COSA: Context-aware Output-Space Adapter for Test-Time Adaptation in Time Series Forecasting

The paper introduces COSA, a context-aware output-space adapter that directly corrects predictions of a frozen forecasting model using residual adjustments and gating. It resides in the Output-Space Correction Approaches leaf, which contains only two papers including COSA itself. This leaf sits within the broader Test-Time Adaptation Mechanisms branch, indicating a relatively sparse research direction focused on lightweight, prediction-level corrections rather than full model updates. The small sibling count suggests this specific approach—direct output correction with minimal parameter overhead—remains underexplored compared to heavier adaptation strategies.

The taxonomy reveals neighboring leaves such as Input-Space and Full-Model Adaptation, which houses six papers employing dual adapters or full parameter updates, and Foundation Model Adaptation, containing four papers on parameter-efficient tuning of pre-trained models. COSA diverges from these by avoiding input transformations or extensive parameter modifications, instead operating solely in output space. The Continuous Online Adaptation branch, with ten papers on incremental learning and ensembling, represents a related but distinct paradigm emphasizing streaming updates without the frozen-model constraint. COSA's design philosophy aligns more closely with efficiency-focused methods than with detection-driven approaches in the Concept Drift Detection branch.

Among twenty-nine candidates examined across three contributions, none were flagged as clearly refuting COSA's novelty. The core adapter mechanism examined nine candidates with zero refutations, the gating-based residual correction examined ten with none refutable, and the adaptive learning schedule also examined ten with no overlaps found. This suggests that within the limited search scope—top-K semantic matches plus citation expansion—no prior work directly anticipates COSA's combination of output-space correction, context-aware gating, and adaptive learning. However, the search scale is modest, and the sparse Output-Space Correction leaf indicates fewer benchmarks exist for comparison.

Based on the limited literature search, COSA appears to occupy a relatively novel position within test-time adaptation for time series forecasting. The sparse sibling count and absence of refutable candidates among twenty-nine examined papers suggest the specific design—frozen base model with lightweight output correction—has not been extensively explored. Nonetheless, the analysis covers a focused subset of the field, and broader surveys or domain-specific venues may reveal additional related work not captured here.