LLM4Causal: Democratized Causal Tools for Everyone via Large Language Model

We sincerely appreciate your careful reading and insightful questions, and we are honored to hear that you believe the proposed method is novel and useful. The following are our point-by-point responses.

Tool selection. Since the major focus of this work is the LLM’s causal reasoning ability, we only set a default method for each identified task for the end-to-end evaluation experiment, e.g., we assign the doubly robust method for all the identified ATE tasks in experiment 4.1. The main purpose of the experiment is to check whether an LLM has been equipped with causal tool calling and interpretation abilities.

However, we agree that method selection is of great importance for practical use and are currently developing a systematic framework to automatically select the method for each query. Some of the potential future approaches could be building a machine-learning-based classification model for method selection [1,2,3], analyzing data structures to check model assumptions, and interacting with users to allow them to specify their preferred methods. All these potential ways have their own benefits and require time and effort to construct models and compare performance. Thus, we will leave the model selection automation as future research directions and keep the current work focused more on the LLM part.

Typos and overlapping captions. We will fix them in the final version. Thank you for pointing this out.

[1] Zhang, J., Jennings, J., Zhang, C., & Ma, C. (2023). Towards causal foundation model: on duality between causal inference and attention. arXiv preprint arXiv:2310.00809.

[2] Gupta, S., Zhang, C., & Hilmkil, A. (2023). Learned Causal Method Prediction. arXiv preprint arXiv:2311.03989.

[3] Lorch, L., Sussex, S., Rothfuss, J., Krause, A., & Schölkopf, B. (2022). Amortized inference for causal structure learning. Advances in Neural Information Processing Systems, 35, 13104-13118.

Thanks for the insightful comments, please let us know if we misunderstood any of them.

LLM to use tools. Our decision not to use LLMs’ inherent abilities for causal decision-making is due to their reliance on knowledge acquired during pre-training. Without such knowledge, LLMs' numerical analysis capabilities on causal tasks are rather limited [1]. When data for causal inference conflicts with LLMs’ stored knowledge, this intrinsic ability can be inadequate. Moreover, the integration of external tools into LLM frameworks is an actively expanding research area, which motivates our approach. The complexities and challenges of this integration, discussed below, highlight the novelty of our work.

Experiment design. First, identifying the correct type of causal question in STEP1 is challenging for both LLMs and humans. Our model excels in STEP1 due to a clearly defined JSON format for causal problems and fine-tuning on our proposed Causal-Retrieval-Bench data. Similarly, interpreting external tools' results in STEP3 is also non-trivial, as even advanced LLMs like GPT-4 can confuse causation with correlation. The satisfactory results in Table 4 are due to our carefully designed interpretation evaluation metrics in Section 4.2 and the Causal-Interpret-Bench data. Thus, our model's achievements in Tables 2, 3, and 4 are non-trivial and not easily predictable.

Parameter size. Indeed, larger LLMs may offer improved capabilities. It is noted that our framework and proposed benchmarks are adaptable to LLMs of any size, not just the Llama-2 7B used in our studies. We selected the 7B model as it was a common choice of prior works and already provided satisfactory accuracy while minimizing computational resources.

Select causal tools. In the experiments, a default causal analysis tool is assigned in step 2 for each causal task to maintain simplicity. While choosing the optimal causal analysis tool can be beneficial, it is beyond our scope and left for future research.

GPT4-turbo usage. For ablation analysis 1 (Table 3), we utilized the function calling feature with self-written schemas (will be provided in the appendix). As for ablation analysis 2 (Table 4), the same prompt as LLM4Causal is fed to the chat completion API for interpretation generation.

[1] Cai, H., Liu, S., & Song, R. (2023). Is Knowledge All Large Language Models Needed for Causal Reasoning?. arXiv preprint arXiv:2401.00139.

We sincerely appreciate your careful reading and insightful comments, and we are honored to hear that you believe the proposed method is effective. The following are our point-by-point responses to your comments.

Individual task handling. Thank you for the constructive suggestion. While it is possible to handle each task individually by fine-tuning separate LLM frameworks, our approach is driven by the needs of a general audience who may lack a background in causal analysis. Determining the specific causal task they aim to perform is a significant challenge for these users, and thus a core component of our proposed framework is using a fine-tuned LLM to understand the user's intention and classify their query into a specific causal task. This classification step then helps identify the appropriate causal functions to be used. Therefore, we believe that fine-tuning for each causal task individually is not suitable. We will include additional discussion in the final version to highlight the importance and motivation for integrating various causal tasks.

Manual consumption. Thank you for another valuable comment. First of all, it is important to note that users can directly use our fine-tuned LLM4Causal for causal analysis without any additional manual effort. Also, for researchers looking to fine-tune other base LLMs, our two well-prepared instruction-tuning datasets can be used directly to enhance any other user-selected base LLM's capability in causal task classification and interpretation, without the need for additional human annotation. Moreover, we agree that manual efforts to prepare a golden dataset for fine-tuning are unavoidable for ensuring dataset quality, as is common in existing literature (e.g., [1,2]). To minimize the manual efforts needed for preparing the JSON data, we proposed an output-first strategy (detailed in Section 3.1.1). This strategy effectively ensures the diversity and accuracy of the generated JSON and significantly reduces the amount of human evaluation required, limiting it to evaluating the corresponding GPT-4-generated causal queries for the Causal-Retrieval-Bench dataset.

[1] Zhang, J. (2023). Graph-toolformer: To empower llms with graph reasoning ability via prompt augmented by chatgpt. arXiv preprint arXiv:2304.11116.

[2] Qin, Y., Liang, S., Ye, Y., Zhu, K., Yan, L., Lu, Y., ... & Sun, M. (2023). Toolllm: Facilitating large language models to master 16000+ real-world apis. arXiv preprint arXiv:2307.16789.

Thanks for the insightful comments and careful reading. We will fix the typos in the final version. Our point-by-point responses to your comments are below:

Model Choices. We have conducted additional experiments using T5-base(220M) and T5-large model(770M) for step 1, under the same setting as we used in the paper. The table below summarizes the accuracy of task classification and entity extraction, showing that smaller models cannot acquire the necessary abilities for our use case.

As shown in empirical results, smaller models are inadequate for our use case, due to larger model’s emergent abilities that may greatly improve the performance [1]. Meanwhile, using a 7B model with fine-tuning is a common practice based on our reference papers [Schick et al., 2023; Qin et al., 2023b; Zhang, 2023].

Overfitting. We want to emphasize that the testing data used in our experiments have varied sentence structures and cover different multi-discipline topics when comparing with the datasets prepared for fine-tuning, as explicitly stated in Section 4. We carefully designed the data generation process (detailed in Section 3.1.1) and implemented prompting techniques such as random demonstration sampling to make sentences further differ from each other. Finally, we want to point out that the scale of our dataset is at the same magnitude as [Qin et al., 2023b and Zhang, 2023] and there are no information leaks between training and inference procedures.

Causality between textual events. LLMs’ causal ability from text events, similar to prompting LLMs on causal link existence, focuses on common sense induction [Rashkin et al., 2018; Wang et al., 2023]. Such internal knowledge is inadequate if the pattern does not match newly arrived data, which motivates us to enhance LLMs' causal ability via external tools. Merging internal (from common sense) and external (from data or specific conditions) causal analysis results would be a valuable future direction. We will include more discussion about the related references in the final version.

[1] Wei, J., Tay, Y., Bommasani, R., Raffel, C., Zoph, B., Borgeaud, S., ... & Fedus, W. (2022). Emergent abilities of large language models. arXiv preprint arXiv:2206.07682.