Black-Mamba introduces a test-time adaptive forecasting architecture based on evidence-gated state tracking under distribution drift. Instead of updating memory after every instantaneous prediction error or surprise signal, it accumulates surprisal over time and triggers an update only when the accumulated evidence suggests a regime change. The authors report competitive or improved forecasting performance across multiple non-stationary benchmarks, while substantially reducing the number of memory updates during inference. The paper combines a base predictor, dynamic memory, mathematical analysis, and biological evidence to motivate selective, event-driven adaptation.
No heat snapshots are available in the last 24 hours.