Post rebuttal update: I appreciate the additional explanation for need of overshooting in empirical methods, and the clarity of response regarding stochastic models. The issue I took was with Sec 2.2, that next-step prediction is insufficient to produce belief states, which is only an issue with approximation error when dealing with empirical results. This is not clearly explained in the paper, but clarified much more nicely in the rebuttal. This would cause me to raise my score from a 3 to a 4 for the misunderstanding, but I still do not find this paper worthy of acceptance. I don't think they are particularly surprising insights, and it seems the sole merit of this paper is an empirical one, and impressive because of performance on complex tasks.

This paper examines the use of generative models for developing representations to improve data efficiency in RL. Specifically, the authors use a generative model that is trained to predict multiple frames into the future (overshooting), and they show that when the model is stochastic (but not deterministic), overshooting leads to useful representations of the environment that can improve RL efficiency. The reviews on this paper were fairly divergent in the first round. Two of the reviewers liked this paper, but one did not feel it provided truly novel contributions, and only brought together previously proposed ideas for using predictive training to improve RL representations. In discussion, the reviewers came to the conclusion that it does demonstrate the utility of overshoot prediction for stochastic models and that an empirical demonstration like this can be useful.

The standard ``serial'' (aka ``singleton'') model of belief contraction models the manner in which an agent's corpus of beliefs responds to the removal of a single item of information. One salient extension of this model introduces the idea of ``parallel'' (aka ``package'' or ``multiple'') change, in which an entire set of items of information are simultaneously removed. Existing research on the latter has largely focussed on single-step parallel contraction: understanding the behaviour of beliefs after a single parallel contraction. It has also focussed on generalisations to the parallel case of serial contraction operations whose characteristic properties are extremely weak. Here we consider how to extend serial contraction operations that obey stronger properties. Potentially more importantly, we also consider the iterated case: the behaviour of beliefs after a sequence of parallel contractions. We propose a general method for extending serial iterated belief change operators to handle parallel change based on an n-ary generalisation of Booth & Chandler's TeamQueue binary order aggregators.

Weaknesses: The absence of any explaining figures makes it hard for the reader to fully understand the underlying architecture of BPNNs. Specifically, it is not clear what the actual target of a BPNN acting on a particular factor graph is! Is it the partition function? Moreover, the intention of the Bethe Free Energy Layer remains obscure. In lines 121-125 you mention: 'When convergence to a fixed point is unnecessary, we can increase the flexibility [...] Additionally we define a Bethe free energy layer using two MLPs that take the trajectories of learned beliefs from each factor and variable as input and output scalars.'

The reviewers felt the paper provides a valuable connection between modern graph neural networks and belief propagation. There was some confusion about how to train the BPNNs, and the paper should be revised to clarify these points. The authors should also expand on discussion about marginal estimation, even if is to highlight limitations of the proposed approach.

General comments: (i) The authors only solve a new special kind of higher order consistency constraints, generalizing soft PN-potentials, but not a truly general class of constraints, as indicated in the title or in the abstract. In case of MAP-inference, which is normally desired, the goal is to obtain a single assignment which satisfies all given linear constraints. The relaxed model the authors optimize is simply a byproduct of looking for marginals instead of MAP-assignments (the added entropy is responsible for this). In case of vanishing entropy one gets the same model. Hence there certainly remains the disadvantage of a parameter in the PN-potential, but now hidden in the entropy.

Major comments: - As mentioned by the authors in the introduction, the SBM is widely used throughout many areas, and not only for detecting communities. The authors introduce no condition on the parameter matrix Q so that their setup includes (in principle) cases as diverse as communities, anti-communities (graphs of a bipartite type) and any other type of connectivity structure. However, I believe that their algorithm (based on belief propagation and thus on the community structure) will reveal only communities. If I am right, they should specify where this underlying assumption comes into play. Is it only that definition 3 becomes useless when the structure is not the one of communities?

The authors develop an very interesting method to the hard problem of computing the partition function of GM. In particular, there is no method to compute Z for general graphs accurately and log(Z) is a particularly interesting component in many fields: it is related to the likelihood in Machine Learning, it gives all the statistical information of a system in statistical mechanic, ... I think therefore that any effort toward that direction is an important contribution. Concerning the work presented here, I have some questions and concerns that I would like the authors to answer: 1) the authors mentioned that BP in general fail to converge on general graph but that more involved convergent alternative exists. I think the author should be more specific, precising if all the convergent algorithms do provide the same estimation of the BP partition function and it they are indeed linear in the system size. The 2-reg loop can be proven to converge in polynomial time but the full loop does not have these garranty.

The community detection problem requires to cluster the nodes of a network into a small number of well-connected'communities'. There has been substantial recent progress in characterizing the fundamental statistical limits of community detection under simple stochastic block models. However, in real-world applications, the network structure is typically dynamic, with nodes that join over time. In this setting, we would like a detection algorithm to perform only a limited number of updates at each node arrival. While standard voting approaches satisfy this constraint, it is unclear whether they exploit the network information optimally.

Belief Propagation (BP) is an important message-passing algorithm for various reasoning tasks over graphical models, including solving the Constraint Optimization Problems (COPs). It has been shown that BP can achieve state-of-the-art performance on various benchmarks by mixing old and new messages before sending the new one, i.e., damping. However, existing methods on tuning a static damping factor for BP not only is laborious but also harms their performance. Moreover, existing BP algorithms treat each variable node's neighbors equally when composing a new message, which also limits their exploration ability. To address these issues, we seamlessly integrate BP, Gated Recurrent Units (GRUs), and Graph Attention Networks (GATs) within the massage-passing framework to reason about dynamic weights and damping factors for composing new BP messages. Our model, Deep Attentive Belief Propagation (DABP), takes the factor graph and the BP messages in each iteration as the input and infers the optimal weights and damping factors through GRUs and GATs, followed by a multi-head attention layer.