acceptodds
Under review as a conference paper at ICLR 2027

Plan Globally, Act Locally: Adaptive Segment Forecasting for Long-Horizon Human Mobility

Abstract

Long-horizon human mobility forecasting must predict a user's future movements over multiple days without accumulating errors from repeated short-horizon rollouts. Human mobility has a natural hierarchy: people make high-level plans across days, while their movements appear as fine-grained location choices over time. Following the principle of planning globally and acting locally, we propose PlanCAD (Plan–Critic Adaptive Diffusion) for long-horizon human mobility forecasting. PlanCAD uses daily plan tokens to capture long-horizon structure and segment-level generation to produce slot-level location predictions. Specifically, PlanCAD uses a large language model (LLM)-based mobility backbone to encode historical trajectories into day-level representations, and a learnable planner generates future daily plan tokens from these representations. A critic scores candidate execution horizons using plan-level and trajectory-level statistics, deciding whether to predict a longer segment or replan earlier. Given the selected plan, a diffusion actor refines future mobility representations in the LLM hidden space and maps them to location predictions. Experiments on real-world mobility datasets show that PlanCAD improves 7-to-7 Acc@3 on Kumamoto from 50.09% to 51.10% over the strongest baseline, with consistent gains on Sapporo and Hiroshima. Ablation studies and analysis further show that complementary critic features support adaptive horizon selection, while plan-conditioned generation improves multi-day forecasting stability.

Primary area:
learning on time series and dynamical systems
Keywords:
Human Mobility Forecasting, Long-Horizon Forecasting

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

This paper introduces PlanCAD to address error accumulation in autoregressive multi-day mobility forecasting and the inflexibility of fixed prediction windows. PlanCAD combines day-level plan tokens, a critic that adaptively selects execution horizons of one to three days, and a diffusion actor that generates half-hourly location predictions in the latent space of a frozen LLM. In the seven-day forecasting setting, PlanCAD raises Acc@3 from 50.09% to 51.10% on Kumamoto, from 48.90% to 50.42% on Sapporo, and from 50.18% to 50.94% on Hiroshima, corresponding to improvements of 1.01, 1.52, and 0.76 percentage points over the strongest baselines.

Critical strengths
  1. The paper addresses the practically important tension between error accumulation in day-by-day autoregressive forecasting and the inability of fixed multi-day windows to adapt to different contexts. By framing long-horizon mobility forecasting as the problem of deciding when to replan, the work focuses directly on the inference process rather than merely replacing the forecasting backbone.

  2. The proposed framework coherently combines day-level planning, adaptive execution-horizon selection, and slot-level location prediction under the principle of “planning globally and acting locally.” This design is supported by competitive long-horizon results: PlanCAD-Full achieves the best performance in 8 of the 12 city, forecasting-horizon, and accuracy-metric combinations in Table 1, including the best Acc@1 and Acc@3 results for all three cities in the 7-to-7 setting.

Critical weaknesses
  1. The critic is supervised using the oracle segment-length label �0��0�, which minimizes only the average cross-entropy over the current candidate segment and does not account for downstream error propagation, complete seven-day rollout quality, or computational cost. Because �1��1� is recomputed from the actor’s current losses and may therefore change as the actor is updated, while the actor itself is trained only on the selected segment, the training procedure may produce unstable winner-take-all behavior or horizon collapse. This concern is reinforced by the ablation results: ORACLE does not clearly outperform the full model or FIXED-1, while the hybrid critic is not consistently better than the plan-only, trajectory-only, or horizon-only variants.

  2. The parameter-matched MATCHED-NO-DIFF variant outperforms the full model on Acc@1, Acc@3, Acc@5, and MRR under standard test conditions, while the diffusion module shows only limited benefits under artificially imposed strong latent-space perturbations. Moreover, inference uses a single denoising step at (t=0) without an iterative reverse-diffusion chain, and the paper does not adequately explain how this procedure fundamentally differs from stochastic residual refinement or why diffusion modeling is indispensable.

Questions for authors
  1. How frequently does the critic select each segment length ((k=1,2,3)) at test time? What are its agreement rate with the oracle, selection regret, and complete seven-day rollout loss? If the oracle itself does not outperform FIXED-1, how can the authors demonstrate that deciding “when to replan” is a primary source of improvement?

  2. Why is the critic supervised using the average cross-entropy over the current segment rather than a full-horizon objective that accounts for subsequent feedback effects and computational cost? Because this moving oracle is generated from the current actor’s own losses, does training exhibit candidate-horizon collapse or frequent switching between candidate horizons?

  3. At inference time, the method fixes (t=0), performs only one denoising step, and does not use a reverse-diffusion chain. What indispensable benefit does the diffusion module provide over a parameter-matched deterministic residual decoder? Can the authors demonstrate consistent gains under naturally occurring representation errors rather than artificially injected Gaussian perturbations?

  4. Please clarify whether the 40k-location vocabularies for the three cities are shared or constructed separately, how overlapping sliding-window samples are generated, whether each user appears across the training, validation, and test splits, and whether the prompt-derived embeddings containing anonymous user IDs and coordinates are generated entirely locally without transmitting the data to third-party services.

Reviewer_j2FnOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes PlanCAD, a Plan-Critic-Act framework for long-horizon human mobility forecasting. It integrates day-level planning, adaptive forecasting horizons and diffusion-based refinement to improve prediction accuracy and efficiency. However, the reported improvements are generally marginal, and their statistical reliability has not been adequately established. Furthermore, the ablation studies raise concerns about the effectiveness of the adaptive horizon selection mechanism and the necessity of diffusion-based refinement. Whilst the proposed direction is promising, I do not believe the current experimental evidence is sufficient to justify acceptance at ICLR (in other words, if it were a 5-point scale, I’d probably lean toward a 2).

Critical strengths
  1. The motivation for adaptive forecasting horizons is interesting and addresses a relevant challenge in long-horizon human mobility prediction.
  2. The paper demonstrates a reasonable understanding of the relevant literature and positions the proposed framework within existing research.
Critical weaknesses
  1. The reported gains are marginal, ranging from -0.86 to +1.70 points compared with the strongest baselines, with an average improvement of only 0.44 points across the 12 metrics in Table 1. However, the authors omit the actual standard deviations (only stating ≤2%) and provide no statistical significance tests. Given the small improvements and only three independent runs, the reliability of these gains remains questionable.
  2. In Table 2, FIXED-1 achieves comparable accuracy and even slightly outperforms PlanCAD-Full on Acc@1. The ORACLE variant also fails to consistently outperform FIXED-1, raising questions about the effectiveness of adaptive horizon selection and the suitability of its supervision signal. Moreover, MATCHED-NO-DIFF outperforms PlanCAD-Full on several key metrics. Although the authors report some robustness benefits, the necessity of diffusion for the primary forecasting task remains unclear.
  3. In Table 3, the full hybrid critic does not consistently outperform simpler feature configurations. In particular, the horizon-only critic achieves comparable performance without using trajectory or plan features, raising questions about the added value of the proposed hybrid feature combination.
Questions for authors
  1. Could the authors provide the actual standard deviations and statistical significance tests for the results in Table 1? Which improvements over the strongest baselines are statistically significant?
  2. Why does the ORACLE variant fail to consistently outperform FIXED-1? Does the oracle supervision objective, based on segment-level cross-entropy, adequately reflect the quality of the final multi-day forecasts?
  3. Given that MATCHED-NO-DIFF outperforms PlanCAD-Full on several key metrics, what concrete benefits does diffusion provide under standard, unperturbed forecasting conditions?
  4. What measurable benefits do the additional plan and trajectory features provide over the horizon-only critic? Are these differences consistent across datasets and statistically significant?
Reviewer_HeXaOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces PlanCAD, a framework that predicts long-horizon human mobility. It combines a day representation from an LLM and a planner that can generate plans using the said day representation. A separate critic is trained to score the horizons to decide if a longer or shorter plan is viable. Another diffusion component then maps the LLM-based representation into location predictions. The method reports a consistent but small improvement over existing baselines.

Critical strengths
  1. Adaptive planning is an interesting approach. Most existing appraoches assumes a fixed prediction step. However, the most optimal strategy is likely to be dependent on how predictable the current context is. This is an idea that could transfer to other prediction tasks as well.
  2. The separation of the global representation and the more fine-grained slot level representation provides a reasonable inductive bias. This aligns well with the intuition about mobility data. As the authors stated, the day routines are likely regular but individual choices are much more variable.
Critical weaknesses
  1. I'm a little unconvinced by the diffusion actor. The paper states that the actor 'refines future mobility representations' and 'maps them to location logits'. Appendix C.3 describes this process in a lot of mathematical notations but seems to be a single denoising step during inference after injecting gaussian noise. This sounds much more like regularization strategy and and doesn't seem to justify a diffusion model. The ablation results in Table 2 also suggests that the model without diffusion (matched-no-diff) outperforms the full model on Acc@1, Acc@3, Acc@5, MRR, and DTW.
  2. The authors introduced the critic, which is supposed to decide which candidate plan (long or short horizon) should be adopted. However, section 4.5 defines the oracle horizon to be 'with the lowest segment-level loss' (line 268). This seems to contradict with the objective that is to reduce long-horizon error accumulation. By selecting the oracle with the lowest immediate loss, it doesn't necessarily minimize the overall total error. In fact, a longer plan might have slightly worse immediate accuracy but still reduce error propagation down the line. It is also unclear how the candidate plans are evaluated. Appendix C.4 states that the oracle minimizes the average cross entropy. Does this mean that the candidate plans are evaluated over different day intervals? One day plan candidate is evaluated on the first day, whereas three-day plan is evaluated on all three days. Wouldn't this systematically penalize longer horizon candidates?
  3. Sections 3 and 4 (and the appendix) could benefit from a more concise set of mathematical notations. For example, some intermediate transformations that are not fundamental to the model introduces significant difficulty when reading and remembering what they are. They sometimes confuses more than they explain. The writing of the result section could also benefit from having more discussions and insightful explanations. The current draft can be overly descriptive and repetitive at places, especially when the numbers are already presented in tables and figures.
Questions for authors
  1. Clarification of my concern of Weakness point 1.

  2. Clarification of my concern of Weakness point 2.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.