LLM and Human Modes of Representation

Abstract: Much work on the cognitive foundations of AI has focussed on comparisons between the ways in which LLMs and humans process information and represent it. One aspect of this comparison involves determining the extent to which LLMs can achieve or surpass human performance on a variety of cognitively interesting tasks. A second explores points of convergence and divergence between LLM and human systems for processing information. Here, I consider some recent research that has addressed both issues in two informational domains. The first is the representation of linguistic knowledge. The second is real world reasoning and planning. While LLMs frequently achieve impressive levels of performance and fluency on linguistic applications, they tend to handle linguistic content in ways that are distinct from human processing. They are also, for the most part, less efficient than humans in learning and generalisation for reasoning tasks.

• The paper presents a comparative analysis showing that while LLMs excel in large-context pattern matching, they lack human-like abstraction and context-sensitive filtering.
• It reveals that LLMs achieve high correlations in sentence acceptability but consistently apply a compression effect unlike human judgments across modalities.
• The study demonstrates that LLMs can surpass humans in structured tasks through prompt engineering yet lag in holistic narrative coherence and domain-general reasoning.

Comparative Analysis of LLM and Human Information Representation

Introduction

The paper "LLM and Human Modes of Representation" (2606.21616) offers a systematic comparative analysis of the ways LLMs and humans process and represent information across two core domains: linguistic knowledge and real-world reasoning and planning. Drawing from a substantial body of recent experimental literature, the author elucidates both specific parallels and divergences in the representational and computational strategies deployed by LLMs and human agents. This synthesis is particularly relevant for understanding the cognitive limits of LLMs, evaluating claims of artificial general intelligence, and guiding future research in cognitive modeling, computational linguistics, and AI alignment.

Representation of Linguistic Knowledge

Sentence Acceptability Judgments

A key finding reviewed involves sentence acceptability—a nuanced, gradient property in human performance. Early work demonstrated that RNNs predict human acceptability ratings with moderate Pearson correlation ($\sim$0.53) once logprob values are appropriately mapped using scoring functions such as SLOR or PenLP. The transition to first-generation transformers (BERT, XLNET) substantially improved predictivity, with Pearson correlations exceeding 0.70 [Lau et al., 2017, 2020]. Recent multimodal experiments indicate that state-of-the-art LLMs, including ChatGPT-4o, InternVL3-8B, and Qwen2.5-7B, can achieve even higher Spearman correlations (up to 0.89) for human acceptability ratings in null (no-context) conditions—though performance varies with context modality and model architecture.

Notably, a compression effect observed in human ratings—where contextual cognitive load compresses responses towards the scale's mean—appears in LLMs' ratings for textual contexts but is absent from human responses in visual contexts. In contrast, LLMs persistently apply this compression across modalities, highlighting a major divergence: LLMs’ context encoding remains less selective across modalities than that observed in human cognition.

Hierarchical Syntactic Structure and Agreement

For syntactic agreement, experimenting with nested structures reveals both convergence and divergence. LSTMs and transformers can recognize subject-verb agreement dependencies, but performance declines with increasing intervening distractors, paralleling—but not matching—the human pattern of difficulty [Linzen et al., 2016]. However, transformers exhibit a catastrophic drop in accuracy for long, hierarchically nested agreements, performing below chance in contrast to human robustness [Lakretz et al., 2022].

Introducing prompting closes this gap considerably. With even minimal prompt engineering (two to eight shots), models like Chinchilla (7B, 70B) can meet or surpass average human performance on deeply nested agreement cases, outperforming in scenarios with high memory demand. The author interprets this as evidence of LLMs’ superior context-window size and perfect token access, whereas humans employ sophisticated filtering and suppression due to bounded memory resources. This inversion of comparative competence underscores fundamental architectural differences: LLMs leverage pattern matching over large contexts; humans rely on structure-sensitive, memory-constrained heuristics.

Narrative Generation

Human narrative discourse retains higher levels of surprisal and coherence than LLM-generated text. Metrics derived from five coherence dimensions demonstrate that, while models approach or exceed human-level performance in surface features (e.g., referential consistency, visual grounding), aggregate scores for discourse coherence remain consistently higher for humans [Ilinykh et al., 2026]. Human narratives show higher perplexity when assessed by open-source models (e.g., Qwen3-VL, LLAMA 4 SCOUT, InternVL3), suggesting greater unpredictability and original structuring in human discourse. LLM-generated content, while fluent and informative, tends toward verbose, pattern-based descriptions lacking deep narrative connectivity.

Reasoning and Planning Abilities

Natural Language Inference (NLI)

LLMs excel at in-domain NLI, achieving human-equivalent or superior accuracies on familiar test sets. However, performance degrades significantly with adversarial or out-of-domain testing, indicating reliance on pattern recognition rather than contextually adaptive, robust inference [Talman et al., 2021]. Recent benchmarks using scientific NLI tasks confirm this: state-of-the-art RoBERTa achieves an F1 of 77.2, while domain experts score 89.33, and even non-expert humans exceed model performance (F1 = 79.78) [Sadat and Caragea, 2024]. This discrepancy is robust to domain shift, further underscoring the lack of generalizable, human-like semantic reasoning in LLMs.

Complex Image Interpretation

Evaluations on diagnostic neuroradiology quizzes reveal that certain LLMs, notably Claude 3.5 and OpenAI 04-mini-high, can match or outperform some human experts—yet this advantage is largely attributable to pattern matching on textual descriptions rather than genuine visual diagnostic reasoning [Albaqshi et al., 2025; Zhou et al., 2026]. The negligible difference in accuracy between text-only and text+image conditions in models, in contrast to human experts’ explicit use of imagery, further delineates extant limitations in multimodal integration and inferential flexibility.

Planning and NP-Hard Tasks

On complex planning—including NP-hard problems such as graph coloring, knapsack, and traveling salesman—LLM performance lags behind heuristic algorithms when faced with task formulations outside of familiar domains [Duchnowski et al., 2025]. Enriching LLMs with symbolic inference modules (e.g., Python-based ILP solvers) improves accuracy (notably for ChatGPT-4o), but generalization remains weak. The inability to efficiently adapt to re-formulated problems confirms that LLMs do not exhibit the flexible, abstract, and domain-independent problem-solving characteristic of human planners.

Implications and Future Directions

The findings synthesized in this paper have significant ramifications for both AI research and cognitive science:

• Distinct modes of representation: LLMs’ successes in linguistically formal tasks are distinct from the functional linguistic competence found in human cognition. Human performance remains superior in real-world reasoning and planning, while LLMs’ advantages lie in large-context pattern integration and dependency resolution.
• Architectural biases: LLMs’ lack of domain-general reasoning is not a mere consequence of insufficient data but an emergent property of current model architectures—transformer-based systems lack the dynamic, context-sensitive filtering and abstraction mechanisms present in human cognition.
• Narrative originality and coherence: Higher perplexity and unified coherence scores in human-produced narratives highlight persistent limitations in even the most advanced LLMs with regard to creativity, global coherence, and discourse planning.
• Cautions for anthropomorphism and deployment: Fluency in surface structure should not be conflated with deep semantic or pragmatic understanding. Anthropomorphic attributions to LLMs are unfounded given the fundamentally different computational underpinnings and error patterns.

Future research should prioritize:

• Architectural innovations to align LLM context management and memory constraints more closely with human cognitive architectures.
• Mechanisms to enhance narrative surprisal and global coherence in multimodal text generation.
• Improvements in few-shot learning and domain-transfer capabilities for real-world reasoning and planning tasks.

Conclusion

"LLM and Human Modes of Representation" develops a rigorous, evidence-driven stance: LLMs are neither mere stochastic parrots nor instantiations of general intelligence. While they frequently equal or surpass human performance in formal linguistic tasks and exhibit impressive narrative fluency, they rely fundamentally on large-context pattern recognition, distinct from human-typical abstraction and reasoning. Persistent shortfalls in multimodal context selectivity, domain-general reasoning, and narrative coherence invite architectural and conceptual advances at the intersection of computational modeling and cognitive science. The work advocates a more nuanced and empirically grounded perspective on the path toward deep language and reasoning competence in AI systems.