Frozen time-series foundation models can follow recurring conditions: PACT edits only embeddings and holds macro-F1 above 0.95 on a four-condition bearing stream
Synopsis
When operating conditions recur, adapting a time-series foundation model by global fine-tuning changes parameters shared across conditions and degrades earlier ones. PACT partitions the frozen source embedding space into local regions, learns a lightweight residual correction only in the regions affected by each new condition, and keeps an explicit identity path that a learned gate softly selects among at inference. On a four-condition bearing-fault stream, using 5% labeled calibration data per arriving condition and no condition identity at test time, average macro-F1 stays above 0.95.
Interpretation
A time-series foundation model facing recurring condition shifts can correct only the affected local regions of its embedding space, leaving the pretrained encoder and classifier head untouched. Prior routes either fine-tuned shared parameters, which sacrifices earlier conditions, or constrained changes by regularization, trading adaptation against retention; condition-specific parameter methods typically rely on hard routing or condition identity. On the frozen Sundial backbone, PACT reaches the highest final average macro-F1 on FD, WISDM, HHAR and UCI-HAR, exceeding the strongest baseline shown by 11.15 and 8.34 percentage points on FD and WISDM, leading by 0.71 points on HHAR, and trailing LwF by 0.09 points on Sleep-EDF.
The correction acts on embeddings before the classifier head, while the encoder, head and previously learned corrections stay frozen, so a new condition cannot overwrite them. For each arriving condition, a low-rank residual correction is learned within each local region of the source embedding space, defined as Voronoi cells around source-embedding k-means centroids, and is initialized to zero, which is the identity. The calibration budget is 5% of the arriving condition's training split on FD, HHAR and Sleep-EDF, up to 450 examples on WISDM and up to 100 on UCI-HAR; at inference no condition identity is given, and a two-layer MLP gate per region applies a softmax weighting over the identity path and accumulated corrections.
The identity path means an earlier prediction can change only when the gate routes its embedding through a non-identity correction, so retention has an analyzable condition. PACT freezes previously learned corrections and updates only their routing, preserving their parameters, but gate refitting can still change predictions on earlier conditions, so unchanged predictions are not guaranteed. On FD, Proposition 1 certifies preservation for 889 of 891 initially correct source predictions, and all 889 certified predictions stayed correct; certificate coverage is 97.5% on HHAR and 71.1% on UCI-HAR, with all certified predictions preserved.
Perspective
The demonstrated setting is time-series classification streams with a homogeneous task and label space and recurring conditions: four FD operating conditions, nine HHAR users, and nine subjects each for WISDM, UCI-HAR and Sleep-EDF, with no condition identity at inference and calibration from a small labeled fraction of each arriving condition's training split.
Preservation is conditional rather than a zero-forgetting guarantee: remaining error comes from routing and the magnitude of the applied correction, and gate refitting can still change predictions on earlier conditions. Overlapping conditions remain hard to separate, and the gaps to oracle routing on HHAR and UCI-HAR indicate room to improve condition assignment; imbalanced or missing calibration classes reduce adaptation quality, with 85% of labels on one class lowering average macro-F1 by 9.08 points on FD.
