Skip to content
InfoResearchPreprintLLM-specific

Condensing Law of Security Information

Published
Record updated
View JSON

Summary

The paper argues that the common assumption that larger LLMs always do better fails for security tasks, because scale lets a model silently repair anomalous inputs before a security component sees them. It formalizes this as a tradeoff between cognitive capability K(W) and security-information retention R(W), and derives bounds that imply a unimodal relationship with model size. The authors call this the condensing law of security information, under which security error is U-shaped and minimized at a finite optimal model size, and report experiments on various real-world LLMs that support it.