EVENT-TRIGGERED BUDGETED MICRO-ADAPTATION IN STREAMING MULTIMODAL FORECASTING

In streaming multimodal forecasting, concept drift often manifests as a change in the relative relevance of modalities [1, 2]. Continuous retraining is computationally expensive and can be unreliable under transient noise. This paper proposes an event-triggered budgeted micro-adaptation strategy for multimodal fusion [5]. We introduce a drift trigger based on fusion-weight optimality over an adaptive window [3] and a micro-adaptation procedure that updates only the fusion layer parameters using a fixed post-drift sample budget. Experimental results show that this approach recovers post-drift accuracy comparable to full retraining (MAE 0.66 vs 0.68) while requiring 8× fewer updates and 20× fewer samples than periodic strategies.