Causal Reflection with Language Models

Abi Aryan, Zac Yung-Chun Liu

Keywords: Causal Reasoning, LLMs, Explainability, Agentic Systems, Reinforcement Learning

TL;DR: Causal Reflection enables agents to adapt and explain by combining dynamic causal modeling with LLM-based reasoning.

Abstract: Large Language Models (LLMs) and traditional Reinforcement Learning (RL) agents lack robust causal reasoning, often relying on spurious correlations. We introduce Causal Reflection, a framework that moves beyond simple reward optimization to build dynamic causal models of an environment. Our approach features a temporal, action-based causal function that models state, action, time, and perturbation to capture delayed and nonlinear effects. We also define a formal $\texttt{Reflect}$ mechanism that identifies mismatches between predicted and observed outcomes, generating causal hypotheses to revise the agent's internal model. Within this architecture, LLMs are not black-box reasoners but structured interpreters, translating formal causal outputs into natural language explanations. This work lays the theoretical groundwork for agents that can adapt, self-correct, and communicate causal understanding.

Submission Number: 164