IC-109GPT-3.5-turbo and GPT-4 produce cycles in inferred causal graphs when using pairwise prompts, with cycle counts growing sharply on larger graphs

Aniket Vashishtha, Abbavaram Gowtham Reddy, Abhinav Kumar, Saketh Bachu, Vineeth N. Balasubramanian, Amit Sharma

SourceCausal Order: The Key to Leveraging Imperfect Experts in Causal Inference

When LLMs are queried with pairwise prompts to determine causal direction between variable pairs, the aggregated edge orientations frequently contain cycles, making the causal order undefined. This failure is severe on larger graphs: GPT-3.5-turbo produces >>3k cycles on the 20-node child dataset and >>5k on the 22-node neuropathic dataset. GPT-4, despite being more capable, still produces >>10k cycles on child, >>10k on neuropathic, 684 on alzheimers, and 15 on covid-19. The paper shows that even upgrading from GPT-3.5-turbo to GPT-4 does not eliminate this failure mode.

Evidence

correlational

Key metric

GPT-3.5-turbo pairwise: child >>3k cycles, neuropathic >>5k cycles; GPT-4 pairwise: child >>10k cycles, neuropathic >>10k cycles, alzheimers 684 cycles, covid 15 cycles

Caveat

Cycle counts are lower bounds (computed as total k-length cycles with k=5) because counting all cycles in a DAG is NP-hard.

Model

GPT-3.5 / ChatGPT-3.5 GPT-3.5-turbo, GPT-4 / ChatGPT4 / GPT-4 Code Interpreter / GPT-4 Technical Report

Concepts

Failure mode

Datasets

bnlearn [eval]

Methods

Chain-of-Thought prompting / Chain-of-Thought (CoT) prompting / CoT prompting / Few-shot Chain of Thought / Few-shot CoT / Chain-of-Thought (CoT-S) / CoT-bag prompting / Few-shot CoT prompting / Rationale prompting / Zero-shot CoT prompting / Wei et al. (2022) chain-of-thought prompting [compared-to]

Related work

Kıcıman et al. 2023 [builds-on]

Related findings

IC-110

Extraction

automatic-extraction