Research map / Large language models / Agents, code and math

Illusory Pattern Perception Drives Spurious Inference in Large Language Models

NeurIPS 2026 · Large language models › Agents, code and math · approach: Mathematical and structural reasoning benchmarks

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Adapts three human illusory-pattern tasks (community correlation, investment correlation, conspiracy belief) to nine LLMs, finds stronger illusion than humans, and uses SAE feature-shift analysis plus activation steering to link the effect to holistic frequency perception and analytic orientation.

Problem

Whether LLMs exhibit illusory pattern perception, i.e., inferring meaningful relationships (group-attribute associations or causal links) from random or insufficient evidence, and how this compares with humans.

Approach

Adapts classic psychology paradigms (illusory correlation, pseudocontingency, conspiracy belief) with human baselines and a neutral-baseline hypothesis test, then uses SAE feature shift between original and correlation-neutral (reminder-prompted) prompts to find internal features linked to the illusion and steers them to reduce it.

Key results

• Average LLM minority choice rate is 0.15 in community correlation vs 0.46 for humans, and 0.31 in investment correlation vs 0.40 for humans.
• Average LLM conspiracy belief rating is 5.37 vs 4.34 for humans (about 23.7% higher on average).
• Signal adjustment on Llama-8B-IT raised community minority-choice rate from 0.17 to 0.35 and lowered conspiracy rating from 5.94 to 4.48, with roughly 5% quality degradation.

Surprising findings

• GPT-5.2 and GPT-4o show stronger conspiracy belief ratings than humans, and Chain-of-Thought prompting only partially mitigates the illusion.
• Under controlled tests, models still select the majority group in MajorLow settings, showing greater representation can override higher positive rate.

Limitations

• Human-only collection via Prolific with 91-99 valid participants per task limits generalizability.
• SAE analysis restricted to three open-source models with available pretrained SAEs and a single layer each.
• Feature categories assigned by two human experts with GPT-4o suggestions, introducing subjectivity.

Experimental setup

Datasets: Human responses collected via Prolific (100 per task)

Benchmarks: community correlation task investment correlation task conspiracy belief task

Metrics: minority choice proportion mean conspiracy rating on 1-7 scale one-sided hypothesis test vs neutral baseline percentage change in SAE feature activation

Baselines: Human participants (Prolific) GPT-3.5 GPT-4o GPT-5 GPT-5.2 DeepSeek-R1-Distill-Llama-8B Llama-8B-IT Gemma-2-9b-IT Qwen2.5-14B-IT Llama-4-Scout-17B-16E-IT

Builds on: illusory correlation pseudocontingency theory sparse autoencoders chain-of-thought

Closest papers by meaning

• Failing to Falsify: Evaluating and Mitigating Confirmation Bias in Language Models (NeurIPS 2026 · similarity 0.87)
• LLMs Know More Than They Show: On the Intrinsic Representation of LLM Hallucinations (ICLR 2025 · similarity 0.86)
• Phenomenal Yet Puzzling: Testing Inductive Reasoning Capabilities of Language Models with Hypothesis Refinement (ICLR 2024 · similarity 0.85)
• Turning large language models into cognitive models (ICLR 2024 · similarity 0.85)
• Generating Computational Cognitive models using Large Language Models (NeurIPS 2025 · similarity 0.85)
• LLM Agents Can Be Choice-Supportive Biased Evaluators: An Empirical Study (AAAI 2025 · similarity 0.85)