LeGIT: LLM Guided Intervention Targeting for Online Causal Discovery

ICLR 2026 Conference Submission24359 Authors

Keywords: Large Language Model, Experimental Design, Causal Discovery

TL;DR: We propose a new framework to leverage LLMs for experimental design in online causal discovery.

Abstract: A fundamental challenge in online causal discovery is designing effective experiments by selecting optimal intervention targets. Conventional numerical methods struggle in the early stages when limited interventional data is available, often yielding noisy or misleading selection guidance. In this work, we introduce the Large Language Model Guided Intervention Targeting (LeGIT), a novel collaborative framework that synergizes the vast world knowledge of LLMs with the precision of numerical algorithms. By analyzing the meta-information of the causal system, it proposes highly informative intervention targets, effectively bootstrapping the discovery process to augment existing numerical approaches, while retaining the convergence guarantees. Evaluated across four realistic benchmarks, LeGIT demonstrates significant improvements in performance and robustness over existing methods, and even surpasses humans. This work establishes that LLMs can play a pivotal role in experimental design, offering a scalable and cost-efficient strategy to accelerate causal and scientific discovery.

Primary Area: foundation or frontier models, including LLMs

Submission Number: 24359