Technical note on Sequential Test-Time Adaptation via Martingale-Driven Fisher Prompting

2025-10-04

AI Summary

This paper addresses robust sequential decision-making under covariate shift in streaming data. Methodologically, it proposes the M-FISHER framework, which unifies real-time distributional shift detection and stable online adaptation. It constructs an exponential martingale from non-conformity scores and leverages Ville’s inequality to guarantee bounded false alarm rate at any time; additionally, it employs Fisher-preconditioned parameter updates to perform natural gradient descent on the distribution manifold, ensuring geometric invariance and KL-divergence minimization. Theoretically, it is the first to integrate martingale theory with Fisher information geometry for test-time adaptation, delivering time-uniform statistical guarantees; it also derives an explicit upper bound on detection delay—O(log(1/δ)/Γ)—that scales inversely with the magnitude Γ of distributional discrepancy. Experiments demonstrate its superior detection efficiency and adaptive stability.

Technology Category

• Machine Learning: Adversarial Learning & Robustness
• Reasoning under Uncertainty: Sequential Decision Making
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Systems and Infrastructure for Web, Mobile and WoT: Data management and stream processing for Web, mobile and wireless applications

Abstract

We present a theoretical framework for M-FISHER, a method for sequential distribution shift detection and stable adaptation in streaming data. For detection, we construct an exponential martingale from non-conformity scores and apply Ville's inequality to obtain time-uniform guarantees on false alarm control, ensuring statistical validity at any stopping time. Under sustained shifts, we further bound the expected detection delay as $mathcal{O}(log(1/δ)/Γ)$, where $Γ$ reflects the post-shift information gain, thereby linking detection efficiency to distributional divergence. For adaptation, we show that Fisher-preconditioned updates of prompt parameters implement natural gradient descent on the distributional manifold, yielding locally optimal updates that minimize KL divergence while preserving stability and parameterization invariance. Together, these results establish M-FISHER as a principled approach for robust, anytime-valid detection and geometrically stable adaptation in sequential decision-making under covariate shift.

Problem

• Detects sequential distribution shifts in streaming data with statistical guarantees
• Bounds detection delay using information gain and distribution divergence
• Implements stable adaptation via Fisher-preconditioned natural gradient descent

Innovation

• Martingale-based sequential shift detection with false alarm control
• Fisher-preconditioned prompt updates for stable adaptation
• Time-uniform guarantees for robust anytime-valid decision-making