ActMap Enables Zero-Overhead Uncertainty Quantification in LLMs
September 11, 2026
ActMap compresses hidden-state trajectories into a compact 96 KiB tensor during a single generation pass. This white-box representation allows for real-time uncertainty quantification without the latency of multiple sampling passes or heavy computation.
HOW THIS AFFECTS YOU
●
builderYou can implement real-time trust scores for LLM outputs without increasing inference latency.
●
researcherThis provides a structured, fixed-size method to probe transformer internal representations for reliability.