(-) My biggest concern is the most naive baseline of simply freezing the input for 5 frames -- what the authors refer to as "persistance". I searched for a proper comparison with it. It appears once in table 2 but that table compares only deterministic regressors which aren't the main proposed method. More broadly, Table 2 contains learned forecasting baselines but excludes RadarDreamer itself. Thus, it does not establish how the proposed generator performs against these predictors under a common evaluation protocol. And then again in table 12 in the appendix where the naive persistence model actually wins over the proposed method in most settings on MAE. RadarDreamer improves CRPS across all four domains, however, average sample MAE worsens in both evaluated modalities for three domains. RaDelft improves on cropped support, while its full-grid RA error worsens. In particular I did not appreciate that Figure 3 shows only gains of the method in MAE. This seeming contradiction between the table and figure pointed to the phrasing in the appendix about the selection of the "displayed anchor" which to my understanding is another way to say Figure 3 shows the validation examples with the overall best result. I ask that the authors replace this visualization with a general case result, or clearly state that this is cherry picked. I also stress the importance of adding persistence as proper baseline in the full evaluation protocol rather than the reduced 32 samples of table 12. Without it, there's too little evidence the generator is doing real work beyond a mere frame duplication baseline.
(-) Those methods referred to as baselines aren't described sufficiently and it is very difficult to understand what exactly they do. Sections 4.1 and 4.2 interpret results before adequately introducing the compared models and experimental controls. The preceding paragraph lists model names without explaining their construction, conditioning inputs, or training objectives. Specifically for "independent" my understanding is there is no communication between the RA and DA, but does the model see both inputs from the input frame? Also, I would consider this ablation rather than a baseline. Table 1 indeed is referred from both 4.1 and 4.2.
(-) The authors never really demonstrate any of the applications mentioned in the intro (perception, planning, etc). The chosen setup of predicting future 5 frames from a single frame is also puzzling. In reality, if one assumes an initial radar measurement to condition the generation this rules out "from-scratch" creation of novel scenes (e.g. through text to radar). So the implicit assumption is that there' s access to recorded radar, which brings up the question -- why a single frame rather than a longer past trajectory? Table 2 shows (in a deterministic setting) that if the past frames contain only RA, then a single RA+DA frame works better. but i see no reason why past frames that have both DA and RA won't work even better and why not use them? Without it I doubt the model is useful as a world model. Specifically, world models are expected to produce longer rollouts than 5 future frames so a single frame conditioning would prohibit utilizing past knowledge that could be instrumental to handling targets under temporary occlusions for example.
(-) The ego motion isn't used in the condition however it governs significant parts of the scene doppler. I would like to see a version of the method that uses ego velocity as a condition. This should be readily available in the data.
(-) flow models have a prescribed loss term. However, the one proposed by the authors in 16+17 modifies it which could (and the authors say so themselves) steer it away from the desired target distribution. The paper should discuss and show this trade-off by measuring the flow model performance with and without it.
(-) Prior radar neural fields work are missing including Neuradar, RadarSplat, RadarFields. in the diffusion/flow literature, the recent work RadarGen is missing. Beyond adding these citations, the paper should explain more carefully how forecasting future observations differs from reconstructing scenes and rendering measurements at new viewpoints, identifying the relevant differences in inputs, outputs, and temporal modeling.
(-) The choice of RA+DA is insufficiently justified. Equation 1 derives these projections from the available full RAD tensor, discarding range-Doppler associations. The paper then makes this induced ambiguity central to its motivation (proposition 1), without demonstrating the benefits of that trade-off. A comparison of RA+DA, RA+RD, and full-RAD conditioning, including prediction accuracy and computational cost, is needed. Full-RAD conditioning could also be used while retaining projected outputs, so generating complete future cubes is not the only alternative.
Minor:
(-) the key insight is somewhat underwhelming. "Doppler exposes motion tendency before displacement becomes visible" and that it is a condition for future motion. This is not an insight as much as a statement about what doppler is.
(-) Another concern is data size. looking at the Data section of this work, this is still small scale. The authors report generation on held out sequences but I would want to see held out scenes, not just sequences.
Writing:
(-) the authors seem to utilize some jargon which I as a computer vision researcher was unfamiliar with, like calling deterministic prediction, a point baseline. Similarly, “Gaussian radar video” is confusing wording. The authors appear to mean Gaussian noise shaped like the encoded future sequence. I'm pretty sure this will throw off other readers.
(-) key definitions are lacking. For example, innovation-balanced is barely explained in the main paper but is still used in captions and tables. This makes the paper very hard to follow.
(-) multiple parts of the paper, especially the intro and abstract are quite hard to follow. For example, the whole paper around Corollary 1 is cryptic. Destd isn't explained.
(-) describing a world model without the agent's action is untraditional
(-) The method is tightly coupled to the specific choice of RA+DA and doesn't discuss how particular vs general this choice is. I think this doesn’t belong under writing – moved it to the major weakness
(-) Proposition 1 formalizes the information loss introduced by the RA+DA projections, but its presentation feels unnecessarily elaborate for a straightforward observation. A simple example would communicate the intuition more clearly in the main text, with the formal statement and proof moved to the appendix.