- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper investigates whether a small number of input prompt tokens can improve frozen time-series foundation models. PaCTS combines shared learnable prompts with low-rank, instance-adaptive prompts generated from input statistics, and refines them through cross-attention over segment statistics with relative positional information. The prompts are prepended to the original patch tokens. The model continues to process the entire visible history; it does not encode earlier observations excluded from the context window. A separate prompt module is trained for each backbone and context-length setting and shared across datasets. The experiments post-train on the training splits of GIFT-Eval, evaluate its 97 test configurations, and transfer to TIME, which the manuscript states has no dataset overlap with GIFT-Eval. At a context length of 4,096, normalized MASE/CRPS for Chronos-2 improve from 0.708/0.496 to 0.697/0.480; PaCTS also outperforms the zero-shot backbone with a context length of 8,192. The paper includes results for additional backbones, component ablations, multivariate inference, and inference costs. The GIFT-Eval results evaluate post-training on temporally held-out portions of the same datasets, whereas TIME tests transfer of the prompt module across datasets.
Critical strengths
- Practical improvements in accuracy and inference cost. Tables 1–2 (pp. 6–7) show improvements in both point and probabilistic forecasting while adding only 153,286 parameters to Chronos-2, approximately 0.128% of its parameter count. Compared with the zero-shot model at a context length of 8,192, PaCTS at 4,096 reduces FLOPs, peak GPU memory, and latency by approximately 49.3%, 18.4%, and 36.8%, respectively. The comparison and measurement conditions are sufficiently clear to motivate PaCTS as a lightweight post-training approach.
- The evaluation extends beyond a single backbone or aggregate score. The incremental ablations in Table 3 show successive benefits from static prompts, adaptive generation, and segment refinement. Results on two additional backbones and across history lengths and forecast horizons provide evidence about applicability. Transfer to TIME without further tuning and multivariate inference after univariate training are also valuable. The explicit distinction between datasets used for adaptation and datasets used to evaluate transfer makes the results easier to interpret.
Critical weaknesses
- Limited novelty and insufficient isolation of the distinctive contribution. The ten descriptors in Table 5 (p. 15) are conventional. Time-LLM uses statistics in textual prompts, Gen-P-Tuning adapts frozen time-series backbones through input-dependent prompts, and CoSPOT constructs basis prompts from spectra. Although backbones and tasks differ, PaCTS’ potential advance lies in combining global low-rank prompts, ordered segment refinement, and shared training across datasets. Table 3 adds information and capacity together, leaving the contributions of basis composition and cross-attention unresolved. A simple encoder using identical global and ordered segment statistics, with matched prompt-token and parameter counts, is missing. Appendix C also uses different training schedules and batch sizes without establishing comparable data exposure or tuning budgets. Given the MASE margin of 0.697 versus 0.703 in Table 1, the gains cannot be clearly attributed to the architecture rather than statistical conditioning or optimization protocols.
- Uncertain limits of statistical conditioning and transfer gains. Histories with different forecasting-relevant dynamics can share identical conditioning descriptors, including global and ordered segment statistics, positions, and masks, and therefore receive identical prompts. Forecasts can still differ because the backbone receives the original patches, but the paper does not test for negative transfer when the descriptors miss relevant dynamics. TIME MASE/CRPS improve only from 0.664/0.558 to 0.661/0.555 (Table 7), without seed dispersion or paired dataset-level uncertainty. Frequencies and horizons sharing a dataset are not independent observations. Claims about replacing longer contexts should remain limited to the evaluated accuracy–cost trade-off: prompts cannot access additional history, and Appendix Tables 13–14 include slightly worse short-context CRPS, such as 0.484 versus 0.481 for PatchTST-FM.
Questions for authors
- Suggested comparison. What gains do the low-rank basis and cross-attention provide over a simple prompt generator using identical global and ordered segment statistics, with matched token counts, parameters, training data, and budgets? Please provide existing controls or add this comparison.
- Protocol clarification. Do PaCTS, LoRA, and full fine-tuning match training records, task and horizon sampling, loss normalization, and data exposure? Please report training costs and validation search ranges, and clarify whether the advantage persists with matched budgets and sampling.
- Existing results. Please report all three seeds, dataset-level paired error changes, win/loss rates, and uncertainty on TIME/GIFT-Eval, accounting for dependencies among frequencies and horizons within datasets. Are the small average TIME gains stable and broadly distributed?
4 comments
Response to Reviewer UbYb (2/3)
Training protocol of the baselines (Weakness 1 and Question 2)
All methods are trained on the same GIFT-Eval training data and train / validation / test splits. In Table 1, full fine-tuning and LoRA follow the official Chronos-2 fine-tuning recipe, which we adopt to avoid reimplementation errors and to use the configuration recommended by the backbone authors.
To further match the protocols, we additionally train full fine-tuning and LoRA within the PaCTS pipeline, matching the training records, task and horizon sampling, loss normalization and data exposure. The settings and the checkpoints are selected by validation loss. As shown in the following table, both baselines perform similarly to their official-recipe results and remain behind PaCTS. Under this matched training budget, training on one GPU takes 4.7 h for PaCTS, 5.6 h for full fine-tuning and 4.7 h for LoRA, with a peak GPU memory of 32.5 GiB, 48.9 GiB and 40.2 GiB, respectively. We also train both baselines within the PaCTS pipeline on TimesFM-2.5 and PatchTST-FM r1 (see our response to Reviewer qCro), where PaCTS also outperforms them.
We will add these results and discussions in the paper.
| Chronos-2 (mean ± std over 3 seeds) | 4,096 MASE | 4,096 CRPS | 8,192 MASE | 8,192 CRPS |
|---|---|---|---|---|
| Zero-shot | 0.708 | 0.496 | 0.704 | 0.491 |
| Full FT, official recipe (Table 1) | 0.703 ± 0.000 | 0.487 ± 0.000 | 0.698 ± 0.001 | 0.482 ± 0.001 |
| LoRA, official recipe (Table 1) | 0.703 ± 0.000 | 0.493 ± 0.000 | 0.699 ± 0.000 | 0.488 ± 0.000 |
| Full FT, PaCTS pipeline | 0.703 ± 0.000 | 0.494 ± 0.002 | 0.699 ± 0.001 | 0.484 ± 0.001 |
| LoRA, PaCTS pipeline | 0.703 ± 0.001 | 0.490 ± 0.002 | 0.698 ± 0.001 | 0.485 ± 0.002 |
| PaCTS | 0.697 ± 0.001 | 0.480 ± 0.001 | 0.695 ± 0.001 | 0.478 ± 0.001 |
Identical descriptors and negative transfer (Weakness 2 and Question 3)
We understand the reviewer's concern that time series with the same descriptors receive the same prompt. The prompt, however, is an additive, gated input and the backbone still processes the raw patches, so their forecasts differ. Moreover, in practice, identical descriptors for different histories are rare: among the 371,330 GIFT-Eval test windows, only 1,305 (0.35%) share their descriptors with a window of a different history, all from two datasets (bitbrains_fast_storage and car_parts) and mostly all-zero or sparse intermittent series, while none of the 111,177 TIME test windows shares the same descriptors. Even on bitbrains_fast_storage, PaCTS still improves over zero-shot Chronos-2 (−1.11% MASE and −6.40% CRPS in the table below).
To examine negative transfer, we report the change of each GIFT-Eval dataset relative to zero-shot Chronos-2, where the 97 configurations are grouped into 23 source datasets (frequency variants and subsets of the same data, e.g. the M4 subsets, are grouped). Each cell is the change relative to zero-shot Chronos-2, as a geometric mean over the configurations of the source dataset and averaged over three seeds.
PaCTS improves both MASE and CRPS on 16 of the 23 datasets, compared with 12 for full fine-tuning and 13 for LoRA. We do observe negative transfer on a few datasets: MASE increases by more than 1% on bizitobs_application, bizitobs_service and car_parts (a single configuration), and full fine-tuning and LoRA also degrade on bizitobs_service and car_parts. These datasets differ from the rest of the training data: the two bizitobs datasets are the only GIFT-Eval datasets sampled every 10 seconds, so their daily period (8,640 steps) exceeds the 4,096-step context, and car_parts consists of sparse intermittent demand series. We will add this discussion of negative transfer to the paper.
| GIFT-Eval dataset | PaCTS Δ MASE (%) | PaCTS Δ CRPS (%) |
|---|---|---|
| covid_deaths | −16.76 | −30.80 |
| bizitobs_l2c | −10.18 | −10.60 |
| us_births | −6.87 | −4.63 |
| m4 | −4.72 | −4.83 |
| loop_seattle | −2.68 | −2.46 |
| m_dense | −1.98 | −2.73 |
| solar | −1.72 | −1.28 |
| bitbrains_fast_storage | −1.11 | −6.40 |
| sz_taxi | −0.93 | −1.34 |
| kdd_cup_2018 | −0.82 | −0.57 |
| restaurant | −0.76 | −0.90 |
| hospital | −0.49 | −4.46 |
| ett1 | −0.41 | −0.24 |
| bitbrains_rnd | −0.21 | −34.46 |
| jena_weather | −0.19 | −0.27 |
| temperature_rain | −0.11 | −0.09 |
| hierarchical_sales | +0.01 | −0.31 |
| electricity | +0.09 | −0.73 |
| ett2 | +0.18 | +0.38 |
| saugeen | +0.33 | +0.29 |
| bizitobs_service | +2.89 | +3.61 |
| car_parts | +5.04 | +4.66 |
| bizitobs_application | +10.04 | +21.06 |
| Datasets improved | 16 of 23 | 18 of 23 |
Response to Reviewer UbYb (3/3)
Uncertainty and dataset-level results on TIME (Weakness 2 and Question 3)
To show that the gains of PaCTS on TIME are stable and broadly distributed, we report the standard deviation over seeds, dataset-level changes and win/loss rates.
As shown in the following table, full fine-tuning and LoRA underperform zero-shot on TIME, whereas PaCTS improves over it by more than the standard deviation over three seeds. Each of the three seeds of PaCTS improves over zero-shot on both metrics on both benchmarks.
At the dataset level (changes relative to zero-shot Chronos-2, geometric mean over the configurations of each dataset, averaged over three seeds; the 98 TIME tasks are grouped into 37 source datasets in the same way), PaCTS improves MASE and CRPS each on 21 of the 37 datasets, better than full fine-tuning (20 / 16) and LoRA (20 / 19). PaCTS is also the only method that improves over zero-shot on average across datasets, and has the smallest worst-case degradation. At the configuration level, PaCTS improves both metrics on 64 of the 97 GIFT-Eval and 45 of the 98 TIME configurations, compared with 47 / 32 for full fine-tuning and 50 / 32 for LoRA. These results show that the improvements of PaCTS are stable and broadly distributed.
The per-dataset changes of PaCTS on GIFT-Eval are given above, and those on TIME below:
| Benchmark | Method (context) | MASE | CRPS |
|---|---|---|---|
| GIFT-Eval | Zero-shot (4,096) | 0.708 | 0.496 |
| GIFT-Eval | Full FT (4,096) | 0.703 ± 0.000 | 0.487 ± 0.000 |
| GIFT-Eval | LoRA (4,096) | 0.703 ± 0.000 | 0.493 ± 0.000 |
| GIFT-Eval | PaCTS (4,096) | 0.697 ± 0.001 | 0.480 ± 0.001 |
| TIME | Zero-shot (4,096) | 0.664 | 0.558 |
| TIME | Full FT (4,096) | 0.665 ± 0.000 | 0.560 ± 0.001 |
| TIME | LoRA (4,096) | 0.666 ± 0.001 | 0.560 ± 0.000 |
| TIME | PaCTS (4,096) | 0.661 ± 0.001 | 0.555 ± 0.000 |
| TIME (37 datasets) | Datasets improved (MASE) | Datasets improved (CRPS) | Dataset-level change vs. zero-shot (MASE) | Dataset-level change vs. zero-shot (CRPS) | Largest per-dataset MASE increase |
|---|---|---|---|---|---|
| Full FT (4,096) | 20 | 16 | +0.16% | +0.37% | +14.59% |
| LoRA (4,096) | 20 | 19 | +0.41% | +0.39% | +19.73% |
| PaCTS (4,096) | 21 | 21 | −0.40% | −0.59% | +5.15% |
| TIME dataset | PaCTS Δ MASE (%) | PaCTS Δ CRPS (%) |
|---|---|---|
| Housing_Inventory | −17.38 | −11.44 |
| Vehicle_Supply | −9.03 | −2.72 |
| Vehicle_Sales | −3.73 | −5.59 |
| Australia_Solar | −3.63 | −2.11 |
| US_Term_Structure | −1.88 | −1.25 |
| epf_electricity_price | −1.76 | −1.59 |
| Commodity_Import | −1.70 | −13.31 |
| OpenElectricity_NEM | −1.32 | +0.82 |
| Commodity_Production | −1.26 | −1.93 |
| EWELD_Load | −1.07 | −6.56 |
| Crypto | −1.00 | −0.38 |
| SG_PM25 | −0.80 | −1.56 |
| azure2019 | −0.52 | −0.82 |
| Port_Activity | −0.50 | −1.27 |
| Oil_Price | −0.46 | −0.69 |
| NE_China_Wind | −0.38 | +0.38 |
| Finland_Traffic | −0.31 | +2.15 |
| CPHL | −0.31 | −0.70 |
| WUI_Global | −0.23 | +0.31 |
| current_velocity | −0.22 | −0.53 |
| Online_Retail_2_UCI | −0.05 | −0.26 |
| SG_Weather | +0.05 | −0.33 |
| MetroPT-3 | +0.06 | −0.03 |
| SG_Carpark | +0.13 | +0.34 |
| Auto_Production_SF | +0.26 | +0.56 |
| Smart_Manufacturing | +0.35 | +0.34 |
| Water_Quality_Darwin | +0.39 | −1.85 |
| Global_Influenza | +0.70 | +3.70 |
| Supply_Chain_Location | +0.81 | +0.70 |
| Coastal_T_S | +0.95 | +0.91 |
| Supply_Chain_Customer | +1.28 | +0.92 |
| ECDC_COVID | +1.45 | +2.57 |
| Uncertainty_1M | +1.58 | +2.29 |
| JOLTS | +1.86 | +1.73 |
| Global_Price | +3.89 | +3.63 |
| Job_Claims | +4.55 | −0.41 |
| US_Labor | +5.15 | +2.66 |
| Datasets improved | 21 of 37 | 21 of 37 |
Claims about replacing longer contexts (Weakness 2)
We agree that the prompts cannot access history beyond the context window, and we will rephrase the contribution as an accuracy–cost trade-off: comparable or better accuracy than zero-shot inference at twice the context length, at substantially lower inference cost.
We thank reviewer UbYb again. We hope the answers and experiments address the concerns and questions of the reviewer.
Dear Reviewer UbYb, thank you again for reviewing our paper. As the discussion period is ending, we would greatly appreciate it if you could let us know whether our responses and the revised paper address your concerns. We are happy to clarify any remaining questions.
Response to Reviewer UbYb (1/3)
We thank reviewer UbYb for the insightful comments. In the following we address the concerns and provide answers to the questions.
Novelty and positioning (Weakness 1)
We agree that existing methods also explore input-space latent prompting. We would like to further clarify our contributions and differences from existing methods in three aspects.
Research target. PaCTS aims to improve a frozen TSFM effectively and efficiently with compact context derived from its own available history, rather than with additional information from other channels (Gen-P-Tuning) or from vision and text (UniCast, CoSPOT). Time-LLM uses input statistics as a text prefix for an LLM rather than constructing latent prompts for a frozen TSFM in our method. To the best of our knowledge, PaCTS is the first method that treats soft prompts as a summary instruction and compact context to improve frozen TSFMs in general.
Training across heterogeneous datasets. To improve frozen TSFMs in general, we train and evaluate a single prompt module jointly on heterogeneous datasets, without dataset-specific adaptation. One module thus serves all datasets without per-dataset training or storage, matches the general-purpose use of TSFMs, and transfers to datasets unseen in training (TIME, see below). In contrast, Gen-P-Tuning, UniCast and CoSPOT train a separate model for each dataset or task.
Architecture design. To make instance-adaptive prompts work across diverse datasets with different context lengths and forecast horizons, we design a statistics-conditioned low-rank prompt generator that explicitly extracts high-level information from the available history, and a cross-attention-based segment refinement that adds finer statistics of ordered segments for long inputs. The same-information control below further shows that this design makes better use of the statistics than a generic generator.
We will add these contribution clarifications in the main paper.
Prompt module architecture ablations (Weakness 1 and Question 1)
To demonstrate the effectiveness of the proposed prompt generation architecture, we conduct experiments by replacing the low-rank generator and the segment cross-attention of PaCTS with a simple MLP generator that receives the same global and ordered-segment statistics, with a matched number of parameters and the same training settings. As shown in the following table (over three seeds), the statistics alone already improve over zero-shot, and the PaCTS architecture adds a further gain on top of the same information and parameter budget. Together with Table 3, the results show that each of the three components of the prompt generator contributes: the shared prompts, the instance-adaptive generation and the ordered-segment refinement, whose segment information helps even with a generic MLP. The low-rank basis and cross-attention design then makes better use of the statistical information than a generic MLP. We will add these new ablations to Table 3.