acceptodds
Withdrawn from ICLR 2027

Asynchronous Multimodal Candidate Reweighting for Multi-Horizon Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multi-horizon prediction of real-world systems often requires reasoning over multiple candidate futures using heterogeneous observations acquired at different times. These candidates may be produced by physical simulators, learned predictors, or ensembles, yet existing fusion methods commonly assign each candidate a single reliability score that is reused across all prediction horizons. This assumption becomes restrictive as candidate futures diverge over time and the relevance of asynchronous evidence changes with the prediction target. We introduce Asynchronous Multimodal Candidate Reweighting (AMCR), a general framework that estimates reliability jointly over candidate identity and target time. Rather than synchronizing observations to a common clock, AMCR preserves their native timestamps together with modality, availability, and quality, and constructs target-conditioned contexts through modality-aware temporal relevance weighting. The resulting candidate-time reliability matrix enables the same candidate bank to be reweighted differently across prediction horizons. We establish a bounded-influence result showing that the effect of temporally distant evidence decays under the proposed temporal gate, together with a sufficient condition for candidate-selection stability. We evaluate AMCR under controlled timestamp perturbations, missing modalities, evidence-quality variation, and horizon-dependent candidate changes, and further instantiate it for physics-based oil-spill evolution using satellite imagery, vessel trajectories, environmental forcing, and simulator-generated futures.

Primary area:
learning on time series and dynamical systems
Keywords:
Multimodal learning, Physics-based simulation, Temporal alignment, Multimodal fusion

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_br34OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors use asynchronously arriving multimodal evidence to assign weights to a set of candidate futures along candidate*horizon, claiming to solve the problem that previous fusion methods reuse a single reliability score across prediction horizons. But the data source is not clearly explained, the data is not reproducible, and all the comparison experiments use self-made weak baselines, so I think the quality is not up to standard.

Critical strengths
  1. Expanding candidate reliability from a fixed score into a candidate*horizon matrix, with separate weights per horizon, fits the way time series are modeled and has practical value.
  2. It integrates satellite imagery, AIS, environmental forcing, and simulators on a real oil spill scenario, which is a meaningful contribution in constructing a real-world setting.
Critical weaknesses
  1. A lot of information about the data is missing, so I cannot judge whether it is real. The source of Yh⋆Y_h^\star is never explained, the data is private and not reproducible, and there are no event counts, no splits, and no variance.
  2. The baselines the paper compares against are really bad. None of them come from existing literature; they are all self-made comparison groups. And the existing literature shows this is not because no methods exist, but because the authors did not compare with them. None of the 5 baselines is a method from the literature, they do not even compare with the work they themselves cite, and they report ECE/Brier without comparing to any calibration method, so the gains do not prove anything.
  3. The abstract mainly advertises asynchronous and four types of controlled experiments, but none of them were actually done. Asynchrony in the method shows up only as a single time decay factor, so the title is just decoration.
Questions for authors

I find this research very unreliable, and nothing in the paper convinced me otherwise. That is the overall impression behind my three questions.

  1. How many events are in the training/validation/test sets? Where does Yh⋆Y_h^\star actually come from? Please provide a complete description that makes the dataset reproducible, otherwise I cannot judge whether these results are real.
  2. Why is none of the 5 baselines a method from existing literature? There are plenty of ensemble weighting and calibration methods out there. Did you not look for them, or did you not want to compare with them? Please explain how these baselines were chosen.
  3. The abstract says experiments about timestamp perturbations, missing modalities, evidence-quality variation, and horizon-dependent candidate changes. Where are those experiments? If asynchrony is just one time decay factor in a formula, then what does "Asynchronous" in the title actually mean?
Reviewer_bfb3OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper identifies several common problems in existing prediction methods, such as how to align asynchronous observations, how to handle differences in multimodal evidence quality, and how to give each modality a “judge” to assess its reliability. More importantly, it further explores how far into the future different evidence can support predictions. The abstract also claims experiments on timestamp perturbations, missing modalities, and evidence-quality variation. Although these experiments are not presented later in the paper, the overall architecture seems reasonably designed. The ablation studies of the proposed components also show substantial effects and clear performance gains. Overall, both the framework and the claims it puts forward are interesting.

Critical strengths

The paper uses established components to address long-standing problems and introduces a novel perspective on how far into the future evidence can support predictions. Based on the evidence’s temporal relevance, quality, and relation to each candidate, the model determines how much influence that evidence should have when predicting outcomes at 24, 48, or 72 hours. This is important because it may change how we approach multimodal model design, especially the design of mechanisms that determine the weights assigned to multimodal evidence.

Critical weaknesses

The experiments listed in the abstract are not presented later in the paper. There are no experiments on timestamp perturbations, missing modalities, or evidence-quality variation. The baselines do not include models proposed by other researchers, and the related work survey is incomplete and overlooks existing work. I am concerned that this could mislead readers into thinking that this direction has not been studied before. The studies cited in the related work section are indeed relevant, but the paper does not compare against their methods, which I find puzzling.

Questions for authors

Dear authors, could you complete the experiments? Based on the current experimental results, I cannot tell whether the proposed components are actually useful, which makes it difficult for me to assess the paper.

Reviewer_ABxvOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a candidate-preserving framework for multi-horizon prediction under asynchronous, heterogeneous multimodal evidence, applied to oil-spill evolution forecasting. AMCR retains a persistent bank of candidate futures and learns candidate–time reliability weights, allowing the same candidate to contribute differently across horizons. The method separates causal evidence availability from temporal relevance, performs relation-aware signed reasoning over supportive, contradictory, and unresolved evidence, and normalizes scores horizon-wise with stability regularization. Experiments show consistent gains over single-simulator, ensemble, uncertainty-weighted, and naive fusion baselines.

Critical strengths

S1: The paper identifies a meaningful limitation of assigning one reliability value to an entire multi-horizon candidate trajectory. Modeling reliability over candidate–time pairs is a natural and practically useful formulation when persistent hypotheses diverge across horizons. The explicit separation of occurrence time from observation/availability time is also conceptually valuable for genuinely asynchronous multimodal evidence.

S2: AMCR provides a coherent candidate-preserving interface that can, in principle, sit on top of physical simulators or learned candidate generators. The relation-aware support/contradiction/ignorance mechanism is reasonably interpretable, and the AMCR-Static and no-contradiction ablations are aligned with two central design choices. The paper also evaluates several complementary aspects—spatial overlap, localization, ranking, rejection, and calibration—rather than relying on a single metric.

Critical weaknesses

W1: The empirical evaluation is not sufficiently specified or independently grounded to support the reported conclusions. The paper does not state the number of events, train/validation/test sizes, temporal/geographic coverage, modality availability statistics, or how the “probabilistic reference” fields and candidate-ranking/rejection labels are constructed. In particular, it is unclear whether the 24/48/72 h references are genuine later observations or simulator-derived targets related to the same pipeline that creates the candidate bank. The paper says that an oracle separately measures candidate-bank coverage and later refers to an “oracle analysis,” but no oracle result is reported. The comparison set is also limited to a single simulator, hand-designed ensembles, and an underspecified naive fusion model. It omits the upstream PVD-AGN weighting described in the appendix and competitive learned temporal/candidate-fusion baselines.

W2: The claimed causal forecasting protocol appears to permit future information relative to the forecast origin. For a prediction initialized at T0, Eqs. (3), (6), and (42) admit every item with \tau_i <= t_h, including observations that arrive after T_0. Consequently, the 48 h prediction may use evidence unavailable when the forecast was issued, and it may use more evidence than the 24 h prediction. This is valid for rolling assimilation/nowcasting only if predictions are explicitly re-issued as evidence arrives. It is leakage for a fixed-origin 24/48/72 h forecast. The experiments do not disclose the evidence cutoffs or compare against a strict \tau_i <= T_0 setting, so the headline advantage of horizon-dependent weighting cannot presently be separated from access to later information. Moreover, the theoretical claims do not resolve this issue: Eq. (26) largely follows from inserting the temporal gate into the attention logit, and Eq. (32) is the generic argmax margin condition, without a derived perturbation bound for AMCR.

Questions for authors

Q1: How are support/contradiction/ignorance, per-horizon candidate-order, rejection, and calibration targets obtained? Please provide the complete training objective and all loss weights, and clarify whether any of these targets are computed from the evaluation reference in a way that could expose future labels to the model.

Q2: Can the authors compare AMCR with (i) the one-shot PVD-AGN reliability weighting described in the appendix, (ii) a capacity-matched learned per-horizon fusion model without the proposed signed graph, and (iii) an established asynchronous/temporal multimodal fusion baseline, all under identical evidence cutoffs? These comparisons are needed to attribute the gains to AMCR rather than model capacity, later evidence, or split variance.

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

Related papers

Discussion (0)

Sign in to comment.