- Rating
- 1
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
This paper considers the problem of epidemic policy planning. The formalism treated in this paper is quite high-level. In it, regions (e.g., states of the US) are connected by a graph structure, whose edges represent flows of mobility between regions. The action space consists of setting certain parameters (e.g., vaccine allocations, financial resources). The goal is to minimise an objective function that trades the effectiveness of the parameters with a notion of cost, while probabilistically guaranteeing the satisfaction of an signal temporal logic specification. The authors propose a method with two stages: first, a graph-based world model (i.e., an RSSM variant) is trained to predict the evolution of the epidemic graph state and then frozen. Secondly, a receding-horizon planner rolls out trajectories from the world model to determine the best action to apply at each time step (and is invoked again from the new state once the actions are applied in the environment). The approach is evaluated on several scenarios ranging from synthetic to scenarios whose simulator parameters are fit based on real-world data. The proposed method generally compares favourably to baselines.
Post-rebuttal update
After reading the authors' replies, I retain my original score. Some details were now provided, but the clarity of the paper remains extremely poor, to the point that it is not understandable as a coherent whole. Generative AI was indeed used extensively but this was not disclosed upon submission.
Critical strengths
S1. The main strength of the work is, in my opinion, the level of abstraction used. The problem framing is such that the method is compatible with how epidemic policies are set in the real-world. Thus, the work has potential to be used in a highly impactful decision-making scenario.
Critical weaknesses
W1. Lack of clarity
The key weakness of the work is its lack of clarity. The writing uses many specialist terms, concepts, and notations that are assumed to be known to the reader, but are often left undefined or assumed to be known. To give a few examples:
- The STL specification is key to the narrative of the work and an important differentiator to other methods, but zero relevant background is provided. I would not expect the average ICLR reader to be familiar with temporal logics; so some key background (incl.) robustness needs to be provided. How is differentiable robustness achieved during optimization, as this appears non-trivial?
- The loss function for training the GWM contains 6 terms, none of which are precisely defined.
- Similarly, the metrics used are not precisely defined (page 6), and beyond standard MAE/RMSE they should be.
W2. Possible undeclared use of LLMs
The writing is locally coherent, but lacking a clear narrative that can be followed sentence by sentence. The general feeling reading the manuscript is that it is LLM-generated to a large extent, and not human-originated writing that was edited using LLMs. This would also explain the lack of clarity I raised in W1.
I have also scanned the code repository provided, and several files (e.g. https://anonymous.4open.science/r/epimind-9706/epimind-code/eval/oracle_dynamics.py) are consistent with the use of agentic tools to write code, run experiments, analyze data, and interpret results. This file has references to some .md files that are not included in the repository but presumably contain prompts or dataset descriptions for use by agents. The comments in this file, as in others, seem to correspond in many cases to justifications of design choices relative to previous iterations of the code.
While this is an assertion that I cannot prove, it appears that a large chunk of this work, both in terms of the manuscript and the code, are LLM-generated, which contradicts the declaration in the paper.
W3. Insufficient statistical evidence
Only 3 and 5 seeds are used respectively for training and evaluation. This is insufficient to draw statistically robust conclusions; the manuscript attempts to defend this pre-emptively but confidence intervals in many results are overlapping. Reporting mean +- sd is inappropriate; confidence intervals should be computed, which at 95% confidence level corresponds to approximately 1.97 standard deviations. This makes the results artificially appear more robust than they are.
Questions for authors
Please address W1, W2, and W3.
Smaller comments / suggestion:
- An MDP problem formulation should be provided, and the STL spec should be moved to the main text, as it is part of the problem definition.
- I would expect the work of Meirom et al., 2021 to be cited as it is highly relevant (http://proceedings.mlr.press/v139/meirom21a/meirom21a.pdf). If it is not used as a baseline, this should be justified.
- Beliefs and latents are introduced before being used in Eq 4. Similarly, various terms are used in Eq 13 at a point in which their purpose is not clear and they are not defined closeby. I'd suggest auditing all math notation and checking whether terms are defined before first use.
- "No graph" baseline: "no graph attention" would presumably be more accurate, as an underlying graph structure is still used.
- 3.2.2 section title: should be "learned rollouts yield"
2 comments
4. Additional Comments
-
Decision formulation and STL specification. We will state the sequential decision problem in the main text, defining the belief state, regional actions, learned transitions, objective, and shared resource constraints. We will present the STL capacity specification alongside this formulation.
-
Baseline selection. We thank the reviewer for pointing us to Meirom et al. (2021), which is closely related to our work on epidemic control over graphs. We will cite it in the related-work section. The two methods are not directly comparable as baselines. Meirom et al. select individuals for testing under a fixed test budget. EpiMind allocates continuous interventions across regions under shared resource budgets. A direct comparison would require replacing Meirom et al.’s node-selection mechanism with a continuous regional allocation rule.
-
Notation and presentation. We thank the reviewer for these suggestions. We will define beliefs and latent states before Eq. (4) and explain the regional-update terms when they first appear in Eq. (13). We will rename the ablation “no graph attention” to reflect that it retains the underlying graph structure and retitle Section 3.2.2 “Learned rollouts yield effective resource-feasible interventions.”
We thank the reviewer for the constructive comments. We clarify the planning formulation, AI use, and statistical evidence.
1. Planning Formulation, Training Objective, and Evaluation Metrics (W1)
STL specification and differentiable robustness. At each planning step h, the capacity margin mh is capacity minus predicted occupancy. We optimize the soft minimum −τlog∑hexp(−mh/τ) with τ=0.5. After resource projection, we check the exact minimum on the joint predicted trajectory. We will present the decision formulation and STL specification together in the main text.
World-model training loss. The table below explains each term listed in the training loss Eq. (10) and how it is used in training.
Evaluation metrics.
2. Scope of AI Assistance (W2)
We thank the reviewer for raising this point. Generative AI assisted with literature retrieval, coding, analysis of experimental results, and manuscript editing. The comments in
eval/oracle_dynamics.pydocument a fidelity check that found approximately 17% relative error in weekly admissions for the earlier oracle implementation. The audit traced this discrepancy to double-counting within-region transmission and using fiscal-adjusted NPI to scale mobility. It also identifies the results that depend on the earlier oracle and need to be rerun. We will include the three Markdown files referenced by the code in the release.3. Paired-Seed Evidence and Uncertainty Reporting (W3)
Table 1 reports three checkpoints. The 29% admissions RMSE reduction is a point estimate, and the RMSE difference remains statistically unresolved. Across 10 paired training seeds in the follow-up evaluation, GF-RSSM improved predictive correlation by 0.024 over the variant without graph attention, with higher correlation in 9 seeds. In a separate controlled intervention test, the neighbor-to-non-neighbor response ratio was 9.13 for GF-RSSM and 1.07 for Action-LSTM. The mean paired difference was 8.06 (95% paired t CI [7.24, 8.88]). The simulator’s ratio was 5.98, so GF-RSSM captured the spatial concentration and overestimated its strength. We will report confidence intervals for the original table results and use paired differences for method comparisons.