acceptodds
Under review as a conference paper at ICLR 2027

Adapt the Data, Not the Model: Input-Space Adaptation for Frozen Time-Series Predictors

Abstract

Conventional unsupervised domain adaptation (UDA) typically relies on updating the predictor's weights to optimize performance on the target domain using a labeled source dataset and unlabeled target data. However, a growing demand for secure deployments, such as closed inference APIs, firmware, or regulatory certification, means models are increasingly released with locked weights. We study domain adaptation in this locked-predictor regime for multivariate time-series models, where the deploying site holds its own labels but source training data may be inaccessible. We introduce PILOT (Per-timestep Input-space adapter for LOcked predicTors), a per-timestep input-space adapter trained against the frozen predictor by back-propagating the task loss using target labels. The framework supports two settings: a source-free variant for strict privacy constraints and a source-assisted variant when data sharing is permitted. We evaluate PILOT across five clinical time-series tasks (encompassing two domain shifts) and five AdaTime sensor benchmarks using a frozen 1D-CNN backbone. PILOT achieves state-of-the-art performance across all comparable frozen-backbone, end-to-end, and test-time adaptation baselines. It yields a +15.3 improvement in AUCPR for AKI, increases the mean Macro-F1 score by +10.1 over the leading test-time adaptation method on AdaTime and matches or exceeds the performance of natively trained target-domain models. Furthermore, training PILOT on a single architecture enables zero-shot transfer to other frozen predictors, allowing the same PILOT module to be reused across different architectures.

Primary area:
learning on time series and dynamical systems
Keywords:
clinical time series, electronic health records, ICU prediction, distribution shift, domain adaptation, frozen models, input-space adaptation, multivariate time series, healthcare machine learning, MIMIC-IV, sensor time series, model deployment

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_ZfNZOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses domain adaptation for multivariate time-series models under a locked-predictor setting, where a deployed model's weights are completely frozen due to regulatory certification or closed inference APIs, but the deploying hospital or site has its own target labels. Instead of updating the predictor's parameters, the authors propose PILOT (Per-timestep Input-space adapter for LOcked predicTors), an upstream adapter that transforms the target input data before feeding it into the strictly frozen predictor. The PILOT architecture uses an encoder-decoder structure built with axial attention blocks (across time and features) and FiLM conditioning for static covariates. It operates in two regimes: a source-assisted setting that uses a k-NN memory bank with cross-attention to retrieve local physiological states from source training data, and a source-free setting for strict privacy constraints where no source data is used. Training follows a two-phase protocol: Phase 1 pretrains the adapter using unmasked identity reconstruction, and Phase 2 optimizes the adapter using target labels by back-propagating the task loss directly through the frozen predictor, combined with auxiliary losses for clinical signal preservation and marginal alignment. The authors evaluate PILOT on five clinical YAIB tasks (three classification: mortality, AKI, and sepsis; two regression: length-of-stay and kidney function), adapting a frozen MIMIC-IV LSTM predictor to eICU and HiRID, as well as five non-clinical AdaTime sensor benchmarks using a frozen 1D-CNN. Across both source-assisted and source-free regimes, PILOT outperforms end-to-end domain adaptation and test-time adaptation baselines, matches natively trained target models, and can transfer zero-shot to other frozen architectures like GRU and TCN.

Critical strengths

I like the core idea of focusing on getting the input space (x) right instead of always trying to modify the model weights (f). This provides a nice perspective that solely chasing model-level generalizability or robustness might not be the only way forward for clinical deployment. Furthermore, the paper does not only test on ICU datasets (including both classification and regression tasks), but also evaluates on the AdaTime sensor benchmarks, which nicely shows that this input-adaptation concept can generalize across different time-series domains.

Critical weaknesses
  • Complex training pipeline and need for clearer component ablations: Learning the input adapter (g_theta) requires a fairly heavy pipeline, including an axial-attention encoder-decoder, FiLM conditioning, an optional k-NN memory bank, a two-phase training protocol, and a combination of six different loss terms. Because there are so many moving parts in both the architecture design and the combined loss function, there should be more thorough ablation experiments in the main text isolating each architectural part and loss term to clearly show which specific component contributes the most to the adaptation performance. Furthermore, even though the downstream predictor is frozen, training the PILOT module itself still requires back-propagating through the frozen model and tuning task-specific hyperparameters. The paper lacks a clear analysis of the trade-off between the compute and effort needed to train g_theta versus the actual performance gain.

  • Incomplete experimental directions and simple frozen backbones: For the ICU experiments in the main text, the authors only show adaptation in one direction (using MIMIC-IV as the source and adapting to eICU or HiRID) rather than testing all source-target pairs. In addition, the frozen predictors used in the experiments are relatively basic architectures like standard LSTMs and 1D-CNNs. Given recent progress in time-series foundation models and state-of-the-art clinical forecasters (such as MIRA from Microsoft), the paper would be much stronger if it showed that PILOT can also adapt modern, large pretrained time-series backbones.

Questions for authors
  • Given that g_theta combines multiple architectural modules and six loss terms, could the authors provide a clearer, step-by-step ablation in the main text showing which specific architecture components and loss terms contribute the most to the adaptation gains?

  • Could the authors provide additional experiments using more modern, state-of-the-art frozen time-series predictors or foundation models (e.g., MIRA or Transformer-based backbones) beyond simple LSTMs and 1D-CNNs? Testing other source-target directions in the ICU benchmarks would also make the evaluation more complete.

  • Could the authors provide a clear trade-off analysis comparing the compute cost and parameter count of training PILOT (g_theta) against the performance gains achieved, as well as against the compute cost of training the frozen predictor from scratch?

  • Figure 2 is currently a bit too abstract and not illustrative enough, especially regarding how the memory bank and k-NN cross-attention retrieval work at each timestep. Could the authors redraw or expand Figure 2 to show this mechanism more clearly?

  • This might be a stretch, but have the authors considered using pretrained LLM-based or foundation embedding models to replace the learned g_theta module, potentially turning this into a completely training-free input adaptation method?

  • Minor wording question: In the abstract and main text, the authors use the term "AUCPR". Do you mean the more standard term "AUPRC" (Area Under the Precision-Recall Curve)?

Reviewer_JpfmOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PILOT, which adapts a frozen time-series predictor by learning an input transformation instead of updating weights. The adapter is a two-phase axial-attention encoder-decoder that rewrites each input window before the frozen predictor consumes it. Its second phase trains on target labels with a task loss back-propagated through the frozen predictor, alongside fidelity, range, MMD and auxiliary-head terms. An optional k-NN source memory bank supplies cross-attention context, and a source-free variant removes every source-tied term while pretraining the first phase on the target cohort. Experiments cover five YAIB clinical tasks with a frozen MIMIC-IV LSTM transferred to eICU and HiRID, and five AdaTime datasets with a frozen one-dimensional CNN. The reported mean classification gain is +4.93+4.93 AUROC, with +15.3+15.3 AUCPR on AKI, and the paper attributes the gain to the through-predictor task gradient rather than marginal alignment.

Critical strengths
  1. Locked-Predictor Setting: The paper separates a labelled-target, frozen-predictor regime from unsupervised adaptation, source-free adaptation, test-time adaptation and parameter-efficient tuning. Figure 1 shows the task gradient passing through the locked predictor into an upstream adapter, and the source-free variant is specified term by term.
  2. Evaluation Breadth: The evaluation spans five clinical tasks across two target cohorts and five sensor benchmarks, with paired bootstrap intervals on the clinical ladder. HiRID transfer deltas exceed the eICU deltas on all three classification tasks, which is a consistent cross-cohort result rather than one favourable cell.
  3. Cross-Architecture Transfer: An adapter trained against an LSTM predictor transfers zero-shot to frozen GRU and TCN predictors in all 16 task-architecture cells. Retention of the LSTM gain is positive everywhere, and the range is wide, so one adapter design does cover several predictor families.
Critical weaknesses
  1. Headline Row Is a Per-Task Composite: The PILOT source-assisted row in Table 1 is not a single configuration but five per-task leaves. The Table 18 caption names the Mortality no-fidelity, AKI no-retrieval, Sepsis no-retrieval and LoS no-MMD cells, and points to a Table 25 source-pretrain leaf for kidney function. Those leaves are scored on the held-out test set that the row reports, and no validation-split ablation or validation-based selection table appears anywhere in the paper. The only eICU validation split fits a temperature, so on the reported evidence the row was selected on the split it reports. The result may well hold, but the row cannot be read as one deployable recipe.

  2. Claims Contradicted by Table 1: The abstract claims state-of-the-art performance across frozen-backbone, end-to-end and test-time baselines, yet CLUDA reaches 76.61 AUROC on Sepsis against 76.31 for PILOT in the same table. That single cell falsifies the sentence on one of the five clinical tasks, and the paper never narrows the claim. Section 4.2 repeats the pattern by calling end-to-end fine-tuning an upper bound, although CLUDA is ahead of that row as well. A label-supervised adapter is also compared with unsupervised test-time adaptation, which inflates the headline margin, and the strongest per-task baseline is the comparator that matters.

  3. Central Mechanism Not Ablated: Section 4.4 attributes the gain to the task gradient flowing through the frozen predictor, but no row of Table 18 sets that loss to zero. The row labelled no target-task loss removes the auxiliary head instead, which Table 4, rule R7 and Appendix K.5 all describe as separate from the main task loss. The central mechanism is therefore supported only by elimination, since no arm removes it while holding everything else fixed. The remaining rows vary retrieval, alignment, pretraining and output mode, and the largest effects come from reconstruction terms, which do not test the through-predictor path.

Questions for authors
  1. Which split selected the per-task leaves in Table 1, and can the authors report the full method as one fixed configuration?
  2. On what quantity and split does the patience-based early stopping in Phase 2 operate, and could test data have influenced it?
  3. Suggestion: add a λtask=0\lambda_{\text{task}} = 0 arm with the auxiliary head and architecture fixed, plus a supervised autoencoder trained without a frozen predictor.
  4. What budget, seed count and stopping rule did the fine-tune and affine baselines receive, and why does the fine-tune row carry no spread?
Reviewer_SjrJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies domain adaptation for pretrained time-series predictors whose parameters cannot be updated. It proposes PILOT, an input-space adaptation framework that learns to transform target-domain time series while keeping the original predictor frozen. It uses a two-stage training strategy consisting of reconstruction pretraining and supervised adaptation with target-domain labels. Experiments on five clinical prediction tasks and five AdaTime benchmarks show improved performance over frozen predictors and competitive results against existing adaptation methods. The paper also demonstrates that adapters trained with LSTM predictors can transfer to GRU and TCN architectures without retraining.

Critical strengths
  1. Comprehensive experimental evaluation. The paper evaluates PILOT on five clinical prediction tasks across two target cohorts and five non-clinical time-series benchmarks. These experiments provide broad empirical evidence for the effectiveness of the proposed framework.

  2. This work shows that an adapter trained with a frozen LSTM can be directly applied to GRU and TCN predictors without retraining. The consistent performance improvements across different architectures suggest that the learned input transformations are not tied to a specific predictor. This is an interesting and potentially useful finding.

Critical weaknesses
  1. Incremental extension of existing input adaptation methods. Prior works such as Voice2Series, Time-LLM, and TATO have already explored input adaptation for frozen models. PILOT extends this idea with per-timestep transformations, axial attention, and source-domain retrieval. However, these additions mainly rely on established techniques, and their advantages over simpler input adaptation designs are not clearly demonstrated. 
  2. PILOT uses labeled target data, while most UDA and TTA baselines do not. This makes the reported performance advantages less convincing. This paper lacks comparisons with simple neural input adapters trained with the same target supervision and comparable model capacity.
Questions for authors
  1. Methodological novelty. What are the key methodological differences between PILOT and existing input adaptation methods such as Voice2Series, Time-LLM, and TATO? Can the authors demonstrate the specific advantages of per-timestep adaptation over simpler input transformation designs?
  2. Fairness of baseline comparisons. PILOT uses labeled target data, while most UDA and TTA baselines do not. Could the authors provide comparisons with simple neural input adapters under the same target supervision and comparable model capacity?
  3. The authors claim that PILOT restores source-domain temporal patterns. Figures 3 and 4 show that most adapted features become further from the source domain in terms of autocorrelation. How does the evidence support the claimed restoration of temporal patterns?
  4. The paper claims that the same adapter transfers across deployments and modalities. However, the experiments appear to retrain adapters for different datasets. Could the authors clarify whether the same trained adapter can transfer directly across domains?
  5. The ablation results in Table 18 show that retrieval and MMD alignment provide limited or inconsistent improvements. Could the authors further clarify which components are essential and provide additional ablations for the individual loss terms?
Reviewer_JpjaOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose supervised domain adaptation for time-series where adaptation happens in feature space. In other words, they learn from target labels how to transform target domain features such that a model trained on the source domain performs well. They assume the model is frozen but we can backpropagate through it. They propose a loss containing six distinct terms: (a) task performance when applying the source-trained model to the transformed features; (b) task performance when applying an auxiliary model to the transformed features; (c) a linear probe that forces latent representations of the adapter to be predictive; (d) reconstruction and (e) range losses that encourage transformed features to match input features and fall within per-feature distributions, respectively; and (f) MMD penalty encouraging similarity between distributions of transformed target features and source features. The authors also emphasize the importance of the architecture, which optionally includes a source memory bank that retrieves source examples most similar to a given target example and "allows selective, feature-level borrowing" from the source examples. Results show that their method is effective, including against models trained on the target domain.

Critical strengths

While I am not convinced by the setup the authors use to motivate the work, I do think the idea of input-space adaptation has merit, as it (a) is (arguably) more interpretable than supervised DA methods that operate in representation space, and (b) decouples adaptation from the source model, which may help pinpoint causes of subsequent degradation.

Results show that the method is effective, and ablations show the value of various components.

There is some generalizable knowledge / takeaways here, one of which is that it is good to learn an auxiliary model on the adapted time-series to avoid vanishing gradients during training.

Critical weaknesses

The setup seems artificial to me. The authors argue that they may not be able to change the source model due to regulatory constraints, which is valid, but the approach they propose is itself an AI model that would seem to require an additional regulatory approval. Moreover, the assumption that the authors can't change the model yet have sufficient access to backpropagate gradients through it is quite specific and is not a circumstance I have seen in practice.

The approach also feels very ad hoc and incremental, with 6 distinct loss components, and the authors spend very little time motivating them. In fact, the discussion of loss terms is mostly relegated to the appendix; I had to go to the appendix to understand the purpose of the linear probe, for example. They do perform ablations to show value of the individual components, though I don't see the ablation for the linear probe, which is the component I'm least clear on.

The most interesting detail to me was the use of the auxiliary head to compensate for vanishing gradients seen with loss (a), presumably early in training. But the authors don't say much about this (i.e. motivation, related work, etc), so it ends up coming across as another ad hoc component.

Most important is that the writing is poor; it's a frustrating read, with countless turns of phrase that likely make sense to the authors, who already know what was done, but do not get the point across effectively to the reader. The authors do not effectively distill the results into takeaways that might be applied by others going forward. I think there is a nice contribution here, but it is buried in a mountain of poor writing that is likely intended to make the contribution seem more substantial, but in my opinion ends up having the opposite effect.

All in all, due to the combination of poor writing and ad hoc, incremental methods, I am not sure what has been learned here, and I am doubtful that this method would ever be used in a real-world setting.

Questions for authors

My questions are implied by the listed weaknesses.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.