Continual adaptation · 2026

PACT

Identity-Preserving Domain-Incremental Adaptation of Frozen Time-Series Foundation Models

A frozen-host adaptation method that corrects distribution shift only where it appears in the embedding space, while retaining an explicit path through the original model.

Piecewise Adaptation via Chart Transport · Manuscript under anonymous review

See how it works View evaluated results

0.958 average macro-F1 on four sequential bearing conditions with frozen Sundial

<0.05 relative forgetting on every one of the five evaluated streams

5% labeled calibration budget per arriving FD, HHAR, and Sleep-EDF domain

Motivation

A new operating condition should not rewrite a good representation.

A time-series foundation model may be trained under one machine load, person, sensor placement, or sleep cohort, then encounter new conditions sequentially. Only a small labeled calibration set may be available, and the condition identity may be unknown when the system makes a prediction.

Fine-tuning shared parameters can recover the new condition but damage earlier ones. Strong retention constraints reduce that damage but can prevent useful adaptation. PACT starts from a different observation: condition shift is heterogeneous across the frozen embedding space. Some local regions move substantially; others are already on the correct side of the classifier boundary and should remain untouched.

Method

An accumulating atlas around an immutable host.

Freeze the host

The pretrained encoder and the source-trained classifier head never change. Source embeddings are standardized once and partitioned into nearest-centroid charts.

Fit local transports

For each arriving condition, PACT learns an identity-initialized rank-8 residual transport per chart using the frozen head and the small labeled calibration set.

Freeze and append

The new transport bank is added to the atlas. Earlier transports, the encoder, and the classifier stay frozen, preventing later domains from overwriting them.

Route without an ID

A chart-specific gate softly mixes the source identity branch with all learned domain transports from the embedding alone. Test-time condition labels are not required.

PACT also derives a margin-based preservation condition: if the gated residual displacement is smaller than the frozen classifier's margin allowance, the original prediction cannot change. Identity routing reproduces it exactly.

Performance

The strongest retention-adaptation balance in the comparison.

5 / 5

Best final average macro-F1

With frozen Sundial embeddings, PACT leads every evaluated dataset after the full domain stream.

+2.4 pts

Average gain over the strongest alternative

Dataset-level gains range from 1.0 point on HHAR to 4.8 points on fault diagnosis.

0.982

Fault diagnosis with 5% labels

In the label-efficiency study with frozen Mantis, PACT reaches 0.982 AMF1 on FD and 0.923 on HHAR.

Sequential comparison with a frozen Sundial backbone, averaged over four domain orders. Higher AMF1 and lower RF are better.
Dataset	Strongest non-PACT AMF1	PACT AMF1	PACT relative forgetting
FD	0.9096	0.9575	0.0022
WISDM	0.8629	0.8774	0.0292
HHAR	0.8551	0.8648	0.0171
UCI-HAR	0.7862	0.8169	0.0469
Sleep-EDF	0.6357	0.6508	0.0229

FD uses four operating conditions; the other datasets use nine-domain streams. The strongest comparison method varies by dataset, most often embedding replay.

What the results show

Routing, not only low-rank adaptation, protects earlier conditions.

Ungated local transports improve plasticity but still apply unnecessary corrections. The learned soft gate produces the best average macro-F1 and the lowest performance-weighted forgetting among the adaptive variants on all five datasets.

Rank 8 matches a full rank-768 map within 0.002 on HHAR and UCI-HAR and performs better on the other three datasets. Across eight frozen backbones, the mechanism transfers unchanged, although final quality still depends on how separable the host representation is.

Growing memoryEach new domain adds a transport bank, and gate refitting retains source and calibration embeddings, though not raw time series.
Overlapping domainsRouting remains harder on HHAR and UCI-HAR, where oracle-domain routing leaves a visible performance gap.
Hard chart assignmentNearest-centroid chart selection can be brittle near chart boundaries; bounded-memory routing and soft chart assignment are natural next steps.