Expert Roundtable on Causal Inference and Large Language Models: Opportunities and Challenges

Introduction : The article presents a curated round‑table discussion on causal inference, focusing on the opportunities and challenges that arise when large language models (LLMs) are applied to causal reasoning tasks.

Q1 – How did the experts become involved with causal inference? The panelists share personal stories: a PhD student discovered causal relationships beyond correlation, a researcher at Didi connected causal methods to user growth and incentive projects, another at Didi AI Lab transitioned from time‑series forecasting to causal work, and a Huawei expert linked causal learning to recommendation systems and counterfactual learning.

Q2 – What development opportunities do LLMs bring to causal inference? The experts note that LLMs can reduce the need for multiple specialized models, lower human and time costs, and handle unstructured data, thereby enriching causal discovery. They also highlight LLMs’ potential as knowledge bases for counterfactual reasoning, semantic understanding, and as convenient interfaces (e.g., ChatGPT). However, concerns remain about bias, the lack of true causal understanding, and the need for domain‑specific benchmarks and knowledge‑graph augmentation.

Q3 – What are the experts’ future plans with the emergence of LLM technology? Responses emphasize integrating causal techniques to improve LLM reliability, using causal methods for model correction, bias mitigation, and enhancing generalization. They discuss building benchmarks to evaluate causal capabilities, combining domain‑specific causal data with LLMs, and exploring how causal reasoning can make LLMs more robust and trustworthy.

Q4 – Current status and needs of causal inference across industries? Practitioners describe applications in finance (cash incentives, risk‑controlled decision making), recommendation systems (using counterfactual samples), and operational monitoring (hardware fault detection). Challenges include sparse data, confounding bias, lack of observable variables, and the difficulty of validating causal assumptions in real‑world tasks.

Q5 – Advice for practitioners The panel advises newcomers to solidify fundamentals, study seminal papers, identify clear decision‑making problems where causal inference adds value, and continuously experiment while learning from failures. They stress the importance of interdisciplinary knowledge (statistics, economics, computer science) and staying updated on emerging techniques.

Expert Profiles : Brief bios of the four interviewees – Dong Zhenhua (Huawei Noah’s Ark Lab, expert in recommendation, search, and causal reasoning), Kuang Kun (Zhejiang University, AI professor focusing on causal inference), Wan Shixiang (Du Xiaoman, algorithm engineer with industry causal experience), and Zheng Jia (Tencent OVBU researcher working on incentive and growth algorithms).