- 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
-
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.
-
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
-
The critic is supervised using the oracle segment-length label , 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 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.
-
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
-
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?
-
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?
-
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?
-
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.