Towards A Universally Causal Agent

Qirun Dai, Xiao Liu, Jiawei Zhang, Chenhao Tan

Keywords: Causal Reasoning, Synthetic Data, Reinforcement Learning, Reasoning Generalization

TL;DR: We propose a three-step framework for building a universally causal agent that can naturally discover and utilize causal structures when addressing real-world high-stakes challenges.

Abstract: Causal reasoning is crucial for large language models (LLMs) to support trustworthy decision-making, enabling them to naturally discover and utilize causal structures when addressing real-world challenges. However, existing efforts still fall short of building universally causal agents, largely due to two fundamental bottlenecks: a lack of comprehensive causal training data, and an insufficient understanding of training algorithms that can induce generalizable causal skills. We propose a three-step approach to bridge this gap: (1) a unified training data framework spanning diverse causal levels and task domains; (2) an exploration of online, on-policy, and test-time algorithms to identify the most supervision-efficient learning methods; and (3) a rigorous evaluation in both causality-driven real-world applications and general reasoning tasks, together with an analysis of how causality-centered training systematically reshapes external reasoning patterns and internal representations.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 41