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