Illusory Pattern Perception Drives Spurious Inference in Large Language Modelsedit

Illusory Pattern Perception Drives Spurious Inference in Large Language Models is a paper by Peihua Mai, Zhuoyan Shao, Xinbao Qiao, Meng Zhang, Xinyue Zhou, and Yan Pang, accepted at NeurIPS 2026. It examines how a perceived pattern in a prompt can lead a large language model to make an inference that is insufficiently grounded in the available evidence.

Authorship note: Peihua Mai, Zhuoyan Shao, and Xinbao Qiao contributed equally as co-first authors.

Overviewedit

The paper places illusory pattern perception among the reliability problems of large language models. A prompt can appear to contain a regularity even when that regularity is a poor basis for the requested conclusion. The resulting risk is that a model follows the apparent pattern instead of evaluating the evidence relevant to the question.

Key takeawaysedit

• Apparent regularity is not evidence. A pattern suggested by a prompt should not by itself determine a model's conclusion.
• Reliable inference depends on grounding. The work highlights the need to check whether an answer follows from the available evidence when a prompt invites pattern completion.
• Prompt interpretation is part of LLM reliability. This failure mode connects the way models read an input to the trustworthiness of their output.

Placementedit

This paper belongs to LLM Reliability and Trustworthy AI. It extends Qiao's broader interest in how limited or misleading evidence affects AI systems.