Research map / Large language models / Language, safety and interpretability
From Tokens to Thoughts: How LLMs and Humans Trade Compression for Meaning
Chen Shani, Liron Soffer, Dan Jurafsky, Yann LeCun, Ravid Shwartz-Ziv
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use contextual hidden-layer embeddings instead of static input embeddings | LLM embeddings · human categorization benchmarks | AMI: 0.45 -> 0.55 | Improved |
| use encoder model (BERT-large-uncased, 340M) instead of much larger decoders | LLM embeddings · human categorization benchmarks | AMI: AMI = 0.60 | Improved |
| increase model scale | LLM embeddings · human categorization benchmarks | AMI with human categories | No clear effect |
| use peak-AMI layer instead of static layer for typicality | LLM embeddings · human categorization benchmarks | Spearman correlation with typicality | Hurt |
| use encoder or static representation models instead of decoder-only models | LLM embeddings · human categorization benchmarks | static typicality correlation: rho<0.15 -> rho 0.25-0.40 | Improved |
| compare LLM clusters with human categories on the L objective | LLM k-means clusters · human categorization benchmarks | L (lower is better) | Improved |
| use encoder models vs decoder models in complexity-distortion | LLM embeddings · human categorization benchmarks | distortion at fixed complexity | Improved |
| lower L (more compression-efficient) as predictor of downstream performance | LLM embeddings · MMLU | correlation with L: r=-0.20, rho=0.51 | No clear effect |
| train OLMo-7B longer (1K to 100K steps) | OLMo-7B · human categorization benchmarks | AMI: near zero -> about 0.45 | Improved |
| continue training from 100K to 500K steps | OLMo-7B · human categorization benchmarks | semantic processing layer: layer 29 -> layer 23 | Mixed |
| change prompt template or pooling strategy (ablation) | LLM embeddings · human categorization benchmarks | results | No clear effect |
| change clustering seed (ablation) | k-means on LLM embeddings · human categorization benchmarks | AMI | No clear effect |
Closest papers by meaning
- Understanding LLM Behaviors via Compression: Data Generation, Knowledge Acquisition and Scaling Laws (NeurIPS 2025 · similarity 0.90)
- When Less is More: The LLM Scaling Paradox in Context Compression (NeurIPS 2026 · similarity 0.89)
- Can Compressed LLMs Truly Act? An Empirical Evaluation of Agentic Capabilities in LLM Compression (ICML 2025 · similarity 0.88)
- The Grounding Gap: How LLMs Anchor the Meaning of Abstract Concepts Differently from Humans (NeurIPS 2026 · similarity 0.88)
- Leveraging Attention to Effectively Compress Prompts for Long-Context LLMs (AAAI 2025 · similarity 0.88)
- Large Language Models Think Too Fast To Explore Effectively (NeurIPS 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
