#LLM Optimization

Beyond Token Count: Semantic Compression for LLMs

Researchers recast LLM reasoning as lossy compression using the Conditional Information Bottleneck (CIB), employing semantic surprisal for efficient token pruning.