing this out, and we will update the details accordingly in the paper. These different versions of the same event described enhance the dataset quality by a significant margin, providing a much more robust evaluation of LLMs. In other benchmarks, keeping only one version often limits the rigorous evaluation.
• Please refer to our detailed comment on human evaluation in the response to Reviewer iBzV.

Response to Question/Clarifications:

Q1) By examples provided in lines 78-80, we wanted to highlight the importance of counterfactual reasoning required to reason causally between the events, where one does not only need to consider the temporal order of events but also needs to think counterfactually by imagining an alternate universe with the same event not occurring will cause the other event to occur or not. In the example, the event “waiting at the luggage belt” may not occur even if the person has “boarded the flight” and “not checked in the luggage”, hence boarding the airplane will not be a direct cause of “waiting at the luggage belt”. This marks the third rung of the causal ladder proposed by Pearl et. al, which is also considered a core feature of intelligence (Penn and Povinelli, 2007; Pearl and Mackenzie, 2018) [lines 23-24].

Q2) The final version of the dataset assumes the activity to happen (also considered in the evaluation prompts as context). Moreover, the marginalization of pre and post-activity events makes it a closed system for rigorous analysis where all the unobserved variables (Figure 1, represented by U, i.e. intention to perform a task) take care of separating out the causal effect from events that may occur in rest of the world. This setup also provides adherence to SUTVA which is not possible in a dataset like COPA. Some of the prior arts (ROCK[https://arxiv.org/abs/2202.00436]) have also clearly highlighted this as a major issue in natural language-based approaches. We believe this work mitigates those issues, implicitly by design, making it a much more reliable framework for claims regarding casual reasoning abilities.

Q3) COLD provides a closed setup that can act as a suitable testbench to perform a rigorous analysis, in comparison with COPA which is an open-ended causal reasoning benchmark with only 1000 samples. Moreover, the presence of an underlying causal graph in COLD helps facilitate enormous causal queries, making use of causal dependency concepts like d-separation. Regarding capturing the causal relationship between two events, COLD provides a medium to cover all three rungs of the causal ladder, 1) association 2) intervention, and 3) counterfactuals (Pearl et. al, 2018). The relationship does not only come from plausibility but also considers the occurrence of an event causing another event. For example “checking-in luggage” will directly cause events like “collect luggage from luggage belt” to occur.

Q4) It is to be noted that temporal precedence is generally assumed essential for defining causation, and it is one of the most important clues that is used to distinguish causal from other types of associations [Mill, 1898, Hill, 1965, Pearl and Verma, 1995]. For our framework, we also consider the topologically sorted order obtained from the observational graphs and use the temporal order to define the causal query triplets, i.e., the cause events will always precede the effect events. We also perform a statistical perspective as suggested in the question without using any LMs. The first two rows of Table 4, show the results for the same with a detailed explanation for the statistical $\Delta$ computation provided in the appendix. Please refer to our detailed comment on human evaluation in the response to Reviewer iBzV.

Thank you for pointing out some of the important directions.

• The scope of this paper was limited to commonsense knowledge and we agree that the data used is simplistic in nature (easily to reason about by humans), and does not deal with events beyond commonsense knowledge. We also highlight the same in lines 72-75 of the paper, that CCR excludes the causal events that are beyond the reach of commonsense knowledge, for example, does “planting trees” have a direct impact on the “rainy season”? or does “providing free education” improve the “economic condition of the country/state”; does “carpooling” directly impact “air pollution”, etc. It is to be noted that the primary goal of this work is to bridge the gap between real-world causal reasoning performed by humans on a daily basis and provide the underlying causal graph for rigorous analysis as done in symbolic approaches. Performing causal reasoning on a broader scale that requires a Randomized control trial (as suggested in the weakness) is challenging for humans and is highly dependent on the methodology rather than reasoning/intelligence.

• Some recent works like ROCK [https://arxiv.org/abs/2202.00436], show causal commonsense reasoning through the perspective of causal theory. Our work provides a medium to integrate the causal theory perspective (previously used by papers like Corr2Cause and Cladder for symbolic queries that are verbalized in natural language prompts to evaluate LLMs) and causal commonsense reasoning. We believe this work will be the first to consider both and will facilitate research in making LLMs more causal in reasoning.

• We are sorry for the unclear explanation provided in the paper. We would like to clarify that prior arts on similar lines (ROCK[https://arxiv.org/abs/2202.00436], COLA[https://aclanthology.org/2023.acl-long.288/]) have made use of ATE estimations to perform a zero-shot evaluation using language models like RoBERTa on datasets like COPA. We followed a similar perspective to estimate ATE between the events that are further used to perform classification using the obtained Delta estimates, i.e. given two options, if the ATE of option1 is greater than ATE of option2, the causal prediction is made as option1, which is further used to compute the accuracy over the dataset. Algorithm 2 uses the causal estimands $\Delta$ to compare the causal strength between the premise event and the choice events. We consider the causal estimand computed between the premise and the available set of choices and predict the label corresponding to the high $\Delta$ values. For a given causal query from the created causal query triplet dataset $\mathcal{D}$, where each data entry $\mathcal{D}\_i$ corresponds to $(p, c_1, c_2, q, l)$, i.e., premise event, choice 1, choice 2, question and the label respectively. As the task is to predict the more plausible cause/effect of the given premise event, we create two event pairs, $(p, c_1)$ and $(p, c_2)$, and compute the causal estimand $\Delta$ for both the pairs using the temporal or the backdoor scheme (described in Algorithm 3. Note that the order of events given to $\Delta_{\mathcal{M}}$ is in $E_1$ and $E_2$ format, i.e. $\Delta_{\mathcal{M}}(E_1, E_2)$. Using the temporal precedence (highlighted as remark above), the cause event will always precede the effect event temporally. Hence, for a causal query with the question as 'cause', the causal estimand is estimated as $\Delta_{\mathcal{M}}(c_1^i, p^i)$, $\Delta_{\mathcal{M}}(c_2^i, p^i)$ and $\Delta_{\mathcal{M}}(p^i, c_1^i)$, $\Delta_{\mathcal{M}}(p^i, c_2^i)$ when the question is 'effect.' Further, based on the estimated $\Delta_{\mathcal{M}}$ scores, the more plausible cause/effect is predicted.

• The examples provided in Figure 3 include Premise: go to store and buy cake mix. Question: Which of the following is an effect? Choice 1: come home with the ingredients. Choice 2: go to kitchen. In the given example, using counterfactual reasoning, we can say that “going to the kitchen” is possible without going to the market (if the ingredients are already available), however when going to the market, the event of “come home with the ingredients.” will always take place, making it a more plausible effect in the given choices. Similarly, example 2 in the figure is also explainable as going to market has no direct relation with heating the oven. Thank you for pointing this out, we will add a more detailed explanation for better clarity of the presented idea.

Response to Questions/Clarifications:

Q1) The primary insight as shown from the results is the LLMs lack causal commonsense reasoning abilities in general. Moreover, to provide a suitable test for the model’s understanding, we also apply the backdoor criterion and validate if it improves the performance. Interestingly, we observe that applying the backdoor criterion does help improve the performance over the simple temporal prediction. As highlighted in Table 4, we do see a significant boost in the performance when applying the backdoor adjustments. We believe this strengthens the claim of using the causal theory perspective to improve the performance of LLMs as causal reasoners.

Q2) We completely agree that this framework can be extended to perform various other evaluations like treatment effect estimation or causal graph discovery, and incorporating these will enhance the evaluation. However, given the density of the paper, we believe these go beyond the scope of this work. We provide a few future directions on similar lines in the main paper. We thank you for pointing out some additional directions, we’ll add these to the camera-ready version of the paper.

Q3) Please refer to the response to the last point of weakness. We hope considering the counterfactual situation will make the causal query justified. We thank you for pointing this out, we will improve the writing to explain these examples in the camera-ready version of the paper.

We thank you for your detailed and insightful response. Please find the response to the mentioned weakness/clarifications below:

• Lack of Human Performance: We would like to mention that validating human performance is a little challenging due to the nature of the causal reasoning task. The nature of counterfactual reasoning requires the human/algorithm to assume a similar alternate world/universe with only a particular happening or not happening to approximate the causal strength between the events. These imaginations can be expressed in statements as highlighted by Pearl et. al, containing an “if” statement in which the “if” portion is untrue or unrealized, which is known as a counterfactual. The “if” portion of a counterfactual is called the hypothetical condition, or more often, the antecedent, making it challenging (cognitively heavy) to conduct a human evaluation. We agree that human performance may be lower than complete perfection and adding human performance in these tables will make the comparisons more informative. However, the true dependency of the events is coming from the underlying causal graph, making the generated causal queries accurate. Previously, works like Jin et. al have released causal queries (https://huggingface.co/datasets/causalnlp/corr2cause, https://huggingface.co/datasets/causalnlp/CLadder) generated from the underlying causal graph that does require causal inference theory understanding to be purely annotated by humans.

• Presentation and Organization of the paper: We agree that the density of contents in the paper is a little high, and the presentation can be improved by a significant margin. We agree that it would be better to keep the estimation of ATE as a separate section, however, due to limited space, we had to merge the experiments and the results section (Section 4 Experiments and Results). We hope you understand that fitting all the details in the main paper is challenging, we would love to have some feedback if you think any particular section would be better suitable to move in the main paper from the appendix. We will make the suggested changes in the camera-ready version of the paper.

• We agree that one of the limitations of our work is the limited set of activities, we also highlight the same in the limitation section of our paper with suitable reasons in [Line 349-355]. We would like to reiterate that finding general commonsense reasoning aactivities/tasks that are well understood by humans remains challenging in general and it would be good to explore in the future by adding more set of real-world activities. Moreover, creating a causal graph for an activity increases as we move toward more long-term tasks. However, as a general test of causal intelligence, our framework provides a suitable platform to validate the reasoning capabilities more rigorously, which is missing from the current literature. We hope, with the enormous causal queries (simulating/coming close to the mini-turing test), the created dataset will serve as a robust testbench for evaluating causal reasoning in LLMs.

Response to Questions/Clarifications:

Q1) In general, LLMs are found to be sensitive toward prompts, and coming up with a suitable prompt may affect the performance. Previously, various LLM benchmarking papers have performed evaluation over causal datasets like COPA, CausalBench, etc. LLMs are few-shot learners ROCK has benchmarked LLMs using the cloze evaluation or the MCQ evaluation. We consider the latter that was also justified for benchmarking commonsense reasoning activities in general.

Q2) We agree that there are multiple ways of estimating probability predictions. Some of the widely accepted ones include the CLOZE test evaluation and the MCQ-based evaluation. As highlighted in the paper, we use multi-choice question-answering (MCQA) [Robinson and Wingate, 2023]. Robinson and Wingate [2023] highlight the advantages of MCQ-based evaluation over cloze evaluation [Brown et al., 2020] (where the LLMs are expected to generate the entire answer in a cloze test), leading to a significant boost in various tasks, including commonsense-based tasks. We agree that there is a risk of estimating bias in probability that may result in variance in prediction results.

To eliminate the bias in the probability predictions, we consider the averaging (similar to https://openreview.net/forum?id=shr9PXz7T0) that uses multiple permutations to balance out the prediction evaluation. Algorithm 3 in the Appendix depicts the process of computing an unbiased estimate for the causal estimand. The causal strength is computed between two events $E_1$ and $E_2$ where $E_1$ is assumed to be preceding $E_2$ temporally. To make an unbiased estimate based on the provided options, we consider normalizing the obtained probability scores by flipping the options and providing the same query prompt to the Language Model.

where $\phi$ denotes the prompt template as shown in Figure 8 (top) and $\phi_f$ denotes the same prompt with flipped options, Figure 8 (bottom). The overall equation helps normalize the prediction probabilities of the 'Increase' option by using the probabilities of the 'Decrease' option. Finally, these normalized scores are computed for multiple trajectories $t_i$ in the backdoor adjustment scheme to compute the causal estimands $p_\mathcal{M}(E2 \mid do(E_1))$ and $p_\mathcal{M}(E2 \mid do(\neg E_1))$ that help estimate the causal strength $\Delta_{\mathcal{M}}$ between the events $E_1$ and $E_2$.