Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces Black-Mamba, a novel test-time adaptive forecasting architecture designed to address the challenges posed by non-stationary conditions in real-world data. By employing evidence-gated state tracking, Black-Mamba selectively updates its dynamic memory only when significant changes in the data distribution are detected, rather than responding to every prediction error. The model demonstrates competitive or superior predictive performance across various benchmarks while significantly minimizing unnecessary memory updates, highlighting the effectiveness of using accumulated surprisal as a signal for adaptation.
Accumulated surprisal can distinguish between persistent distribution shifts and transient noise, leading to more efficient forecasting adaptations.
Forecasting under real-world conditions is inherently non-stationary, as the conditional distribution of future observations evolves over time. Recent test-time adaptive sequence models address this challenge by updating internal states during inference, but tie adaptation to instantaneous prediction errors or surprise. This coupling can conflate persistent distribution shift with stochastic innovations, leading to unnecessary updates and inefficient adaptation. We introduce Black-Mamba, a test-time adaptive forecasting architecture that formulates online adaptation as evidence-gated state tracking under distribution drift. The model augments a base predictor with a dynamic memory updated when temporally accumulated surprisal provides sufficient evidence of a regime change. This turns adaptation into a selective, event-driven process rather than a continuous one. Across multiple forecasting benchmarks with non-stationary dynamics, Black-Mamba achieves competitive or improved predictive performance compared to existing test-time adaptation methods while significantly reducing the number of memory updates during inference. Together with mathematical analysis and biological evidence, these results suggest that accumulated surprisal provides a principled signal for distinguishing persistent drift from transient noise, yielding more efficient and robust adaptation.