AI Paper Atlas

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

ICLR 2026 · Large language models › Language, safety and interpretability · approach: Linguistic encoding in LLM representations

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 changedWhereMeasuredEffect
use contextual hidden-layer embeddings instead of static input embeddingsLLM embeddings · human categorization benchmarksAMI: 0.45 -> 0.55Improved
use encoder model (BERT-large-uncased, 340M) instead of much larger decodersLLM embeddings · human categorization benchmarksAMI: AMI = 0.60Improved
increase model scaleLLM embeddings · human categorization benchmarksAMI with human categoriesNo clear effect
use peak-AMI layer instead of static layer for typicalityLLM embeddings · human categorization benchmarksSpearman correlation with typicalityHurt
use encoder or static representation models instead of decoder-only modelsLLM embeddings · human categorization benchmarksstatic typicality correlation: rho<0.15 -> rho 0.25-0.40Improved
compare LLM clusters with human categories on the L objectiveLLM k-means clusters · human categorization benchmarksL (lower is better)Improved
use encoder models vs decoder models in complexity-distortionLLM embeddings · human categorization benchmarksdistortion at fixed complexityImproved
lower L (more compression-efficient) as predictor of downstream performanceLLM embeddings · MMLUcorrelation with L: r=-0.20, rho=0.51No clear effect
train OLMo-7B longer (1K to 100K steps)OLMo-7B · human categorization benchmarksAMI: near zero -> about 0.45Improved
continue training from 100K to 500K stepsOLMo-7B · human categorization benchmarkssemantic processing layer: layer 29 -> layer 23Mixed
change prompt template or pooling strategy (ablation)LLM embeddings · human categorization benchmarksresultsNo clear effect
change clustering seed (ablation)k-means on LLM embeddings · human categorization benchmarksAMINo clear effect

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy