acceptodds
Under review as a conference paper at ICLR 2027

PA-RNet: Perturbation-Aware Residual Network for Robust Multimodal Time Series Forecasting

Abstract

In real-world applications, multimodal time-series forecasting faces a key challenge: textual information is often useful but unreliable. Auxiliary texts may contain irrelevant, ambiguous, incomplete, or structurally corrupted content, making direct text integration prone to introducing noisy semantic signals and degrading forecasting performance. Therefore, robust multimodal forecasting requires a model that can exploit useful textual context while suppressing misleading perturbations.To address this challenge, we propose PA-RNet, a perturbation-aware residual network for robust multimodal time-series forecasting. Before cross-modal fusion, PA-RNet refines numerical features and corrects textual representations to mitigate the influence of perturbations while preserving task-relevant information. It then aligns the refined textual context with temporal patterns in the numerical series, improving forecasting robustness under noisy multimodal inputs.We establish the Lipschitz continuity of PA-RNet with respect to textual embeddings and provide theoretical insights into how perturbation-aware refinement can reduce expected prediction error. We further assess its robustness through supplementary experiments with injected textual perturbations. Results across diverse domains demonstrate that PA-RNet consistently outperforms state-of-the-art baselines while maintaining stable forecasting performance under both original and perturbed textual conditions.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Forecasting, Multimodal Learning, Robustness

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_KoS1OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies robustness to unreliable textual context in multimodal time-series forecasting. PA-RNet combines a gated numerical refinement module, causal attention with low-rank textual residual correction, and cross-modal attention. The paper also presents theoretical analyses of sensitivity to textual perturbations and evaluates forecasting performance across nine backbones under several corruption types and intensities. The problem is practically relevant, and the experimental coverage is substantial. However, inconsistencies in the method description and limitations in the theoretical and empirical evidence make the central robustness claims difficult to assess.

Critical strengths
  • Practical relevance. The paper addresses unreliable textual context in multimodal forecasting, an important challenge when auxiliary text is noisy, incomplete, or misleading.
  • Broad experimental coverage. Evaluation across nine forecasting backbones, multiple corruption types and intensities, text-cleaning baselines, and masking conditions supports the method’s compatibility and practical investigation.
Critical weaknesses
  • Inconsistent method descriptions. The architecture, equations, ablations, and hyperparameters contain conflicting references to spectral/FFT operations and numerical versus textual components, making it unclear what was implemented and tested.
  • Insufficient evidence for improved robustness. The theory assumes conditions not established for the learned model, while the experiments lack clear comparisons of corruption-induced degradation and uncertainty estimates needed to distinguish small improvements from variation.
Questions for authors
  1. What exact modules were implemented and removed in each ablation? Please reconcile the spectral/FFT terminology, numerical-versus-textual branch descriptions, and hyperparameter definitions with the equations.
  2. Can you report uncertainty estimates across the five seeds and paired differences against the strongest baseline?
  3. Are corruption experiments conducted with fixed checkpoints or corruption-specific retraining? Under matched training conditions, how much does each method degrade relative to its original-input performance?
  4. What evidence establishes that the learned textual correction satisfies the assumptions of Proposition 2? If these conditions remain assumptions, can the theoretical contribution be framed more narrowly?
Reviewer_AtBxOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies robust multimodal time series forecasting under unreliable textual information, where auxiliary descriptions may contain irrelevant, incomplete, ambiguous, or corrupted content. To mitigate the negative effects of such perturbations, the authors propose PA-RNet, which combines numerical feature refinement, causal semantic correction, and cross-modal attention to improve the integration of textual and numerical information. The paper further provides a theoretical analysis of the model's sensitivity to textual perturbations and evaluates its forecasting performance under both original and synthetically perturbed textual conditions. Experimental results across multiple datasets and forecasting backbones demonstrate competitive performance compared with existing methods.

Critical strengths

(1) The paper considers noisy and unreliable textual information in multimodal forecasting.

(2) The authors provide a formal analysis of Lipschitz continuity and investigate conditions under which textual representation refinement can reduce prediction errors.

(3) The experiments cover multiple forecasting backbones, datasets, and textual perturbation types and intensities.

Critical weaknesses

(1) The proposed solution mainly combines numerical gating, low-rank residual refinement, causal self-attention, and cross-modal attention, all of which are established techniques.

(2) This work does not explicitly distinguish corrupted textual information from valid semantic content, nor does it introduce a specific objective for identifying or correcting semantic inconsistencies.

(3) Proposition 1 establishes Lipschitz continuity of PA-RNet with respect to textual embeddings. However, the derived bound contains the factor (1+L_Psi), and therefore does not demonstrate that the proposed semantic correction reduces sensitivity compared with an otherwise identical model without correction.

(4) The additive zero-mean noise assumption is not clearly connected to the discrete semantic perturbations considered experimentally. The notation used to compare corrected and uncorrected predictions also appears inconsistent with the earlier definition of the full prediction function.

(5) Several dataset-backbone combinations show no improvement or even worse performance.

(6) The baseline does not adequately cover existing methods. For example, Cross-MoE has investigated model-agnostic text–time-series integration and relevance-based filtering on Time-MMD, while Aurora employs modality-guided attention for multimodal time-series forecasting.

Questions for authors

(1) The manuscript does not clearly identify a novel perturbation-specific mechanism or explain what fundamental limitations of existing fusion methods remain unresolved.

(2) How does the causal semantic correction module distinguish informative textual content from semantically corrupted or contradictory descriptions?

(3) The causal semantic correction module does not employ an explicit semantic correction objective, what evidence demonstrates that it performs meaningful error correction rather than simply learning another representation transformation?

(4) Can the authors provide robustness-specific ablations under multiple perturbation types and intensities, including comparisons with parameter-matched residual adapters and gating mechanisms?

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

The paper addresses problem of training a multi-modality time-series forecasting that can utilise the useful textual context but reduce the perturbation that can be misleading. The proposed network is trained to learnt the task-dependence residual corrections. The experiments are conducted in various of common benchmark in multi-modality time-series forecasting.

Critical strengths
  • Figure. 1 introduces the proposed framework and well-present the approach.
  • Experiments show the improvement in forecasting performance of the approach on average in different datasets.
  • Table 4 is able to show the effects of each components introduced to the framework (PA-RNet)
Critical weaknesses
  • The paper can be improved by re-organising the sub-sections, because the main architectures (which should be presented in section 3.2 - perturbation-aware forecasting framework) seem to be placed in the subsection 3.3-currently is written as theoretical foundations.
  • The perturbation ratio ρ\rho were used in the experiments, but were not introduced in the main approach.
  • Citation format could be improved.
Questions for authors
  • Could the author explain what is the perturbation ratio used in experiments since this quantity has not been described in the approach?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.