[Paper Review] Mind Scramble: Unveiling Large Language Model Psychology Via Typoglycemia

Miao Yu, Junyuan Mao+8 more|arXiv (Cornell University)|Oct 2, 2024

Artificial Intelligence in EducationComputer Science3 citations

TL;DR

This paper introduces LLM Psychology, a novel research framework that uses typoglycemia—scrambled text where only the first and last letters of words are preserved—to probe the cognitive mechanisms of large language models (LLMs). By applying this psychological paradigm, the authors reveal that LLMs exhibit human-like processing patterns, including context-dependent comprehension and task-specific robustness, while also uncovering unique, consistent 'cognitive fingerprints' across models, with deeper analysis of attention mechanisms explaining these behaviors.

ABSTRACT

Research into the external behaviors and internal mechanisms of large language models (LLMs) has shown promise in addressing complex tasks in the physical world. Studies suggest that powerful LLMs, like GPT-4, are beginning to exhibit human-like cognitive abilities, including planning, reasoning, and reflection. In this paper, we introduce a research line and methodology called LLM Psychology, leveraging human psychology experiments to investigate the cognitive behaviors and mechanisms of LLMs. We migrate the Typoglycemia phenomenon from psychology to explore the "mind" of LLMs. Unlike human brains, which rely on context and word patterns to comprehend scrambled text, LLMs use distinct encoding and decoding processes. Through Typoglycemia experiments at the character, word, and sentence levels, we observe: (I) LLMs demonstrate human-like behaviors on a macro scale, such as lower task accuracy and higher token/time consumption; (II) LLMs exhibit varying robustness to scrambled input, making Typoglycemia a benchmark for model evaluation without new datasets; (III) Different task types have varying impacts, with complex logical tasks (e.g., math) being more challenging in scrambled form; (IV) Each LLM has a unique and consistent "cognitive pattern" across tasks, revealing general mechanisms in its psychology process. We provide an in-depth analysis of hidden layers to explain these phenomena, paving the way for future research in LLM Psychology and deeper interpretability.

Motivation & Objective

• To investigate whether large language models (LLMs) exhibit human-like cognitive processes in text comprehension, beyond superficial statistical patterns.
• To address the lack of systematic, intrinsic cognitive mechanism analysis in LLMs by adapting psychological experiments like typoglycemia.
• To develop a democratized, no-data-needed benchmark for evaluating LLM robustness and cognitive behavior across diverse tasks.
• To uncover model-specific, consistent cognitive patterns across tasks, suggesting a general underlying psychological mechanism in LLMs.
• To provide micro-scale interpretability through hidden layer analysis, linking behavioral phenomena to internal model dynamics.

Proposed method

• Migrate the typoglycemia phenomenon—where humans read scrambled words with intact first and last letters—from human psychology to LLMs as a cognitive probe.
• Design and apply three levels of scrambling: character-level (e.g., replacing inner letters), word-level (reordering letters within words), and sentence-level (reordering words with preserved first/last letters).
• Evaluate LLMs across diverse tasks (e.g., BoolQ, summarization, translation) under original and scrambled inputs to measure accuracy, token usage, and inference time.
• Use controlled typoglycemia patterns (e.g., Char-DEL-INT_1, Char-REO-INT) to systematically vary input perturbations and assess model resilience.
• Perform ablation and attention visualization on hidden layers to explain behavioral differences and identify model-specific cognitive fingerprints.
• Compare LLMs (e.g., GPT-4, Llama-3.1, Gemma-2-2B, Gemma-2-9B) across tasks to reveal consistent, unique cognitive patterns.

Experimental results

Research questions

• RQ1 Do LLMs process scrambled text in a way that mirrors human cognitive mechanisms, particularly context-based word reconstruction?
• RQ2 How does typoglycemia exposure affect LLM performance across different task types, especially complex logical tasks like math or reasoning?
• RQ3 Can LLMs be characterized by unique, consistent cognitive patterns across diverse tasks, suggesting an intrinsic psychological profile?
• RQ4 To what extent do LLMs rely on statistical patterns versus deeper structural understanding when processing typoglycemic inputs?
• RQ5 What internal mechanisms in the model’s hidden layers explain the observed behavioral differences in typoglycemia tasks?

Key findings

• LLMs demonstrate macro-scale human-like behaviors: reduced accuracy and increased token usage and inference time when processing scrambled text, especially in complex tasks.
• Different LLMs show varying robustness to typoglycemia, with GPT-4 and Llama-3.1 maintaining high accuracy even under significant scrambling, indicating model-specific resilience.
• Complex logical tasks (e.g., math, BoolQ) are significantly more challenging for LLMs in scrambled formats, suggesting that such tasks rely more on precise syntactic and semantic structure.
• Despite robustness, some models show counter-intuitive improvements in accuracy on certain scrambled inputs, hinting at data-driven behavior rather than true comprehension.
• Each LLM exhibits a unique, consistent cognitive fingerprint across tasks, indicating a stable internal psychological mechanism underlying its processing style.
• Hidden layer analysis reveals that attention patterns and feature representations shift systematically under typoglycemia, explaining behavioral differences and enabling deeper interpretability of LLM cognition.