In short, I neither understand the problem framework nor details of the solution method. I am not understanding the validation tasks either. I cannot judge the contribution of this paper nor its possible drawbacks. Specific comments are below:

• the paper explicitly positions itself as operating on a POMDP environment, even mentioning it in the abstract. But POMDPs are neither defined nor ever seemingly explicitly used in the paper. The only dynamical model that is introduced is a POMDP only in a very trivial sense: if nothing else, both its transitions and observations are deterministic

• on the topic of the presented dynamical system, the paper calls it a "Structured State Space Model" and says that these are "sequence-to-sequence models well-suited for tasks that require significant memory retention and are particularly effective at processing long sequences due to their computational efficiency". But this model seems to be just a standard Linear Time-Varying (LTV) control system! Control design for LTV systems is challenging, but has certainly been explored since the 1950s or earlier. In general, in the context of agent planning, if these are agent dynamics, I don't see what makes them particularly "well-suited" for any task. After all, the model should not depend on the task: the model is whatever represents the agent's dynamics. If this is simply a learning model, then the agent's dynamics are not ever defined.

• there seems to be a lack of awareness of classical control (let alone planning on POMDPs -- a line of work which is never truly mentioned); the authors call usual linear control systems "time-invariant SSMs" and speak about recent studies that explore the conditioning of system matrices on the input sequence. Again, this is not recent work -- stability of linear systems is a classical introductory control topic

• I do not understand the formal problem that this paper is solving. It does not seem to be ever defined and is mostly just described as "long-term planning". Some questions that come to my mind are: Are there rewards? Is there a reachability task? Does the agent move? What are its dynamics?

• I also do not know what the agent knows about its environment. If its dynamics are just linear and known, I don't understand why any learning is necessary: optimal control laws for reward maximization (at least with a particular reward structure) can possibly be derived analytically.

• the details of the proposed solution approach are murky to me. Let me just give one example. Section 3.3 says that "unlike in MDPs, predicting actions solely from adjacent frames in POMDPs can be unreliable". Doing so is not in fact unreliable, it is theoretically impossible: both in general MDPs and POMDPs, there is no unique mapping from a transition (s,s') to an action a that might have caused this transition. To address this issue (and I am not sure what it means to address it, given that the problem simply does not have a solution), the paper says it will use "Transformer encoders". Why? How does that work?

• I do not understand the tasks, which are never truly described (the paper does not even provide a full name for MNIST) -- agent motion, agent knowledge, possible "long-term" rewards, etc. never seem to be defined. The paper says that "our model extends diffusion-based planning models into the realm of meta-learning", but this topic is never discussed.

Novelty

The paper addresses the long-term horizon problem by applying known sequence processing techniques (RNNs, Transformers, SSMs) to known decision-making models, i.e., Diffusors. The paper often refers to the computational complexity of these memory techniques, but these are well-known facts, e.g., RNNs are sequential during training but fast during inference, while Transformers are parallelizable but scale quadratically during inference. Thus, I consider the main contribution as an application of known techniques rather than technical innovation.

The introduced benchmarks seem interesting but their evaluation lacks a comparison with other approaches, which is necessary to assess their suitability for testing long-horizon planning, i.e., Are they sufficiently difficult? Do other approaches really struggle on these new domains, as stated in the paper? See Significance below.

Clarity

The abstract teases generalization as a problem of existing approaches. However, the main challenges addressed in the paper are only focused on long-horizon planning and computational complexity (during training and inference).

Significance

The experimental evaluation of the paper is a pure self-evaluation with POMDPDiffusors without further context.

The paper does a lot of conceptual comparison with prior works, such as world models and alternative diffusion approaches, such as Diffusion Forcing. However, none of these approaches is compared within the experimental evaluation, which leaves many questions open to assess the significance of the work:

1. How does the POMDPDiffusor fare in traditional POMDP benchmarks like Pocman, Battleship, etc., compared with prior approaches?
2. Do prior approaches really scale that badly, as stated in the paper? We need to see the numbers - not only the words
3. Do prior approaches really struggle in the new benchmark domains, i.e., are the benchmark domains really justified? Again, we need the numbers - not only the words.

Without any further evidence regarding these questions, the true advancement of the work remains unclear.

Minor

• At the end of page for a UNet model is refered to out of nowhere which has not been mentioned and explained before.
• In Related work a reference is missing in "Efficient World Models"

The paper is difficult to understand and the contributions are not quite fleshed out.

As someone who is not particularly familiar with the background (in particular, the "Diffuser", I suppose?), it is difficult for me to infer exactly what the contribution is and how it works. In particular, the text is currently imprecise both in English as well as math. Examples include:

• The proposed method to "model memory" is explained as "through cross-attention computation during the denoising process", and otherwise does not seem to give any details.
• It is claimed that, by "separating memory and planning", the complexity reduces from one O notation to the other, but unclear where these come from.
• The state-space is defined as transforming an input x in R^{T x D} to output y in R^{T x D} (where x and y have the same size) but it is not quite ever really clear what x and y would be for the POMDPDiffuser (most likely due to lack of my background).

This makes me believe (perhaps wrongly) that the proposed method is a combination of supervised learning of state-space models to represent histories and diffusion models to learn policies for these histories from decision data. This, without additional contributions - which may be there but not understood - seem to reduce to behavior cloning on sequences, which is somewhat lacking in novelty.

Lastly, the experimental evaluation does not seem to be very convincing. In particular, there seem to be no baselines, other than ablations on their own approach, and I must assume offline RL methods for POMDPs exist (it was not claimed otherwise in the paper).

1. This paper claims to improve upon SSMs, RNNs, and Diffusers; however, it primarily integrates these models by using an SSM as the memory encoder and a diffusion model for planning with the memory.
2. It does not address the issues associated with transformers as stated in the introduction. The model still relies on a transformer encoder for action selection, predicting the full action sequence from past trajectories.
3. The new evaluation benchmark does not appear to provide enough innovation to be considered a genuinely new benchmark.

• The paper lacks a solid theoretical analysis explaining why diffusion models are suitable for long-memory and long-planning tasks in POMDPs. There is no discussion on the convergence properties, limitations, or potential pitfalls of applying diffusion models in this context. As a reader, I was hoping to see it atleast in the appendix section of the paper.
• The paper acknowledges that the proposed method struggles with more complex tasks but does not delve into why this is the case. I would suggest adding a section/few lines on how it might be addressed in future work.
• The experiments seem to be very simplistic.

(Follow up weaknesses in the Questions section)

• The writing is rushed and a bit disconnected.
• Contributions mainly take the form of the empirical results presented, comparing different configurations of known techniques, without new theoretical/algorithmic insights.
• Given the status of the writing, it's difficult to appreciate the empirical results without significant effort - I'm reading the experiments section without fully understanding the methodology and I have to keep going back to the (rushed) prior sections.
• I seems unlikely those serious issues with the presentation can be addressed without a major revision.