OpenNovelty.org

From Tokens to Thoughts: How LLMs and Humans Trade Compression for Meaning

ICLR 2026 Conference SubmissionAnonymous Authors
OpenReview Score: 6.0Download Report PDF
CompressionHuman and Machine CognitionInformation TheoryConcepts
Abstract:

Humans organize knowledge into compact categories that balance compression with semantic meaning preservation. Large Language Models (LLMs) demonstrate striking linguistic abilities, yet whether they achieve this same balance remains unclear. We apply the Information Bottleneck principle to quantitatively compare how LLMs and humans navigate this compression-meaning trade-off. Analyzing embeddings from 40+ LLMs against classic human categorization benchmarks, we uncover three key findings. First, LLMs broadly align with human categories but miss fine-grained semantic distinctions crucial for human understanding. Second, LLMs demonstrate aggressive statistical compression, achieving ``optimal'' information-theoretic efficiency, while humans prioritize contextual richness and adaptive flexibility. Third, encoder models surprisingly outperform decoder models in human alignment, suggesting that generation and understanding rely on distinct mechanisms in current architectures. In addition, training dynamics analysis reveals that conceptual structure develops in distinct phases: rapid initial formation followed by architectural reorganization, with semantic processing migrating from deeper to mid-network layers as models discover more efficient encoding. These divergent strategies, where LLMs optimize for compression and humans for adaptive utility, reveal fundamental differences between artificial and biological intelligence, guiding development toward more human-aligned AI.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

The paper applies the Information Bottleneck principle to quantify compression-meaning trade-offs in LLM versus human conceptual representations, analyzing embeddings from over 40 models against human categorization benchmarks. It resides in the 'Information Bottleneck Principle in Semantic Categorization' leaf, which contains only three papers total. This is a relatively sparse research direction within the broader taxonomy of 21 papers across multiple branches, suggesting the work occupies a focused niche at the intersection of formal information theory and empirical LLM-human comparison.

The taxonomy reveals neighboring work in adjacent leaves: 'Entropy-Based Conceptual Importance Quantification' examines entropy measures in structured representations like AMR graphs, while 'Information-Theoretic Brain-LLM Alignment' uses compression metrics to align neural and model representations. The sibling papers in the same leaf—Evolution Compression Categorization and Color Categories Compression—apply information-theoretic frameworks to human cognitive evolution and perceptual domains respectively. The current work appears to bridge these by directly comparing LLM and human compression strategies across semantic categorization tasks, diverging from purely human-focused or purely neural alignment studies.

Among 28 candidates examined across three contributions, no clearly refuting prior work was identified. The information-theoretic framework contribution examined 10 candidates with zero refutations, the digitized benchmark contribution examined 8 with zero refutations, and the empirical findings on divergent optimization strategies examined 10 with zero refutations. This suggests that within the limited search scope, the specific combination of Information Bottleneck analysis applied to multi-model LLM-human comparison on classic categorization benchmarks appears relatively unexplored, though the analysis does not claim exhaustive coverage of the literature.

Based on the limited search of 28 semantically related candidates, the work appears to occupy a novel position combining formal information-theoretic analysis with large-scale empirical LLM-human comparison. The sparse population of its taxonomy leaf and absence of refuting candidates within the examined scope suggest distinctiveness, though the analysis acknowledges it cannot rule out relevant work outside the top-K semantic matches or citation network examined.

Taxonomy

Core-task Taxonomy Papers
21
3
Claimed Contributions
28
Contribution Candidate Papers Compared
0
Refutable Paper

Research Landscape Overview

Core task: comparing compression-meaning trade-offs in LLM and human conceptual representations. The field structure reflects a multifaceted investigation into how both artificial and biological systems balance efficient encoding with semantic richness. Information-Theoretic Frameworks for Compression-Meaning Analysis anchor the taxonomy, drawing on principles like the information bottleneck to formalize how compression shapes categorization and concept formation, as seen in works like Evolution Compression Categorization[10] and Color Categories Compression[13]. Human-LLM Alignment Studies examine empirical correspondences and divergences between model and human representations, spanning neural alignment (Brain-LLM Alignment[7]), linguistic patterns (Linguistic Patterns Human LLM[2]), and object concept formation (Human-like Object Concepts[5]). Compression Mechanisms in LLM Architectures explore architectural innovations such as adaptive tokenization (AdaTok[14]) and latent memory systems (Latent Memory Reasoning[8]), while Cognitive Mechanisms and Theoretical Frameworks situate these questions within broader theories of semantic representation (Semantic Representation LCCM[19]) and sensorimotor grounding (Sensorimotor Regularities Alignment[20]). Critical Analyses and Limitations, including Six Fallacies[4] and NotebookLM Misalignment[9], interrogate the validity of human-LLM comparisons and highlight persistent gaps. Particularly active lines of work center on whether compression principles that govern human categorization—such as efficient coding under resource constraints—also explain LLM behavior, and whether observed similarities reflect genuine cognitive alignment or superficial pattern matching. Tokens to Thoughts[0] sits within the Information-Theoretic Frameworks branch, specifically addressing the Information Bottleneck Principle in Semantic Categorization. It shares conceptual ground with Evolution Compression Categorization[10], which examines how evolutionary pressures shape compression-meaning trade-offs in human cognition, and Color Categories Compression[13], which applies similar information-theoretic tools to perceptual categorization. Where these neighbors emphasize human cognitive evolution and perceptual domains, Tokens to Thoughts[0] appears to bridge the gap by directly comparing how LLMs and humans navigate compression-meaning trade-offs, potentially offering a unified framework for understanding representational efficiency across biological and artificial systems.

Claimed Contributions

Information-theoretic framework for comparing LLM and human conceptual representations

The authors develop a unified framework based on Rate-Distortion Theory and Information Bottleneck principles to systematically measure and compare how LLMs and humans balance compression efficiency against semantic meaning preservation in conceptual organization. The framework introduces an L objective that combines information-theoretic complexity with geometric distortion to evaluate representational strategies.

10 retrieved papers
Digitized cognitive psychology benchmarks for evaluating conceptual alignment

The authors digitize and publicly release classic human categorization datasets from foundational cognitive science studies, comprising 1,049 items across 34 categories with membership and typicality ratings. These benchmarks provide high-quality empirical grounding for evaluating whether LLMs understand concepts as humans do.

8 retrieved papers
Empirical findings on divergent optimization strategies between LLMs and humans

Through analysis of 40+ LLMs against human benchmarks, the authors reveal that LLMs achieve superior information-theoretic efficiency but miss fine-grained semantic distinctions crucial for human understanding. Encoder models surprisingly outperform decoder models in human alignment, and training dynamics show conceptual structure develops through rapid formation followed by architectural reorganization.

10 retrieved papers

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

Information-theoretic framework for comparing LLM and human conceptual representations

The authors develop a unified framework based on Rate-Distortion Theory and Information Bottleneck principles to systematically measure and compare how LLMs and humans balance compression efficiency against semantic meaning preservation in conceptual organization. The framework introduces an L objective that combines information-theoretic complexity with geometric distortion to evaluate representational strategies.

Contribution

Digitized cognitive psychology benchmarks for evaluating conceptual alignment

The authors digitize and publicly release classic human categorization datasets from foundational cognitive science studies, comprising 1,049 items across 34 categories with membership and typicality ratings. These benchmarks provide high-quality empirical grounding for evaluating whether LLMs understand concepts as humans do.

Contribution

Empirical findings on divergent optimization strategies between LLMs and humans

Through analysis of 40+ LLMs against human benchmarks, the authors reveal that LLMs achieve superior information-theoretic efficiency but miss fine-grained semantic distinctions crucial for human understanding. Encoder models surprisingly outperform decoder models in human alignment, and training dynamics show conceptual structure develops through rapid formation followed by architectural reorganization.

Table of Contents

From Tokens to Thoughts: How LLMs and Humans Trade Compression for Meaning | Novelty Validation