From Tokens to Thoughts: How LLMs and Humans Trade Compression for Meaning — part 3/4

Cleaned-text code points [31513, 47341). Bibliography omitted here; source.md includes it.

Implications. These dynamics demonstrate that the compression-oriented strategy observed in fully trained models develops from the earliest stages of training. The rapid initial alignment followed by efficiency-focused reorganization suggests models are inherently biased toward statistical compression rather than semantic richness. Achieving human-like representations may require not just different final objectives but fundamentally different learning dynamics that actively maintain semantic diversity throughout development.

## 6 Discussion and Conclusion

We investigated how LLMs and humans navigate the compression-meaning trade-off in conceptual representation. Using information-theoretic analysis of 40+ models against classic cognitive benchmarks, we reveal fundamental differences in their representational strategies.

Key Findings. LLMs achieve broad categorical alignment with humans (AMI ≈ \approx 0.55), successfully partitioning semantic space into recognizable categories. However, they fail to capture the internal structure that makes these categories cognitively useful, as typicality correlations remain weak ( ρ < 0.2 \rho<0.2 ) across model families. Most strikingly, when evaluated on the compression-meaning trade-off, LLMs consistently achieve lower (better) ℒ \mathcal{L} scores than human categories, indicating they optimize for statistical efficiency over semantic richness. This pattern holds across architectures, though encoder models surprisingly outperform decoder models in human alignment despite being orders of magnitude smaller. Training dynamics analysis reveals rapid category formation followed by architectural reorganization that shifts semantic processing from deep to mid-network layers, suggesting efficiency optimization continues throughout training.

Implications. These findings challenge the assumption that statistical optimality equals understanding. LLMs excel at their training objective, which is minimizing prediction error, but this drives them toward representations that sacrifice semantic nuances. Encoder models’ superior alignment with human representations questions the current paradigm of unified, scaled decoder models, suggesting that language understanding and generation may require distinct architectures and rely on different processes . Our framework provides quantitative tools for monitoring the compression-meaning balance in future systems.

Conclusions. Our findings reveal an apparent paradox, showing that LLMs are simultaneously better and worse than humans. This occurs because LLMs and humans employ divergent strategies: statistical compression versus semantic richness , likely reflecting different optimization pressures. While LLMs process billions of tokens efficiently, humans enable flexible reasoning and generalization. Progress toward human-like AI may require preserving the apparent “inefficiencies” that support cognitive flexibility. We provide theoretical understanding, practical metrics, and high-quality digital benchmarks to develop more human-aligned representations. We encourage the community to utilize the data and metrics for future research towards making AI more human-like.

## 7 Ethics statement

Our study relies exclusively on publicly available LLMs and digitized datasets from classic cognitive psychology experiments ( Rosch, 1973a ; Rosch, 1975 ; McCloskey and Glucksberg, 1978 ) . No new human subject data was collected, and all benchmark data we release have been properly attributed and curated to preserve research integrity.

We do not foresee privacy, security, or fairness risks arising from our analyses. Our contribution is methodological and theoretical, focusing on representational trade-offs between humans and LLMs. Nevertheless, we acknowledge that insights into model-human divergences could influence how future systems are designed. We caution that optimizing solely for statistical efficiency without considering semantic richness may exacerbate risks of misinterpretation or oversimplification in socially sensitive applications.

We declare no conflicts of interest or external sponsorship that could bias the reported findings.

## 8 Reproducibility statement

We have taken several steps to ensure reproducibility. All digitized human categorization datasets used in our analyses are publicly released in machine-readable form (Appendix B.1). Detailed model specifications, including architectures, scales, and hyperparameters, are provided in Appendix B.2, and we document embedding extraction procedures, pooling strategies, and prompt templates in Appendix B.3-B.4. Full experimental results, including layer-wise analyses, clustering metrics, and training dynamics across checkpoints, are reported in the appendices (B.5-B.14). Our theoretical framework and derivations are described in Section 4, with complete definitions and formulations provided to enable replication. We will release the code for dataset processing, embedding extraction, and evaluation upon acceptance (to preserve anonymity).

## 9 Acknowledgments

We are grateful to our teacher and mentor, the late Prof. Naftali Tishby, for his profound contributions to information theory, as well as for inspiring us with the depth, elegance, and excitement of this field. This work was supported in part by the Koret Foundation grant for Smart Cities and Digital Living.

## Appendix A Cognitive Intuition

### A.1 A Cognitive Intuition of the compression-meaning tradeoff

The compression-meaning tradeoff refers to the cognitive tension between representing concepts with maximal efficiency (i.e., minimal information) and preserving the semantic richness needed for flexible generalization, inference, and communication. For instance, upon hearing the sentence “There was a large brown Labrador barking loudly near the playground,” a person will often encode a simplified memory, such as “big scary dog near kids.” This is not to suggest that humans cannot recall the full sentence, but rather that we typically retain the most meaningful elements to enable efficient reasoning and generalization; without such abstraction and simplification, leveraging past experience for learning and prediction would be more difficult.

We acknowledge that our metrics, while capturing the information-compression trade-off and geometric efficiency in embedding space, do not directly measure all aspects of human-style conceptual abstraction or reasoning. Our aim is not to claim full equivalence with human cognition (nor do we think anyone can or should make such claims), but rather to provide a quantitative, interpretable proxy that highlights where LLMs and humans converge or diverge in how they compress and organize semantic information. We view these metrics as one lens among many, and we are careful in the paper to frame our findings as providing insights into human-like patterns rather than definitive evidence of human-style conceptual processing.

### A.2 Cognitively-Inspired Inductive Biases

Future models could more closely align with human conceptual structure by incorporating cognitively motivated inductive biases and representational mechanisms. Hierarchical and compositional structure could enable models to capture nested relationships between categories, reflecting the way humans organize knowledge from superordinate to basic-level concepts (e.g., animal → mammal → dog → Labrador). Feature- and relation-based biases could help models focus on meaningful perceptual or functional attributes and relational patterns, rather than relying solely on statistical co-occurrence.

Additionally, theory- or causally grounded priors that draw on humans’ intuitive understanding of how objects interact or behave, could constrain learning in complex domains and support more flexible generalization. Incorporating a hybrid exemplar-rule approach, combining memory of specific examples with abstracted rules, would further approximate human category learning. Modular architectures, in which specialized sub-networks handle different aspects of conceptual representation, could enhance generalization and reduce interference between unrelated features. Finally, meta-learned priors distilled from symbolic or program-based representations offer a way to embed structured, human-like concept hypotheses directly into neural models, allowing them to generalize more like humans across novel situations.

Together, these inductive biases offer a path for models that not only compress information efficiently but also organize knowledge in a manner that mirrors human conceptual richness, capturing graded typicality, family resemblance, and hierarchical relationships. While integrating these biases may trade some compression for fidelity, they provide an exciting opportunity to reduce meaning distortion and bridge the gap between statistical efficiency and human-like conceptual understanding.

## Appendix B Limitations

While this study offers valuable insights, several limitations should be considered. • Our analysis primarily focuses on English; generalizability across languages with different structures is an open question.

• Human categorization data as a benchmark may not fully capture cognitive complexity and could introduce biases.

• Our IB-RDT objective is applied to specific LLMs; other models or representations might behave differently.

• Our analysis is limited to textual input and does not explore image-based representations.

Future work could address these by expanding to other languages, exploring alternative cognitive models, and testing these principles on different architectures or in real-world applications.

## Appendix C Dataset Access Details

The aggregated and digitized human categorization datasets from Rosch (1973a) ; Rosch (1975) ; McCloskey and Glucksberg (1978) are made available in CSV format at: https://huggingface.co/datasets/CShani/human-concepts .

## Appendix D LLM Details

• BERT family: deberta-large, bert-large-uncased, roberta-large ( Devlin et al., 2019 ; He et al., 2021 ; Zhuang et al., 2021 ) .

• QWEN family: qwen2-0.5b, qwen2.5-0.5b, qwen1.5-0.5b, qwen2.5-1.5b, qwen2-1.5b, qwen1.5-1.5b, qwen1.5-4b, qwen2.5-4b, qwen2-7b, qwen1.5-14b, qwen1.5-32b, qwen1.5-72b, qwen2.5-72b ( Bai et al., 2023 ; Yang et al., 2024 ) .

• Llama family: llama-3.2-1b, llama-3.1-8b, llama-3-8b, llama-3-70b, llama-3.1-70b ( Touvron et al., 2023a ; Touvron et al., 2023b ; Grattafiori et al., 2024 ) .

• Phi family: phi-1.5, phi-1, phi-2, phi-4 ( Javaheripi et al., 2023 ; Abdin et al., 2024 ; Abouelenin et al., 2025 ) .

• Gemma family: gemma-2b, gemma-2-2b, gemma-7b, gemma-2-9b ( Team et al., 2024 ; Team et al., 2025 ) .

• Mistral family: mistral-7b-v0.3 ( Jiang et al., 2023 ) .

• GPT family: gpt2, gpt2-medium ( Radford et al., 2019 ) .

• DeepSeek family: DeepSeek-R1-Distill-Qwen-1.5B, DeepSeek-R1-Distill-Qwen-7B, DeepSeek-R1-Distill-Qwen-14B, DeepSeek-R1-Distill-Qwen-32B, DeepSeek-R1-Distill-Llama-8B ( DeepSeek-AI, 2025 ) .

• OLMo family: Olmo-7b ( olmo2024 ) .

• ViT family: Clip ViT-B/32, Clip ViT-B/16 ( Radford et al., 2021 ) .

• Classic static embeddings: GloVe, Word2Vec ( Pennington et al., 2014 ; Mikolov et al., 2013a ; Mikolov et al., 2013b ) .

## Appendix E Contextual Prompts and Pooling Strategies

Contextual embeddings of LLMs require feeding words into the model through a prompt. Because tokenizers often split a word into multiple tokens, and since some items in our datasets consist of two or more words, we face a design choice regarding how to aggregate token representations. In our methodology, we adopt average pooling over the actual tokens, ensuring that all subword pieces contribute equally. Figures 4 and 5 reveal that the average pooling strategy achieves consistent performance and demonstrates the tightest distribution, making it the most reliable choice for our research.

For prompts, we selected a neutral template, "This is a {word}. " (with a trailing space), designed to minimize any additional semantic bias on the target item. Figures 6 and 7 show this prompt to balance performance and consistency, making it ideal for baseline comparisons.

In this section, we explore alternative pooling strategies and evaluate a diverse set of prompt templates across multiple models.

##### Pooling Strategies.

We compare four common approaches: • Avg - mean over all tokens representing the word

• First - representation of the first token

• Last - representation of the last token

• Sentence Avg - mean over the full sentence embedding

##### Prompt Templates.

To test robustness, we design eight templates spanning different linguistic framings:

• "This is a {word}."

• "This is a {word}. " (with trailing space)

• "The concept of {word} is"

• "When we think of {word}, we consider"

• "A typical {word} would be"

• "Examples of {word} include"

• "The category {word} contains"

• "One kind of {word} is"

##### Models.

We evaluate eight representative LLMs covering major architectures: • bert-large-uncased (BERT family)

• deberta-large (DeBERTa family)

• gemma-2-2b (Gemma family)

• Llama-3.2-1B (Llama family)

• Mistral-7B-v0.3 (Mistral family)

• phi-2 (Phi family)

• Qwen2.5-1.5B (Qwen family)

• roberta-large (RoBERTa family)

## Appendix F Static vs. Contextual AMI Exploration

To understand how LLMs develop conceptual alignment with human categories, we examine the progression from static to contextual embeddings. Figure 8 presents three complementary views of this progression across different model scales and architectures.

The left subplot shows Static AMI scores, which represent the conceptual alignment achieved by models’ input embeddings before any contextual processing (i.e., the E matrix embeddings of the target word). These scores reveal that even at the most basic level, LLMs encode semantic information that supports human-like categorical grouping. Remarkably, static models like Word2Vec and GloVe achieve static AMI scores that rival the peak contextual performance of modern LLMs, suggesting that fundamental conceptual structure is captured early in the learning process.

The middle subplot displays Average AMI across all layers, providing a measure of overall semantic representation quality throughout the network. This metric shows the typical performance a model achieves across its entire depth, offering insight into how consistently different layers maintain conceptual alignment. The improvement from static to average AMI demonstrates that contextual processing generally enhances rather than diminishes semantic understanding.

The right subplot reveals Peak AMI , representing the optimal conceptual alignment achieved by any single layer. This metric identifies where in the network conceptual understanding is maximized, typically occurring in middle-to-late layers before declining in the final layers. The progression from static to peak AMI shows that contextual processing not only preserves but significantly enhances the conceptual alignment present in static embeddings.

Several key insights emerge from this multi-metric analysis. First, all models demonstrate above-chance alignment even in their static embeddings, confirming that basic semantic structure is a fundamental property of learned representations. Second, the consistent improvement from static to peak AMI across all model types suggests that contextual processing universally enhances conceptual understanding rather than creating it de novo. Third, encoder architectures of different types (BERT, ViT encoders, and static models) achieve comparable or superior performance to much larger decoder models, highlighting that architectural factors and pre-training objectives significantly influence conceptual alignment quality beyond mere model scale.