LeCun et al. Paper: LLMs Over-Compress at the Cost of Semantic Richness, Explaining Opacity

ziv_ravid · x · 2026-11-04

Chen Shani and Ravid Shwartz-Ziv reshare their ICLR'26 paper "From Tokens to Thoughts: How LLMs and Humans Trade Compression for Meaning" (with Dan Jurafsky and Yann LeCun), noting they argued LLMs' growing opacity stems from over-compression before the discourse took off.

Using an Information Bottleneck framework to compare human conceptual structure with embeddings from 40+ LLMs:

LLMs broadly align with human category boundaries but fall short on fine-grained semantic distinctions.
Humans keep 'inefficient' representations preserving contextual nuance; LLMs aggressively compress for more optimal information-theoretic compression at the cost of semantic richness.
Surprisingly, encoder models beat much larger decoder models on human alignment, suggesting understanding and generation rely on distinct representational mechanisms.
Training dynamics show two phases: rapid concept formation, then architectural reorganization as semantic processing migrates from deep to mid layers.

The upshot: LLMs optimize for compression, humans for adaptive utility — a fundamental gap between artificial and natural intelligence.

Original post →

LeCun et al. Paper: LLMs Over-Compress at the Cost of Semantic Richness, Explaining Opacity

More from Models