Perceptual Latency Choreography in Probabilistic Generative Interfaces
Investigating how progressive token streaming, speculative execution previews, and skeleton choreography mask inferential latency and foster user trust in generative UI systems.
As generative AI transitions from terminal chat prompts into embedded direct-manipulation software, system latency remains the primary friction point between user intent and computational execution. This paper investigates perceptual latency choreography - a design framework combining speculative interface states, progressive semantic chunking, and tactile feedback to mask model inference delays.
Through a controlled comparative experiment (N=240), we measured task completion velocity and subjective trust across three latency handling paradigms. Interfaces incorporating predictive layout reservation and staged confidence reveals achieved a 42% decrease in perceived delay and an 88% boost in user willingness to delegate multi-step operations to autonomous agents.
Users do not measure latency in milliseconds; they measure it in continuous forward feedback and conversational momentum.
