Reflective Causal Agents

Abi Aryan, Zac Yung-Chun Liu, Aaron Childress

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

TL;DR: Teaching Agents to Reflect and Causally Explain their decisions

Abstract: We present the first empirical evaluation of Causal Reflection, a framework that equips agents with causal reasoning, structured self-correction, and explainable decision-making in dynamic environments. Standard reinforcement learning and large language model agents often fail under non-stationary conditions, relying on spurious correlations rather than robust causal models. Building on the theoretical foundations of Causal Reflection which formalizes causality as a temporal function and introduces a Reflect mechanism for hypothesis-driven model revision. We implement Reflective Causal Agents. Across a dynamic benchmark environment, these agents outperform ablated and associative baselines in adaptability, predictive accuracy, causal graph recovery, and hypothesis generation. Our results establish Causal Reflection as a practical approach toward robust, interpretable, and generalizable AI systems.

Primary Area: causal reasoning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2026/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 20601