acceptodds
Under review as a conference paper at ICLR 2027

ProFIT: Prototype-Guided Latent Diffusion for Few-Shot Irregular Time Series Forecasting

Abstract

Reliable forecasts are often required for a newly deployed weather station, a newly monitored patient, or a new wearable user before sufffcient task-speciffc history can be collected. This setting is particularly challenging when observations are sparse and asynchronous: an unseen task may differ from the training tasks in both its underlying dynamics and its sampling mechanism, whereas existing irregular forecasters typically require substantial within-task history and most few-shot models assume regularly sampled inputs. We propose ProFIT, a Prototype-Guided Latent Diffusion framework that infers a continuous-time task function from only a few observed events. ProFIT ffrst encodes complete meta-training tasks directly as sets of timestamped variable-value events, learning a global function latent space without interpolation or regulargrid assumptions. For a new task, a sampling-aware matcher combines observed values with sampling statistics to retrieve representative task prototypes, which guide conditional latent diffusion over plausible function latents. A forecast-alignment objective further constrains generated latents to preserve the predictive semantics of complete-task representations. The resulting latent is decoded at arbitrary query times to produce probabilistic forecasts. Experiments on multiple realworld irregular time-series datasets show that ProFIT consistently achieves strong forecasting performance. These results highlight the potential of prototype-guided latent diffusion for forecasting unseen irregular processes from extremely limited observations.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Few-Shot Forecasting, Irregular Time Series Forecasting, Latent Diffusion, Prototype-Based Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_jYhKOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ProFIT addresses entity-disjoint few-shot forecasting of irregularly sampled multivariate time series: given only K asynchronous observation events (t, v, x) from an unseen entity (station/patient/subject), forecast values in a future window. The method learns a complete-task function latent via a variational encoder-decoder (attention pooling over event embeddings with multi-scale log-Fourier time features; FiLM-modulated Gaussian decoder), clusters standardized posterior means into P prototypes, and at test time routes the K-event context to a convex combination of prototypes using both a context embedding and explicit sampling statistics (span, gap moments, variable frequencies). A conditional latent diffusion prior (with time-progressive weakening of prototype guidance) generates task latents, trained with a forecast-alignment objective (predictive KL between teacher and generated latents) plus prototype anchor/usage-balancing losses.

Critical strengths
  1. A well-specified and practically relevant problem formulation. Entity-disjoint transfer with K-shot counting events (not windows), nested support sets, and fixed target windows is a clean, reproducible protocol that faithfully captures the "newly deployed station/patient" scenario. The explicit statement that this evaluates transfer rather than controlled sampling-shift interventions is honest scoping.
  2. Sampling-aware routing is a sensible and somewhat novel touch. Augmenting the context embedding with explicit sampling statistics (valid ratio, gap mean/std, variable frequencies) so that "similar-valued contexts with different spans, gaps, or coverage" route differently is a simple idea that directly targets a real failure mode of prototype/meta-learning approaches on irregular data.
  3. Efficiency reporting is exemplary for this venue's standards. Table 4 reports parameters, memory, latency with std over repetitions, and DDIM step–accuracy trade-offs, including the honest observation that non-diffusion baselines are faster but less accurate. Many comparable papers omit this entirely.
  4. Complete-task latent as teacher with diffusion prior at test time is a coherent way to get a proper posterior over task dynamics under extreme sparsity, and the forecast-alignment loss (KL on predictive distributions rather than latent vectors) is the right alignment target.
Critical weaknesses
  1. All results are single-run point estimates. The paper states this explicitly. Few-shot evaluation is notoriously high-variance, and several margins are small (e.g., Activity K=25: 1.591 vs. ASTGI 1.690; PhysioNet K=30: 0.964 vs. Hi-Patch 0.912). Without multi-seed error bars, the headline "lowest error on USHCN and Activity for every context budget" cannot be statistically assessed. This is the most consequential weakness.
  2. The PhysioNet claim is overstated relative to Table 1c. Hi-Patch beats ProFIT at K=5 (0.984 vs. 1.145) and at K=30 (0.912 vs. 0.964) in MSE. "Competitive" is technically defensible, but the abstract's framing (best on two datasets, "competitive" on the third) hides that ProFIT loses 2 of 4 K-settings on PhysioNet to a simpler non-generative baseline.
  3. Baseline adaptations look fragile. ST-GFSL and ASTGI produce catastrophically bad numbers at small K (USHCN K=1: 81.9 and 82.1 MSE, ~14× ProFIT), suggesting the "native interface" adaptation of these methods to event-level irregular few-shot input is broken rather than merely weak. Similarly, how UniTS (a regular-grid model) consumes asynchronous events "without interpolation or forward filling" is not specified. If baselines are handicapped by the interface, the margins are hard to interpret.
Questions for authors

See weaknesses above

1 comment

Responses to Weakness 1: We thank the reviewer for pointing out the insufficiency of single‑run point estimates. We have supplemented experiments with model seeds 42, 43, and 44 for USHCN, Activity, and PhysioNet, reporting the mean ± sample standard deviation of normalized MSE, RMSE, and MAE for each K. The results and related discussion are added in Table 12, Appendix J.

Dataset and scaleKMSE (mean ± SD)RMSE (mean ± SD)MAE (mean ± SD)
USHCN (MSE ×10⁻³; RMSE/MAE ×10⁻²)15.808 ± 1.1057.597 ± 0.7424.532 ± 0.534
55.785 ± 1.0857.582 ± 0.7294.520 ± 0.523
155.770 ± 1.0747.573 ± 0.7234.514 ± 0.520
305.762 ± 1.0677.568 ± 0.7184.510 ± 0.517
Activity (MSE ×10⁻²; RMSE/MAE ×10⁻¹)251.650 ± 0.0671.284 ± 0.0260.992 ± 0.026
501.601 ± 0.0561.265 ± 0.0220.973 ± 0.024
751.566 ± 0.0591.251 ± 0.0230.961 ± 0.025
1001.524 ± 0.0681.234 ± 0.0270.949 ± 0.028
PhysioNet (MSE/MAE ×10⁻²; RMSE ×10⁻¹)51.131 ± 0.0241.063 ± 0.0116.126 ± 0.044
151.025 ± 0.0311.012 ± 0.0155.738 ± 0.122
300.949 ± 0.0310.974 ± 0.0165.459 ± 0.144
600.877 ± 0.0370.936 ± 0.0205.200 ± 0.197

Responses to Weakness 2: This paper studies the few‑shot forecasting problem for irregular multivariate time series on new tasks with inherent correlations. Strictly speaking, the PhysioNet dataset is not particularly suitable for our model, or rather, not entirely aligned with our problem. ProFIT's design goal is to leverage accumulated experience from training tasks to help unseen tasks with only a few irregular observations build predictive representations. The "unseen tasks" require inherent correlation with training tasks. For example, in the traffic sensor scenario, the newly added sensors and original sensors are in the area, and traffic flows are correlated. Thus, even if the target task lacks sufficient history, the model can forecast using reusable variation patterns from other tasks. USHCN weather station network and Activity human motion sensor network fit our definition. However, patient heterogeneity in PhysioNet causes a lack of this inherent correlation. Given limited availability of public irregular time‑series datasets, and our model's results being reasonable, we included it in the experiments. We will modify inappropriate performance descriptions like "competitive" in the revised manuscript to express our model's experimental results more appropriately.

Responses to Weakness 3: It should be noted that the catastrophic errors of ST‑GFSL and ASTGI at K=1 are not artifacts of broken interfaces, but reflect their architectural incompatibility with extreme event‑level sparsity. ST‑GFSL learns task‑specific transferable metrics by constructing spatio‑temporal graphs. When K=1, there are not enough nodes and edges to build any meaningful graph, causing metric learning to degenerate. Similarly, ASTGI relies on adaptive spatio‑temporal graph interactions. With only one asynchronous event across variables, it cannot compute cross‑variable dependencies, and its aggregation mechanism collapses. Regarding input for regular‑grid models like UniTS, introducing extra interpolation or padding would alter the original model and compromise fairness. Therefore, we use no interpolation or forward filling; instead, each asynchronous event (ti,vi,xi)(t_{i},v_{i},x_{i}) is mapped to a continuous‑time input token. Specifically, UniTS receives an ordered sequence containing observed value xix_{i}, variable identity mask viv_{i}, and exact elapsed time Δti=ti−ti−1\Delta t_{i}=t_{i}-t_{i-1}. This time‑aware tokenization is a standard and fair way for regular‑grid models to process irregular sequences without violating actual temporal gaps.

Therefore, the margins are fair and interpretable. Baselines specifically designed for continuous‑time or irregular data (e.g., APN, TimeFlow, Hi‑Patch) do not collapse at K=1; instead, they degrade gracefully. This proves our event‑level interface is fair and functional. To further validate the comparison, we have supplemented additional experiments in Table 1 with recent works adapted for irregular forecasting (Rubato, LightGTS) and state‑of‑the‑art general foundation models (TOTO 2, Moirai) under the same fair interface, proving the superiority of our work.

Reviewer_SJwhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes a context-guided approach for irregular time-series modeling based on latent diffusion models. In addition to addressing the general challenges of irregular time-series, the authors aim to tackle a difficult setting where the test-time entity is unseen during training, and its latent dynamics must be inferred from K observations. The approach consists of two stages. The first stage, Complete-Task Function Latent Learning, learns a latent representation over an irregular time-series using a neural process with VAE-style learning. This representation is trained to reconstruct observation values conditioned on observation time and type. In the second stage, the Stage-1 latent representations of the training entities are clustered into prototypes using K-means. The K observations of a new entity are then encoded and used to query these prototypes, creating a prototype anchor. A conditional diffusion model—conditioned on both the encoding of the K observations and the prototype anchor—is trained to forecast the latent representation of the test-time entity, which is then decoded into observations. At test time, multiple latent representations are sampled. Experiments on USHCN, Activity, and PhysioNet compare forecasting accuracy against six baselines across various context sizes, alongside component ablations and a hyper-parameter sensitivity study. ProFIT outperforms the baselines on USHCN and Activity, while achieving mixed results on PhysioNet.

Critical strengths
  1. The proposed approach does not rely on special tricks for irregular time series like interpolation or grid. It is closer to a clean or standard conventional neural process formulation.
  2. The evaluation results on USHCN and Activity shows the competitiveness of the proposed approach and covers different context window sizes. The ablation study on different critical component of ProFIT justifies their contributions and necessities.
Critical weaknesses
  1. Major presentation issues: This is the only paper in my review batch, and the only major AI conference submission I have ever read, that leaves the sections unnumbered. This is a major presentation issue that needs to be fixed, as it makes it difficult for reviewers to refer to specific sections when providing feedback. In addition, the Stage-2 section, "SAMPLING-AWARE PROTOTYPE-GUIDED LATENT DIFFUSION PRIOR," is dense and hard to follow. It covers three distinct topics—k-means clustering, routing through the prototypes, and conditional diffusion model training—without clear transitions or separations.

  2. Minor presentation issues: Table 3 and Table 8 are identical. Additionally, the formatting of the equation blocks leaves a significant amount of wasted white space in the METHOD section. The overall presentation could be much more compact.

  3. The use of "K-Shot/few-shot" framing to describe the number of observations in an irregular time-series context window is non-standard and deviates from the established meaning of few-shot learning.

  4. The forecasting accuracy on the PhysioNet dataset is not impressive, trailing behind Hi-Patch on MAE for every choice of K. Furthermore, on the USCHN dataset, the value of K appears to have no impact on the model's prediction error, which is unexpected.

Questions for authors
  1. Could the authors explicitly confirm whether an "entity-disjoint unseen task" means the test entity differs from training entities in "both its dynamics and its sampling pattern" (as noted in the Abstract)? Specifically, does this encompass out-of-distribution testing? If so, which experiment demonstrates this? The Introduction states that the protocol is “not a separately controlled intervention on sampling-pattern or numerical-dynamics shifts.” If it does not encompass out-of-distribution testing, how does the "entity-disjoint unseen task" setting on the PhysioNet dataset differ from the entity-level held-out split used in the benchmark protocols of Hi-Patch and t-PatchGNN? In those works, Activity is split over recordings rather than subjects. What does an “entity” refer to for Activity in this work's evaluation protocol?

  2. Nothing prevents ProFIT from being evaluated under the standard irregular-forecasting benchmark protocol used by the cited baselines (e.g., Hi-Patch), as the context encoder accepts any number of events. These evaluation protocols feature standardized observed windows, fixed horizons, and training/test splits. Please report ProFIT's performance under this protocol and compare it against the baselines.

  3. Fig. 4 (c) shows four non-zero routing weights from four prototype representations. Eq. 16 masks all but the top-k logits. The experimental setup states that P=4. Does this mean k=P=4 in this experiment on the USHCN data?

3 comments

Responses to Weakness 1: Thank you for the reviewer’s comments. We sincerely apologize for the missing section numbering and dense formatting. We have restored all numbering and restructured Section 4.2 into three clear subsections, adding explicit transitions to improve readability.

Responses to Weakness 2: In the revised manuscript, we have removed duplicate tables and compressed the METHOD section by merging consecutive equation blocks into compact align environments, significantly reducing whitespace.

Responses to Weakness 3: We agree that using "K-shot" to denote the number of observations is non‑standard. We have globally revised the terminology to "few‑shot evaluation with a context of K observed events" and added a clear definition in Section 5.1 to avoid ambiguity.

Responses to Weakness 4: First, regarding the insufficient performance on PhysioNet, strictly speaking, PhysioNet dataset is not particularly suitable for our model, or rather, not entirely aligned with our problem. ProFIT's design goal is to leverage accumulated experience from training tasks to help unseen tasks with only a few irregular observations build predictive representations. "unseen tasks" require inherent correlation with training tasks. Thus, even if the target task lacks sufficient history, the model can forecast using reusable variation patterns from other tasks. USHCN weather station network and Activity human motion sensor network fit our definition. However, patient heterogeneity in PhysioNet causes a lack of this inherent correlation. Given the scarcity of public irregular time‑series datasets, and our model's results being reasonable, we included it in the experiments.

Second, due to small values of K, few‑shot learning may exhibit some irregular behaviors due to randomness. For example, UniTS performs best at K=15 and degrades at K=30; K=5 is worse than K=1. ST‑GFSL performs best at K=1, and K=5 is better than K=15. TimeFlow performs better at K=1 than K=5.

Responses to Question 1: First, ProFIT is to use training task experience to help unseen tasks with limited irregular observations build predictive representations. For example, in USHCN, we train on specific (seen) weather stations and treat test stations (no training data available as for seen stations) as new tasks, aiming to predict them using few new task data. There is an inherent correlation between new and existing tasks (despite different locations, they share meteorological correlations), and each station's dynamics and sampling patterns naturally differ. In fact, our experiments were not specifically designed to investigate OOD‑related problems. As stated , our protocol is "not a separately controlled intervention on these shifts." This means we do not artificially construct synthetic OOD test sets. To clarify potentially misunderstood definitions: here, an "entity" refers to an independent data source (e.g., a weather station, a patient, or the five individuals A/B/C/D/E in Activity), while a "task" refers to time series data generated by these entities. The specific observations along the time dimension are called "events."

Second, regarding the difference from Hi‑Patch and t‑PatchGNN benchmark protocols: while our task partitioning (entity‑level hold‑out) aligns with these baselines, the forecasting setting is fundamentally different. For USHCN, Hi‑Patch uses a standard 24‑month continuous observation window, receiving all continuous observation patches (hundreds of dense time steps) to predict the next 6 months. In contrast, under our 70%/30% temporal split, ProFIT is strictly limited to only K discrete events sampled from history (e.g., K=1 to 30), with no interpolation, to predict target events in the subsequent 30% window.

Responses to Question 2: We would like to clarify that our work is not a standard irregular time series forecasting model; the problem and baselines are different (as answered above). This causes performance variations across USHCN, Activity, and PhysioNet. Since no identical problem model currently exists, we use related baselines.

And, according to this comment, we have supplemented exploratory comparisons under standard observation windows. We directly applied the trained ProFIT to these protocols. Results are as follows (Due to character limits, only part of the table is shown below; the full version and related discussion are provided in Table 10, Appendix G).

USHCNMSE(×10⁻¹)MAE(×10⁻¹)
ProFIT6.3153.055
Hi‑Patch5.07 ± 0.173.06 ± 0.06
t‑PatchGNN5.23 ± 0.023.24 ± 0.19

Responses to Question 3: In the case study, a k=P=4 selection is used, without specifically optimizing the k selection. In the model architecture design and main experiments, we adopt the selection method in the paper's formula.

Reviewer_SJwhOpenReview ↗

Post Rebuttal Comment

I would like to thank the authors for improving the paper's presentation and clarifying their response to my first question. However, my concerns have not been fully addressed, and the response has introduced new ones:

  1. Regarding the performance on PhysioNet, while I agree with the authors' explanation that "patient heterogeneity in PhysioNet causes a lack of this inherent correlation," this implies a significant limitation to the proposed method. I am concerned that the claimed contributions may only hold for specific types of data.

  2. In their response to Q1, the authors state that "ProFIT is to use training task experience to help unseen tasks with limited irregular observations build predictive representations." However, the term "unseen tasks" is ambiguous. For example, is a record from a different patient in the PhysioNet dataset considered an unseen task? If the authors assume an inherent correlation between new and existing tasks, how does an unseen or new task differ from a standard test sample in a typical machine learning setup?

Given the remaining ambiguity regarding the paper's actual contributions, I cannot recommend acceptance.

We appreciate the reviewer’s valuable comments and would like to take this opportunity to further clarify the relevant concerns.

  1. First, machine learning models are usually designed for specific research problems and are naturally tailored to certain tasks and data distributions. As long as such data captures the representative characteristics of a specific subclass of problems, and the model demonstrates advantages for solving this subclass of tasks through evaluations under this setting, its value and contribution can be justified. To illustrate the motivation of our work, Allow us to provide an informal example. An area initially has 100 traffic monitoring sensors. Because traffic flow varies greatly between day and night, the sampling frequency of each sensor is irregular. Now, 50 additional sensors are just deployed. How can we forecast for them with extremely limited data? Actually, this is a practical problem we met and serves as the primary motivation for our work. Abstracting this problem, this paper aims to study the few‑shot forecasting problem for irregular multivariate time series on unseen tasks. To date, we have not found any research work with identical problem setting. Regarding the performance results on the PhysioNet dataset, the performance gap arises from a moderate misalignment between the dataset’s application scenario and the target problem of our work, rather than from specific data types. In the revised manucript, we have further revised the relevant statements to more accurately reflect the contributions of our model.

  2. as to "unseen tasks", we also agree that this term is somewhat ambiguous. We use this term to refer to the time-series data generated by specific entities, e.g. newly sensors in the preceding example, newly added weather stations in meteorological monitoring networks, or a patient whose data is excluded from training and validation in the PhysioNet dataset. Here, “unseen” characterizes the specific entities that are newly introduced into the system and lack historical data describing themselves. The term “task” is adopted to follow the terminology convention in few-shot learning. In the revised manuscript, we have further clarified the definition of "unseen task" to avoid the ambiguity.

Reviewer_wasjOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

ProFIT studied entity-disjoint few-shot forecasting of irregular time series. It encodes complete meta-training tasks into a shared function-latent space and uses sampling statistics to route unseen tasks to representative latent prototypes. A conditional latent diffusion with prototypes then generates task-specific function latents, guided by a forecast-alignment objective to preserve predictive semantics. Experiments on USHCN, HumanActivity, and PhysioNet show competitive or good forecasting under K -shot protocols.

Critical strengths
  1. The paper is clear and easy to follow.
  2. prototypes and diffusion follow existing works and are reasonable.
Critical weaknesses
  1. Computational Overhead and Scalability: Despite the efficiency analysis showing speedups via DDIM-5, the overall framework requires training a VAE, clustering prototypes, and running a diffusion model for inference. The computational complexity and memory footprint are significantly higher than non-generative baselines. The paper does not adequately justify this massive increase in computational cost given the marginal or negative gains on certain datasets.
  2. Unconvincing Performance on PhysioNet: While ProFIT excels on USHCN and Activity, it underperforms compared to simpler baselines like Hi-Patch and TimeFlow on the PhysioNet dataset, which is a critical benchmark for irregular clinical time series. The authors dismiss this as "competitive," but failing to improve upon existing methods on a primary benchmark raises concerns about the robustness of the diffusion prior in high-dimensional, noisy clinical settings.
  3. Weak Forecasting Baselines and Architectures: The paper fails to compare against modern, highly optimized irregular or few-shot forecasting foundation models (e.g., TOTO-2, or Moirai, LightGTS, in zero-shot/few-shot/transfer settings, or RUBATO in irregular settings). Demonstrating superiority only over older or less capable baselines further weakens the claim of performance.
  4. Limited Methodological Novelty and nothing new from this paper: prototypes and diffusion (and flow matching) are well studied in time-series analysis.
Questions for authors

See Weaknesses

3 comments

Responses to Weakness 1: Thank you for the reviewer’s comments. In fact, the problem our work aims to solve and the baselines are somewhat different. Allow us to explain this with an example: An area initially has 100 traffic monitoring sensors. Because traffic flow varies greatly between day and night, the sampling frequency of each sensor is irregular. Now, 50 additional sensors are just deployed. How can we forecast for them with limited data? This is a practical problem and serves as the primary motivation for our work. Abstracting this problem, this paper aims to study the few‑shot forecasting problem for irregular multivariate time series on unseen tasks. To date, we have not found any research work with identical problem setting. We adopt closely related baseline methods for comparison.

The use of prototypes and diffusion is specifically targeted at solving the aforementioned problem. The detail discussion please refers to Responses to Weakness 4.

Regarding the computational cost, ProFIT's overall cost is well within a reasonable range. As detailed in newly added Table 11, Appendix H, training the entire framework (including the VAE, latent diffusion, and K‑means clustering) takes less than 2 minutes (92.9 seconds), with a peak GPU memory footprint of merely ~75 MiB.

ComponentTime (s)
Data loading (outside training total)0.004
Stage 1: VAE training (60 epochs)22.187
Stage 2: context and latent‑diffusion training (60 epochs)70.722
    Latent statistics (within Stage 2)0.017
    Stage‑1 latent collection (within Stage 2)0.016
    K‑means prototype construction (within Stage 2)0.003
Complete training (Stages 1–2)92.909

Responses to Weakness 2: We have also previously noticed the performance gap on PhysioNet. Strictly speaking, PhysioNet dataset is not particularly suitable for our model, or rather, it does not perfectly align with the problem we aim to solve. ProFIT's design goal is to leverage accumulated experience from training tasks to help unseen tasks with only a few irregular observations. The so‑called "unseen tasks" need to have an inherent correlation with training tasks. Therefore, even if the target task has not accumulated sufficient history, the model can still forecast using reusable variation patterns from other tasks. Weather stations in USHCN and the human motion sensors in Activity perfectly fit our problem definition. But, heterogeneity of patients in PhysioNet leads to a lack of inherent correlation. Considering the limited availability of irregular time‑series datasets, and given our model's results are not poor, we also included it.

We have modified inappropriate performance descriptions like "competitive" to express more appropriately.

Responses to Weakness 3: According to this comment, we have added new baselines as shown below. (Due to character limits, only part of the table is shown; the full version are provided in Table 1)

MethodMSE K=1K=5K=15K=30RMSE K=1K=5K=15K=30MAE K=1K=5K=15K=30
Moirai106.77450.84120.55618.61532.67622.54814.33713.64419.12311.4767.5107.378
Toto 2.08.9078.9258.9638.9639.4389.4479.4679.4685.4705.4765.4895.490
Rubato8.9788.9909.0239.0689.4759.4829.4999.5235.5505.5435.5425.547
LightGTS10.00511.8987.4175.95610.00310.9088.6127.7185.6565.7924.5544.829
ProFIT5.8085.7855.7705.7627.6217.6067.5967.5914.5324.5204.5144.510

Responses to Weakness 4: Since traditional dense models fail when the context is extremely sparse, we utilize prototypes to provide a bounded anchor for global dynamics across tasks. However, standard prototypes ignore sampling patterns, so we specifically designed a sampling‑aware routing mechanism that incorporates event intervals and coverage statistics, ensuring the model knows not only the "values" of the sparse data but also "how" it was sampled. Furthermore, because sparse contexts naturally correspond to multiple plausible futures, we adopted a diffusion prior. To prevent this diffusion from becoming blind and unconstrained, we introduced a novel time‑decayed prototype conditioning, smoothly transitioning guidance from high‑noise global patterns to task‑specific recovery. Finally, since latent similarity does not guarantee similar predictions, we introduced a forecast‑alignment objective to explicitly couple the generative latent space with future predictive distributions. Therefore, the prototype and diffusion components are not merely stacked together, but deeply coupled and specifically adapted to solve the information scarcity problem in irregular few‑shot forecasting.

Thanks for the rebuttal.

I still have questions on whether foundation models only use test K-shot or see the training datasets the same as ProFIT.

And given the limited methodological novelty, I will keep my score.

We thank the reviewer for raising this comment.

  1. The base models do not merely use K samples from the test task; instead, like ProFIT, they are trained on the full local training set. Specifically, for Moirai and LightGTS, we start from their official pre-trained weights and fine-tune them using exactly the same training-task splits as ProFIT, with model selection performed on the identical validation-task splits. Throughout this process, the ground-truth targets of the test set are strictly isolated from all models. For Toto 2.0, since its official release does not yet provide a fine-tuning pipeline, we evaluate its pre-trained weights under the officially supported zero-shot setting.

  2. As to the limited methodological novelty, We would like to further clarify this point. Our work investigates a new problem centered on addressing sparsity, asynchronous sampling, and forecasting with extremely few samples. We adopt prototypes and diffusion because certain properties of these techniques fit the aforementioned challenges. For example, prototypes are naturally suited for cross-task few-shot routing, providing a bounded anchor when the context is extremely scarce. And, diffusion models are specifically chosen because sparse contexts naturally correspond to multiple plausible future trajectories. However, prototypes and diffusion alone cannot resolve the problem. We propose novel mechanisms to construct the proposed model (As mentioned in the Response to Weakness 4), and extensive experiments validate its effectiveness. To better clarify the target problem and contributions of our work, we have further revised the manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.