acceptodds
Under review as a conference paper at ICLR 2027

Beyond Context Concatenation: Scalable, Manifold-Aligned Retrieval for Time-Series Foundation Models

Abstract

Retrieval augmentation equips frozen time-series foundation models with external memory, enabling them to handle distribution shifts without fine-tuning. However, existing methods face three critical bottlenecks: in-context token concatenation incurs quadratic attention complexity in sequence length and retrieval budget, causing a 46.70x FLOP increase. Simply adding covariates does not significantly improve performance because the internal representations do not align. At the same time, external memory stores cause metrics to shrink along high-variance latent axes, and numerical retrieval methods cannot predict macroeconomic shocks. In this work, we resolve these limitations through a scalable, manifold-aligned retrieval architecture. Our framework eliminates quadratic context expansion by routing candidates via an injection hierarchy into latent cross-attention or output-space Wasserstein barycenters, scaling linearly with the retrieval budget. We align future planning covariates with the representation manifold via a pre-trained projection adapter under parameter-freezing guardrails, and integrate dense text embeddings to guide retrieval and modulate fusion gating during macroeconomic shocks. Empirical benchmarks show latent injection achieves representation parity with token concatenation across Chronos-2, Moirai-2.0, and TimesFM-2.5 backbones while preventing out-of-memory collapse at production batch sizes. On the 30 planning tasks of fev-bench, aligned covariate retrieval yields a 52.09% CRPS gain over unaligned concatenation. Also, memory scaling across to sequences confirms power-law error decay () in agreement with theoretical manifold bounds, and dual multimodal grounding reduces macroeconomic shock error by 7.72% MSE. Structured latent injection and manifold alignment thus provide an efficient, scalable foundation for zero-shot time-series forecasting.

Primary area:
learning on time series and dynamical systems
Keywords:
Time-series forecasting, retrieval-augmented generation, zero-shot forecasting, latent injection, manifold alignment, covariate retrieval, distribution shift, scalable retrieval

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_bY2zOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a retrieval-augmented framework for time-series foundation models, combining latent cross-attention and output-space Wasserstein barycenters for efficient retrieval integration. It further introduces supervised covariate alignment and multimodal text grounding, alongside a theoretical analysis of memory scaling. Experiments across multiple foundation model backbones demonstrate improvements in forecasting accuracy and computational efficiency.

Critical strengths
  1. The paper addresses a practically important problem by investigating how external retrieval can improve frozen time-series foundation models without substantially increasing inference costs.

  2. Exploring different retrieval injection strategies is an interesting direction. In particular, latent cross-attention and output-space Wasserstein barycenters provide alternatives to computationally expensive context concatenation, while covariate alignment and text grounding enable the incorporation of additional information.

  3. The experimental results demonstrate promising forecasting performance across multiple benchmarks and foundation model backbones, with improvements over existing approaches.

Critical weaknesses
  1. The limitations under unseen distribution shifts are not sufficiently discussed. While retrieval from external memory can improve forecasting, its effectiveness may be limited when relevant historical examples are unavailable or historical patterns no longer hold. A discussion of these limitations would better clarify the applicability of the proposed framework in real-world forecasting scenarios.

  2. The theoretical assumptions underlying Proposition 2 are not fully consistent with the empirical dimension estimates. Proposition 2 assumes r≪dintr \ll d_{\mathrm{int}}, whereas the reported PCA rank (r=43) exceeds the estimated intrinsic dimension (dint=12.49d_{\mathrm{int}}=12.49). This also reverses the stated relationship between their corresponding convergence exponents. The authors should clarify how the theoretical assumptions relate to the empirical estimates.

  3. The choice of key hyperparameters is not sufficiently justified. The paper sets the alignment loss weight to 1.0 and the text retrieval weight to 0.5 without explaining the rationale behind these choices. A brief justification would improve the methodological clarity.

Questions for authors
  1. The limitations under unseen distribution shifts are not sufficiently discussed. While retrieval from external memory can improve forecasting, its effectiveness may be limited when relevant historical examples are unavailable or historical patterns no longer hold. A discussion of these limitations would better clarify the applicability of the proposed framework in real-world forecasting scenarios.

  2. The theoretical assumptions underlying Proposition 2 are not fully consistent with the empirical dimension estimates. Proposition 2 assumes r≪dintr \ll d_{\mathrm{int}}, whereas the reported PCA rank (r=43) exceeds the estimated intrinsic dimension (dint=12.49d_{\mathrm{int}}=12.49). This also reverses the stated relationship between their corresponding convergence exponents. The authors should clarify how the theoretical assumptions relate to the empirical estimates.

  3. The choice of key hyperparameters is not sufficiently justified. The paper sets the alignment loss weight to 1.0 and the text retrieval weight to 0.5 without explaining the rationale behind these choices. A brief justification would improve the methodological clarity.

Reviewer_BQ6xOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper address the efficiency bottleneck and multimodal alignment in retrieval-augmented time series foundation models. The proposed method combines latent cross-attention and output-space barycenter fusion, linearizing the computational complexity to O(k)O(k). The efficiency is validated through stress testing, and forecasting gains are demonstrated vias CRPS improvements across 30 tasks on fev-bench. Although the paper targets a practical problem, its core methodological novelty is limited. Further, the manuscript suffers from excessive formulation and term inflation that are unnecessary for the core method, making the paper hard to follow and reducing transparency.

Critical strengths

This paper targets a critical efficiency bottleneck in retrieval-augmented time-series foundation models, and the proposed latent cross-attention and output-space injection reduce the latency and memory usage significantly.

Critical weaknesses
  1. Limited novelty and weak motivation. The proposed method seems to be a combination of several existing engineering tricks, and the motivation for combining these specific components is unconvincing.
  2. The writing is difficult to follow. This paper introduces heavy mathematical notion and formal propositions that mostly over-complicate basic statistical operations, while their relationship with the core motivation and model design is not adequately discussed. The literature review overlooks key prior work on time-series manifold learning, despite the paper’s heavy reliance on these geometric concepts. Additionally, the figures are overly text-dense and uninformative. The experimental section lacks qualitative forecast visualizations (e.g., prediction vs. ground truth curves), leaving the empirical evaluation incomplete and difficult to verify.
  3. Excessive academic over-packaging. Simple implementation details are repeatedly wrapped in inflated jargon. For example, concatenating features via MLPs in termed “covariate-conditioned”. This excessive term inflation increases cognitive. The paper would benefit significantly from a major re-organization focused on clarity and transparency.
Questions for authors
  1. Please clarity the core novelty, and explain why the heavy formulation in Section 3.2 are necessary for the proposed model.
  2. Real-world databases contain heterogeneous sub-manifolds. Does the empirical scaling exponent show notable variance across different domains?
  3. Output-space quantile averaging assumes unimodal distribution families. How does output barycenter fusion perform under extreme multimodal future scenarios compared to latent-space cross-attention?
Reviewer_QRqfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents a retrieval framework for time-series foundation models that incorporates historical examples through cross-attention or output-space aggregation. It also uses known future covariates and textual context to guide retrieval.

Critical strengths
  • Addresses a relevant problem: using additional historical information without putting all retrieved sequences into the model’s input context.
  • The approach makes sense, particularly the use of known future covariates to find more relevant historical examples.
  • Includes experiments across multiple forecasting backbones and reports substantial improvements in forecast accuracy.
  • The comparison between incorporating retrieved information through cross-attention and combining forecasts in output space is interesting.
Critical weaknesses
  • The novelty needs to be explained more clearly. Cross-attention is already used to incorporate external representations without expanding the main input sequence, including in related time-series work. I see value in the proposed combination, but it is difficult to identify what is new in the mechanism itself. A more direct comparison with the closest related approaches would help.
  • It is unclear how much of the improvement comes from retrieval. The method adds supervised adapter training, covariate information, and retrieval together. Comparing against the same trained adapter and covariates without retrieval would help separate these effects. I would also like to see a stronger covariate-aware baseline.
  • The evaluation does not make it clear enough when the approach helps. The aggregate results are encouraging, but a breakdown across tasks would show whether the improvements are consistent or concentrated in particular settings. Lines 414-415 refer to a task-wise breakdown in Appendix C, but I could not find those results there.
  • The training and evaluation setup needs a clearer explanation. The adapters are trained on series from the same task collection used for evaluation, although the paper describes separate entities and time periods. The paper should make clear what is unseen at evaluation time and what “zero-shot” means in this setting.
  • The paper is difficult to follow in places. Figure 1 includes too many components and results to provide a clear overview. A simpler figure and one concrete example of the retrieval and forecasting process would help.
Questions for authors
  • What is the main methodological difference from the closest attention-based retrieval approaches? Which comparison most clearly demonstrates this contribution?
  • How much improvement remains when using the same trained adapter and covariates without retrieval? Are there settings where retrieval provides little benefit or reduces accuracy?
  • What explains the advantage of output-space aggregation over latent cross-attention, and when would you recommend each? It would also help to clarify which method and configuration Table 9 represents: its ETTh1 CRPS of 0.5912 differs from the values reported in Table 8?
  • What is precomputed for the efficiency experiments, and what is included in the measured runtime? In particular, how are the candidate forecasts for output-space aggregation obtained?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.