acceptodds
Under review as a conference paper at ICLR 2027

FutureWay: Adaptive Future-Informed Waypoints for Long-Horizon Diffusion Prediction

Abstract

Long-horizon diffusion prediction aims to forecast how information spreads on social networks, which requires resolving local ambiguity and correcting trajectory drift during autoregressive rollout. Yet existing methods primarily learn next-participant prediction on ground-truth histories, leaving them poorly equipped to correct from rollout errors. To address this challenge, we propose FutureWay, a framework that turns recorded futures into decision-specific guidance on model-generated histories. Future information can help distinguish locally plausible participants and correct decisions after rollout deviation, but the useful guidance horizon depends on the current diffusion state. During training, multi-horizon future-informed teachers provide complementary corrective signals. A prefix-only projection analysis characterizes the optimal approximation to future-informed guidance using observable history, motivating student-aware horizon weighting that balances corrections beyond a pretrained anchor against compatibility with the evolving student. Representation-aware on-policy self-distillation transfers this guidance into student decisions and representations. At inference, FutureWay uses only the current cascade prefix and social graph without future information. Experiments on three real-world benchmarks show consistent improvements in both long-horizon future-participant coverage and order-sensitive trajectory recovery, with an average relative gain of 10.56% over the strongest comparators.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Social Networks, Social Diffusion, Public Opinion, Future-Informed Guidance

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

The paper studies autoregressive prediction of information cascades: given users who have already participated and a social graph, predict the next participants. Standard next-participant training uses ground-truth historical prefixes, which will cause a distribution mismatch since inference is conditioned on imperfect model predictions.

FutureWay addresses this mismatch by generating student rollouts and constructing corrective training examples. The earliest historical participant missing from a generated history becomes the corrective label. Later recorded participants provide privileged inputs to teachers with different look-ahead horizons. Their predictions are adaptively combined and distilled into the student through prediction and representation losses. Experiments on Twitter, Weibo, and Quora show improvements in future-participant coverage and order-sensitive trajectory recovery.

Critical strengths
  • The paper studies a fundamental problem of error accumulation in social opinion diffusion.

  • The experiments are validated on real social network data and show improvements by mitigating the distribution mismatch between training and test time.

Critical weaknesses
  • The following claim seems not true: “Second, treating multiple future participants as positive targets encourages coverage of future participants rather than discrimination of the exact next participant.” If the model can perfectly learn pθ(vt+1∣C^t,G)p_\theta(v_{t+1} \mid \hat{C}_t, G), it should be able to predict the exact next participant. Cascade data contain temporal order that can be preserved through ordered multi-step supervision.

  • The term “guidance” is a bit misleading. The paper also does not properly define what guidance means in social network opinion diffusion. If I understand correctly, it is nothing other than teacher-provided training supervision. However, the term can be confused with the guidance used in diffusion models.

  • The last paragraph of the introduction is difficult to follow because so much abstract jargon is introduced, including “recorded futures,” the sudden appearance of a teacher–student model paradigm, “representation-aware on-policy self-distillation,” and so on.

  • The proposed idea is clever and starts from a good motivation to solve the mismatch between training and testing conditions, but the writing makes it hard for the audience to follow. I think a simple figure or a toy example demonstrating the method would make it clear enough.

  • The motivation for why we need a teacher model is unclear. Why can't we directly minimize $ \mathcal{L} = -\log p_\theta(C \mid \hat C, G)$ to learn the correct answer? The paper provides an ablation study on this, but the gains from using a teacher seem marginal.

Questions for authors
  • Can the authors provide some intuition for why the NLL does not decrease when guidance from a longer horizon is provided, given that one would expect more information to be introduced? Is it dataset-related?

  • In the experiments, “each dataset uses a fixed TEST cohort containing sufficient future participants for its largest horizon.” Does this mean we need to know the complete historical cascades to apply this algorithm? In reality, forecasting has no fixed horizon. Is the algorithm still applicable?

Reviewer_8GBTOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes a next-participant prediction method that incorporates future information during training, in addition to the observed past history. It trains multiple teachers using different future horizons and uses their output distributions to guide the student in correcting prediction errors during long-horizon trajectory prediction.

Critical strengths
  1. The paper presents an important challenge in long-horizon prediction, where early errors can accumulate and affect subsequent predictions.

  2. The paper provides reasonable experiments on three real-world datasets.

Critical weaknesses
  1. Insufficient experimental details for Figure 2. The introduction presents Figure 2 as empirical motivation, but the experimental setup is not clearly described. The authors should clarify whether the results are obtained from real-world data or simulations and provide essential experimental details, at least in the figure caption.

  2. Missing supporting references (Lines 78–82). The authors make several claims about the limitations of existing approaches without providing sufficient supporting evidence or citations. Please clarify the basis for these claims and include relevant references.

  3. Unclear contributions in the introduction (Lines 95–107). The final paragraph of the introduction contains several technical terms (terms in italics) without sufficiently explaining their meaning, making the proposed approach difficult to follow. I suggest simplifying this paragraph and clearly listing the paper's main contributions to improve clarity and highlight its novelty.

  4. Insufficient Justification for Design Choices. The paper claims that future-informed teachers provide additional guidance for correcting student prediction errors. However, since the ground-truth future sequence is already available during training, it is unclear whether the benefits of future-informed teacher predictions justify the complexity of the proposed framework. Is it necessary to incorporate future trajectory information through multiple teachers, or could simpler approaches achieve similar benefits? The authors should better justify this design choice. Could the observed improvements arise from other effects, such as soft-label regularization, rather than meaningful future-informed guidance?

  5. Ablation with prefix-only Teacher. Related to Weakness 4, do the authors isolate whether the performance gains come specifically from future-informed teacher guidance? In particular, have they compared against a prefix-only teacher under otherwise identical settings and simpler alternatives such as label smoothing?

Questions for authors
  1. Could the authors provide a more intuitive explanation of how proposed future-informed teacher predictions help correct student prediction errors? Although the paper presents the mathematical formulation, I believe the authors could do a better job of explaining the underlying intuition and the specific roles of each component. Currently, it is difficult for readers to understand how these components work together to achieve the claimed benefits and to assess the individual contribution and significance of each component.

  2. Why do the authors choose to distill only the teacher's output predictions rather than directly matching teacher and student intermediate representations? Is there a specific motivation or empirical justification for this design choice?

  3. Suggestion: Figure 3 is dense and difficult to follow. It appears to visually reproduce the methodology description rather than provide meaningful intuition about the key steps and how the components work together.

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

This paper proposes FutureWay, a framework for long-horizon information diffusion prediction that addresses error accumulation during autoregressive forecasting. The method introduces multi-horizon future-informed teachers to provide corrective supervision on student-generated histories, student-aware horizon weighting to adaptively combine teacher guidance, and representation-aware on-policy self-distillation to transfer future information into a prefix-only predictor.

Critical strengths
  • The paper identifies an important limitation of conventional next-participant prediction: autoregressive forecasting accumulates errors because models are primarily trained on ground-truth histories rather than their own generated trajectories.
  • The proposed framework provides a reasonable approach to transferring future-informed guidance into a predictor without requiring future information at inference time. The combination of multi-horizon teachers and representation-aware distillation is technically interesting.
  • The authors evaluate multiple datasets and prediction horizons, report results across five random seeds, and include comparisons against fixed-horizon teachers, uniform mixtures, learnable mixtures, and different distillation objectives. These experiments provide useful evidence for the proposed design.
Critical weaknesses
  • Theorem 1 establishes a standard forward-KL projection identity, while the state-dependent horizon advantage follows from a general expectation–maximum inequality. These results characterize properties of future-informed distillation but do not establish that the proposed student-aware KL weighting improves long-horizon prediction. In particular, under ideal teachers sharing the same underlying data distribution, the prefix-only projections coincide across horizons. The connection between the theoretical results and the effectiveness of the specific weighting rule therefore remains insufficiently established.
  • The method constructs corrective targets by removing already generated participants from the ground-truth cascade and selecting the earliest remaining participant. Although this provides a well-defined recovery objective, a model-generated trajectory that has deviated from the true cascade may correspond to a different diffusion state. It is unclear whether forcing the model to recover missing ground-truth participants reflects valid propagation dynamics or primarily optimizes the retrospective participant-coverage metric.
  • The proposed weighting rule favors teachers whose distributions are better aligned with the evolving student than with the pretrained anchor. However, this may reinforce already learned predictions rather than prioritize genuinely corrective guidance. While the ablation results support adaptive weighting, the paper does not directly establish that the selected weights correlate with actual horizon-specific corrective utility. A comparison against oracle horizon assignments and alternative weighting criteria would clarify the mechanism.
  • The evaluation defines the observed prefix as the first 30% of the complete ground-truth cascade, implicitly relying on the final cascade length, which is generally unknown during real-time forecasting. Moreover, the fixed-horizon evaluation retains only 35.76% of Twitter and 43.91% of Weibo test cascades. The reported results therefore characterize retrospective continuation on selected cascades rather than fully realistic online diffusion forecasting.
Questions for authors
  1. Can the authors provide a more direct theoretical or empirical connection between the student-aware weighting score and improvements in long-horizon prediction performance?

  2. How does the proposed correction mechanism behave when the generated trajectory deviates substantially from the ground-truth cascade? Does recovering the earliest missing participant always represent a meaningful next action?

  3. How closely do the learned horizon weights match the horizon-specific corrective utility measured by a target-aware oracle?

  4. Can the authors evaluate under an online observation protocol that does not rely on knowing the final cascade length, and report results on a broader test cohort?

  5. How does FutureWay compare with alternative on-policy distillation or imitation-learning approaches under matched training computation?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.