- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3(+1)
- Soundness
- 3(+1)
Read review
Paper summary
The paper introduces SeriesPFN, an end-to-end prior-data fitted network (PFN) that performs time-series classification and extrinsic regression through in-context learning. It supports univariate and multivariate time series and predicts targets for new examples by conditioning on labeled context examples, without updating its parameters. Its main contributions are (1) a unified synthetic task prior for classification and regression and (2) a shared encoder and in-context learning architecture with task-specific target embeddings and prediction heads.
The method is validated primarily through benchmark comparisons and ablation studies. The reported results are strong, although a clearer explanation of the rationale behind some design choices would help readers understand how the authors arrived at the final method and when those choices are expected to generalize.
Critical strengths
SeriesPFN achieves strong reported performance on both classification and extrinsic regression using a single jointly trained checkpoint: the best mean accuracy rank on UCR112 and UEA26, and the second-best mean RMSE rank on TSER62. These results support the practical value of the unified approach.
Critical weaknesses
Limited explanatory insight into the unified approach. The ablations support the method's effectiveness, but leave open which target-relevant temporal characteristics are shared between classification and regression and how joint training exploits them. A concrete real-data example would help explain the observed transfer benefit. Similarly, the relationship between the two-endpoint regression prior and the diversity of real regression tasks remains insufficiently illustrated. These gaps limit the insight readers can gain beyond the reported performance improvements.
Questions for authors
-
Class-label embeddings and class capacity (clarification). Appendix B.4 describes assigning task labels to learned class slots, and Table 5 specifies a maximum of 64 slots. How should readers reconcile this capacity limit with Section 4.3's description of the retrieval decoder as non-parametric in the number of classes? Could the same checkpoint support more than 64 classes, or would that require modifying and retraining the model?
-
Number of
[CLS]tokens (design justification). Section 4.2 explains that the four tokens gather information and are pooled and concatenated into a 512-dimensional series representation. What motivated choosing four tokens rather than one or another small number? Was this choice based on an architectural requirement or an empirical comparison, and how sensitive is performance to the token count when representation size and computational cost are accounted for? -
Regression prior expressivity (clarification). Section 3.4 generates regression tasks by linearly interpolating between two endpoint parameter settings. How restrictive is the two-endpoint linear parameterization, and what evidence suggests that it provides sufficient diversity for real regression tasks?
-
Motivation for joint training (clarification). Section 5.3 suggests that classification benefits from representations learned through regression, and the single-task ablations provide some supporting evidence. Could the authors provide a concrete illustrative example from real data showing a temporal characteristic useful for both classification and regression, and explain how learning it through regression could benefit classification? Even a small case study would help readers understand the motivation for joint training beyond the aggregate performance results.
-
TiCT preprocessing (clarification). Could the authors confirm whether TiCT's input time series were resampled to length 512 before inference? The local TiCT release enables length normalization with a target length of 512 and applies
scipy.signal.resample(data, max_len, axis=2)when the input length differs, for arrays shaped(samples, channels, time). Appendix C.1 already discusses differences in the train/test split protocol; confirming the preprocessing would help rule out an additional source of the reported performance discrepancy.
Suggestion for future work. A sensitivity study of the 50/50 classification/regression and source-dependent/combiner-dependent task mixtures would help assess how robust the results are to these sampling choices.
5 comments
Response to zU1W (2/2): Shared features and inference
Q4.
[...] a concrete illustrative example from real data showing a temporal characteristic useful for both classification and regression [...]
An illustrative example is beat timing. BIDMC32HR predicts continuous heart rate from PPG and ECG signals. If we grouped its targets into rate intervals, the same beat timing would support classification. Learning to resolve the continuous rate then also provides the information needed to distinguish the groups. Our prior is built in the same way: a regression target varies continuously the same generating parameters that a class label varies discretely.
This example illustrates the motivation; we did not run it as an experiment. The single-task ablations give indirect support. In the joint model, about half of the training tasks are classification, so a classification-only model trained for the same 100k steps sees roughly twice as many classification tasks. It is still 0.39 accuracy points worse on UCR112 (p = 0.003). The revision adds a classification-only model trained for 50k steps, which sees about as many classification tasks as the joint model. It is 2.64 points worse on UCR112 and 4.48 on UEA26 (both p < 0.001; Appendix E.1). This model also trains for half as long, so some of this difference may come from the shorter training.
In both comparisons the joint model is better at classification. This is what we would expect if learning from regression tasks also helps classification, as in the heart-rate example above. We have not measured which features the model shares between the two tasks, so the paper states this as a hypothesis (Section 5.3).
Q5.
Could the authors confirm whether TiCT's input time series were resampled to length 512 before inference?
Yes. TiCT inputs are resampled to length 512 when needed using scipy.signal.resample. Our wrapper uses (samples, time) arrays, so resampling is along axis 1, the same temporal axis as axis 2 in the released (samples, channels, time) layout.
Suggestion for future work.
A sensitivity study of the 50/50 classification/regression and source-dependent/combiner-dependent task mixtures would help assess how robust the results are to these sampling choices.
Appendix E.1 tests the single-task and source-only/combiner-only endpoints. Studying intermediate mixture weights remains future work.
Thank you again for the detailed questions, which helped us explain the reasoning behind the design more clearly.
Follow-up on the author response
Thank you for the detailed response and manuscript revisions. The explanations of the 64-class capacity, the four-token architectural choice, and TiCT’s preprocessing address my clarification questions. The additional single-task ablations and per-dataset analysis of the regression gap also provide useful context.
The beat-timing example is helpful: it makes the motivation for joint training more concrete by illustrating how the same temporal characteristic could support both classification and regression. I could not find this example in the revised manuscript. Could the authors include it in the paper so that readers can benefit from this explanation? It would be useful to preserve the distinction between this illustrative motivation and experimental evidence identifying which features the model actually shares.
The expressivity of the regression prior and the mechanism of transfer remain open questions, but I appreciate that the authors acknowledge these limitations and distinguish hypotheses from tested conclusions. Overall, the response strengthens my assessment, and I maintain my recommendation of weak acceptance.
Response to zU1W (1/2): Architecture and regression prior
Thank you for the review and the detailed questions. We appreciate your recognition of the "strong reported performance" on both tasks and the "practical value of the unified approach".
In brief: we give a real-data example of a cue shared by classification and regression, and we explain how the two-endpoint prior relates to real regression tasks, including a limitation that the revision now documents. We respond to each question below.
Q1.
The current checkpoint supports at most 64 classes; supporting more would require larger label embeddings and training with that capacity. "Non-parametric in the number of classes" describes only the retrieval decoder: its learned projections do not depend on the class count (Appendix B.4). The 64-class limit comes from the separate learned label embeddings (Appendix B.1 and Table 5). TabPFN-3, Section 2.2 makes the same distinction.
Q2.
Four 128-dimensional summary tokens concatenate directly into the 512-dimensional ICL representation while keeping the encoder width at 128 (Appendix B.3). This follows the four-token design in TabICL, Section 3.3, TabICLv2, Appendix A.2, TimEE, Appendix B.1, and TabPFN-3, Appendix C, which explicitly specifies the same 4 × 128 = 512 arrangement. It was an architectural choice that we adopted from these models. We did not compare token counts, so we cannot say how sensitive performance is to this choice. With direct concatenation, changing the count also changes the ICL width; keeping that width fixed would require a compensating projection or a different encoder width.
Q3.
The interpolation is between two settings of the generating process, not between two time series. The interpolated parameters then pass through the combiners, temporal filters, and nonlinear activations of the graph, so a straight path in parameter space does not make the series depend linearly on the target (Section 3.4; Appendices A.5 and A.6). For example, a bilinear combiner multiplies two weighted inputs. If both sets of weights are interpolated linearly, its output already depends quadratically on the target.
There is still a restriction: within one task, the parameters follow a single straight path. Sources, graphs, and endpoints differ between tasks.
The evidence that this construction transfers to real tasks is the TSER62 result: SeriesPFN is trained only on these synthetic tasks and ranks second among all compared methods. The gap to the first, TS2TabPFN, comes mostly from the 14 datasets that the other compared methods predict most accurately. There, SeriesPFN's RMSE is 120% higher, against 9% on the other 48 (Appendix D.8). In the examples we examined, the published target definitions involve averages, totals, or maxima over the observation period. We hypothesize that such tasks are rare in our prior, because its targets set the parameters of the process that generates a series and are not computed from the series. We have not yet tested this hypothesis.
We have also not trained with curved or multi-anchor paths, so we cannot say whether they would reduce the gap. They remain future work, as the conclusion notes.