Technical Summary: LLM and Human Modes of Representation

Problem Statement

The paper addresses the polarized debate regarding the cognitive foundations of Large Language Models (LLMs). One perspective views LLMs as breakthroughs toward General Artificial Intelligence (GAI), while the opposing view dismisses them as "stochastic parrots" merely reproducing training data. The central problem is to move beyond these extremes by empirically comparing how LLMs and humans process information and represent knowledge. Specifically, the paper investigates whether LLMs achieve human-level performance through similar cognitive mechanisms or through distinct, potentially superficial, processing strategies. The inquiry focuses on two primary domains: the representation of linguistic knowledge and real-world reasoning/planning.

Methodology

The paper synthesizes recent experimental research across six specific tasks, comparing human performance against various LLM architectures (including RNNs, Transformers, and Vision-Language Models). The methodology involves:

1. Sentence Acceptability Ratings:

   • Human Baseline: Utilizing crowd-sourced ratings for grammatical and ungrammatical sentences in null, relevant, and irrelevant textual contexts (Lau et al., 2017; 2020).
   • Visual Context Extension: Extending this to visual contexts using GPT-5 generated images (relevant vs. irrelevant) and testing both human raters and Vision-Language Models (VLMs) like ChatGPT-4o, InternVL3, Qwen2.5, and LLaVA-1.5 (Jang et al., 2026).
   • Metrics: Spearman and Pearson correlations between model outputs (logprobs or generated scores) and human ratings.

2. Hierarchical Syntactic Structure (Subject-Verb Agreement):

   • Task: Evaluating the ability to identify subject-verb agreement in sentences with increasing numbers of intervening distractor NPs and varying depths of embedding.
   • Models: Testing LSTMs, GRUs, CNNs, and Transformers (GPT-2, RoBERTa, Chinchilla) under zero-shot, few-shot (2-shot, 8-shot), and prompted conditions (Linzen et al., 2016; Bernardy & Lappin, 2017; Lakretz et al., 2022; Lampinen, 2024).

3. Narrative Generation:

   • Task: Generating stories based on film image sequences.
   • Metrics: Five quantitative metrics for discourse coherence (coreference continuity, implicit discourse relations, topic change, character persistence, visual grounding) aggregated into an Arithmetic and Geometric Mean Narrative Coherence Score (NCS).
   • Perplexity Analysis: Comparing the perplexity of human-generated stories against model-generated stories and scraped online text (Ilinykh et al., 2026).

4. Reasoning and Planning:

   • Natural Language Inference (NLI): Testing fine-tuned transformers and prompt-based LLMs on in-domain and out-of-domain (subdomain shift) inference tasks, comparing results against human experts (Talman et al., 2019; 2021; Sadat & Caragea, 2024).
   • Complex Image Interpretation: Evaluating diagnostic accuracy on neurological images (CT/MRI) with and without textual descriptions, comparing models (e.g., Claude 3.5, OpenAI 04-mini-high) against human medical experts (Albaqshi et al., 2025; Zhou et al., 2026).
   • Planning: Assessing performance on NP-hard optimization problems (graph coloring, traveling salesman, knapsack) with reformulated natural language specifications to test generalization, comparing LLMs against greedy heuristics and symbolic solvers (Duchnowski et al., 2025).

Key Contributions and Results

1. Linguistic Knowledge Representation

• Acceptability and Context: While LLMs achieve high correlations with human acceptability ratings, they exhibit a "compression effect" in visual contexts that humans do not. Humans filter out irrelevant visual context when judging sentence acceptability, whereas VLMs retain the image as part of the extended context, influencing their processing. This suggests a fundamental difference in modality filtering.
• Syntactic Hierarchy: Humans struggle with agreement tasks as the number of distractors and embedding depth increases due to memory limitations. In contrast, LLMs (particularly with prompting) often surpass human performance on these tasks. The paper attributes this to LLMs' capacity to maintain large context vectors with perfect access to information, whereas humans rely on filtering and suppression.
• Narrative Coherence: Human narratives consistently achieve higher integrated coherence scores and higher perplexity (surprisal) than LLM-generated narratives. While LLMs produce fluent text, they lack the originality and deep cohesion found in human discourse, often resulting in verbose, conjunctively descriptive narratives.

2. Reasoning and Planning

• Inference (NLI): LLMs perform well on in-domain NLI tasks but show significant performance drops in out-of-domain testing. Adversarial testing reveals reliance on surface pattern matching rather than deep lexical semantic reasoning. Even with domain-specific fine-tuning, LLMs lag behind human experts in generalizing across diverse knowledge domains.
• Image Interpretation: Models often achieve high diagnostic accuracy on medical images, but this success is driven by textual descriptions of symptoms rather than visual analysis. When images are removed, performance remains statistically similar, and error analysis indicates failures in diagnostic reasoning rather than image identification.
• Planning: On NP-hard optimization problems with reformulated specifications (out-of-domain), LLM accuracy declines rapidly as task complexity increases. They generally underperform compared to symbolic solvers (ILP) and greedy heuristics, indicating a lack of robust general reasoning capabilities for hard optimization problems.

Significance and Claims

The paper argues that the experimental evidence refutes both extreme views: LLMs are neither "stochastic parrots" incapable of complex tasks nor agents of General Intelligence.

• Distinct Mechanisms: LLMs achieve impressive fluency and often surpass humans in specific linguistic tasks (e.g., agreement in complex structures) but do so through mechanisms distinct from human cognition. They rely on powerful pattern recognition and extensive context retention rather than the sophisticated filtering and suppression mechanisms humans use to manage limited memory.
• Formal vs. Functional Competence: The results support the distinction between "formal linguistic competence" (handling purely linguistic structures) and "functional linguistic competence" (real-world inference and reasoning). LLMs excel at the former but struggle with the latter, particularly when tasks require generalization beyond training distributions.
• Limitations in Reasoning: The paper concludes that LLMs lack deep, domain-general reasoning. Their success in inference and planning is largely dependent on fine-tuning or few-shot prompting within specific domains. When faced with out-of-domain variations or complex optimization, they revert to pattern matching rather than logical deduction.

The paper posits that future research should focus on constraining LLM training and processing to better converge with human linguistic behavior, modifying designs to improve narrative coherence and surprisal, and enhancing few-shot training for domain transfer in reasoning tasks. These efforts would not only improve AI utility but also illuminate the nature of human language acquisition and representation.