acceptodds
Under review as a conference paper at ICLR 2027

Network World Models as Environments for Algorithm Design on Complex Systems

Abstract

World models, which simulate an environment and predict how it changes under actions, are increasingly used in real-world applications such as robotics. Complex systems call for the same tool because the effect of an action is not immediate. Seeding nodes for a campaign, or immunizing nodes against an epidemic, changes little on its own; what matters is the outcome that unfolds over the steps that follow. Designing an algorithm that selects such actions to maximize expected performance on a task is inherently iterative, and every candidate must be scored by the outcome it produces. Obtaining that outcome has relied on simulation, whose cost becomes a bottleneck when candidates are evaluated over many sampled trajectories. We propose an action-conditioned Network World Model that learns a network's diffusion dynamics under interventions over time, applies each action to the network, and predicts the outcome that follows. It serves as a fast evaluator inside an algorithm design loop in which a coding agent designs and refines executable algorithms using feedback from full rollouts, action-level credit, and counterfactual probes over alternative interventions. Across eight network tasks and five diffusion models, the designed algorithms match or exceed the strongest reported baseline in 118 of 121 settings while enabling up to 14.5 times faster rollouts than Monte Carlo simulation.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
network world models, automated algorithm design, graph learning, diffusion dynamics

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_1i1yOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes an action-conditioned network world model for LLM-based algorithm design on graph diffusion tasks. And a graph neural model is trained with a process-constrained head to predict subsequent dynamics. A coding agent then refines executable algorithms using rollout rewards, trajectory diagnostics, leave-one-out action credit, and counterfactual probes. The paper provides a conditional selection-regret bound and evaluates eight tasks and five diffusion processes. Independently scored algorithms match or exceed the strongest listed baseline in 118 of 121 settings.

Critical strengths
  1. The application is meaningful: designing reusable algorithms for budgeted network interventions is more persistent than selecting an action sequence for one task instance.

  2. The paper offers a coherent interface between explicit node/edge interventions, stochastic diffusion rollouts, and a coding agent. Applying immediate edits exactly while exposing intervention-specific trajectory and counterfactual feedback is a potentially valuable extension beyond fixed-seeding diffusion estimators. Its strongest contribution appears to be this integrated algorithm-design environment.

Critical weaknesses
  1. The practical benefit of learning the world model. Section 3.2 motivates learning unobserved dynamics parameters, but Appendix I.1.1 keeps transmission weights visible, and Appendix E.1 already provides a vectorized exact Oracle. Figure 4 reports acceleration over NDlib without comparing this Oracle, while Appendix G changes the search loop and feedback alongside the evaluator. A controlled comparison keeping the search loop and feedback fixed, and replacing only the learned evaluator with the exact Oracle, would directly establish whether learning improves final algorithm quality or reduces total cost. This comparison should include data generation and training costs, accounting for reuse across searches.

  2. The novelty discussion understates the closest precedents. MONSTOR already learns single-step graph diffusion and stacks it for multi-step evaluation inside IM algorithms, rather than merely predicting a final scalar. Memon et al., which is cited, already refine an LLM intervention policy using world-model rollout feedback. Schmidt et al. also model explicit intervention change points with a graph surrogate. The paper has meaningful distinctions, but should establish these narrower advances rather than implying the broader learned-simulator/feedback paradigm is missing from prior work.

Questions for authors
  1. The title suggests broad applicability to complex systems, while the evaluation focuses on network diffusion and influence propagation. Could you justify the broader scope or narrow the title and claims accordingly?

  2. Appendix B.2 describes a 50-word search memory and growing attempts/population tables. How are context limits, history truncation, and candidate retention or retrieval handled as the serial search grows beyond the reported 10 rounds?

  3. What benefit does learning a world model provide over direct algorithm search with an exact simulator? Please compare the same search loop and feedback using the learned evaluator versus the vectorized exact Oracle (Appendix E.1), reporting final quality and total runtime, including data generation and training, under matched search budgets and hardware.

In all, I see some novelty in this work and would consider raising my score if my concerns are properly addressed.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 1i1y [1/3]

We thank the reviewer for the feedback and for recognizing the significance of the application and novelty of our work. We address each point below.

W1 & Q3: Network World Model compared to the exact oracle

The practical benefit of learning the world model. Section 3.2 motivates learning unobserved dynamics parameters, but Appendix I.1.1 keeps transmission weights visible, and Appendix E.1 already provides a vectorized exact Oracle. ... What benefit does learning a world model provide over direct algorithm search with an exact simulator? Please compare the same search loop and feedback using the learned evaluator versus the vectorized exact Oracle (Appendix E.1), reporting final quality and total runtime, including data generation and training, under matched search budgets and hardware.

Learned edge weights: The Network World Model does not receive the true edge transmission probabilities. Its input weight is wuv=1w_{uv} = 1 on every edge, in both training and rollouts. The true probabilities are used only by the simulator that generates the training data and by the oracle that scores the final programs for a fair comparison. The model therefore observes only which nodes become infected, and learns the transmission probabilities from the network's structure and these observed infections.

Some confusion may have arisen from ambiguous symbols and wording, which we have rewritten at lines 167-169, 189-192, and 2225-2228 to clarify that the model learns the transmission probabilities, which are distinct from the true probabilities used by the oracle simulator. We thank the reviewer for bringing this up.

The oracle simulator is a reference: The oracle is a vectorized simulator of the diffusion process with its true parameters. It exists only when every parameter of the process is known, so it cannot be used in real deployments, where the dynamics may be unknown. In our benchmarks, it plays the role of the real world: the final algorithms of our method, every baseline, and every algorithm-discovery system are scored on it for a fair comparison. Because it knows the exact dynamics, search with the oracle is the best case and essentially bounds what learned dynamics can achieve.

The oracle cannot be used in settings where no exact model exists. For example, the dynamics may be available only as observed cascades, as in our experiments on APS, Taoke, and Digg, where the Network World Model is trained directly on these cascades. Or a simulator may exist, but its dynamics are not given as explicit edge-level transmission probabilities, as in agent-based epidemic simulators, where spread emerges from individual behavior. In that case, no exact vectorized implementation can be built, and a candidate can only be scored by repeated simulation, so Monte Carlo is the fair comparison. Against Monte Carlo, the Network World Model is 4.0×4.0\times faster per rollout on Power Grid and 14.5×14.5\times faster on Digg (Figure 4), and breaks even after about five searches.

When the full model of the diffusion process is known, the oracle is fastest, because it uses the true parameters directly. The oracle and the Network World Model are both faster than Monte Carlo simulation. In the same design loop, the Network World Model finds algorithms within 0.13 percentage points of those found with the oracle on both networks (table below), and within a median of 0.05% over all 25 paired cells of our simulated tasks (at most 1.6%).

Setup: IM on Network Science and CND on Power Grid, 10% budget, IC. Only the evaluator changes. All three evaluators get score-only feedback (each candidate's score and its comparison with the current best), because a Monte Carlo search with diagnostics, credit, and probes would take too long. This also motivates our faster Network World Model. Given more time, we will soon add the full-feedback comparison for all three evaluators. The reported training cost includes data generation and model training time.

EvaluatorIM, Network Science (%) ↑\uparrowCND, Power Grid (%) ↓\downarrowTime per evaluation of 200 rollouts (s), NS / PGTraining cost, NS / PG
Oracle (true parameters)36.6219.160.22 / 3.54none
Monte Carlo (NDlib)36.5819.165.86 / 23.85none
Network World Model36.6119.161.62 / 6.7715 min / 3.3 h
AuthorsOpenReview ↗

Response to Reviewer 1i1y [2/3]

W2: Positioning against the closest prior work

The novelty discussion understates the closest precedents: MONSTOR already learns single-step graph diffusion and stacks it for multi-step evaluation inside IM algorithms; Memon et al. already refine an LLM intervention policy using world-model rollout feedback; Schmidt et al. also model explicit intervention change points with a graph surrogate.

Our contributions: (1) An action-conditioned world model that applies interventions at any step (seeding, node and edge removal, blocking, weight changes) and rolls the dynamics forward, serving as a learned environment for algorithm design on complex networks. (2) A coding agent that writes and refines reusable, executable algorithms in this environment from step-level feedback, across eight tasks. Each prior work covers only part of this:

  • MONSTOR (Ko et al., 2020) uses a GCN to estimate a seed set's influence under IC, replacing Monte Carlo inside greedy influence maximization. It is an estimator within a fixed algorithm.
  • Memon et al. (2026) is a perspective on world models for epidemiology. Its case study fine-tunes a language model to forecast state-level hospitalization trends from policy vectors and refines an LLM policy with its rollouts. It has no network and no network interventions, and its output is a policy for one setting.
  • Schmidt et al. (2024) build a GNN surrogate of a metapopulation epidemic model for forecasting. Its only intervention is a change in contacts at a fixed number of change points.

The related work section now states these distinctions (lines 112-119).

Q1: Scope of the work

The title suggests broad applicability to complex systems, while the evaluation focuses on network diffusion and influence propagation. Could you justify the broader scope or narrow the title and claims accordingly?

Here, "complex systems" means complex networks, the established framing for dynamical processes on networks (Newman, 2003; Boccaletti et al., 2006; Pastor-Satorras et al., 2015). These are systems whose state evolves through interactions along a network and changes under actions taken on it, as the introduction states. The introduction now states this scope more firmly (lines 71-74).

Our evaluation covers more than just influence propagation, with eight tasks spanning four problem families and five diffusion processes (IC, LT, CLT, SIR, SIS):

  • spreading: influence maximization, static and adaptive;

  • containment:

    • critical node detection against an outbreak
    • blocking a rumor with a competing cascade
    • epidemic control with recovery and reinfection, where actions include vaccination, quarantine, and contact reduction
  • inverse problems: locating the sources of an outbreak, and reconstructing who infected whom from a partial observation

  • forecasting: predicting the growth of real logged cascades, with no simulator

  • The networks span several domains, from 10310^3 to more than 10510^5 nodes:

    • social and communication: Digg, Email-EU, Brightkite, Deezer
    • collaboration and citation: Network Science, NetHEPT, Cit-HepTh, APS
    • human face-to-face contact: Infectious SocioPatterns
    • infrastructure: the Power Grid, and the Oregon1 Internet topology
    • peer-to-peer and trust: Gnutella, PGP
AuthorsOpenReview ↗

Response to Reviewer 1i1y [3/3]

Q2: Context limits and candidate retention

Appendix B.2 describes a 50-word search memory and growing attempts/population tables. How are context limits, history truncation, and candidate retention or retrieval handled as the serial search grows beyond the reported 10 rounds?

The context stays bounded: Each call to the coding agent sends the system turn (program contract and rules), the opening turn (task statement and network profile), and the 10 most recent exchanges.

  • One exchange is a round's prompt (operator instruction, attempts and population tables, memory and feedback) plus the model's program.
  • Once the program is scored, it is shrunk to a diff against its parent.
  • Older exchanges are dropped. Past a 10-round search, the model only sees the final 10 rounds. In a 5-round or 10-round search, for example, search drops nothing.

The search memory is rewritten every round at 50 words. The attempts table keeps the first attempt and the most recent. Candidates are kept outside the conversation. The population table lists every scored candidate with its reward, paired strength and mechanism line, ranked by paired strength. It grows by one short line per candidate.

For retrieval:

  • Every round's prompt includes the current best program in full as the parent, with its diagnostics, so the best candidate stays in context however many rounds pass.
  • Other members are retrieved by paired strength: under refine, the two strongest are shown as alternatives, and under crossover, a partner is drawn by rank.

We have added a more detailed description of context limits with the coding agent algorithm design loop to Appendix B.2.

We thank the reviewer again for the detailed and constructive review, and we hope that the clarifications and additions address the reviewer's concerns. We are glad to answer any further questions.

Reviewer_JPUzOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses the design of algorithms that choose interventions on networks (e.g., seeding, node removal, immunization, blocking) whose effect unfolds through a diffusion process. Designing an algorithm for this is iterative, and every candidate has to be scored by its outcome. Today that score comes from Monte Carlo simulation, which the authors identify as the bottleneck. It proposes a Network World Model: a GraphSAGE-based one-step transition model. The model serves as the evaluator inside an LLM coding-agent loop that writes and refines executable programs. A simulation-lemma bound relates the world model's one-step error to algorithm-selection regret. Experiments cover eight tasks and five diffusion models on networks of up to more than 10510^5 nodes. The paper reports that the designed programs match or exceed the strongest baseline in 118 of 121 settings, transfer across networks, and run faster than Monte Carlo simulation.

Critical strengths
  1. Breadth of the end-to-end evaluation. Eight tasks and 5 diffusion models, scored on one shared referee. This uniform protocol makes the comparisons across tasks consistent and shows that a single design loop can be applied to quite different network problems.

  2. Readable and transferable programs. The returned algorithms are interpretable procedures rather than black-box policies. In the completed transfers, programs designed on one network perform on par with programs designed directly on other networks of the same task. This supports the value of designing algorithms rather than instance-specific solutions.

Critical weaknesses
  1. The contribution of the learned world model is not isolated, and the claims are broader than the setting tested. All simulated tasks use diffusion processes whose functional form is given, the readout is constrained to that form, the true transmission probabilities are available to the generated programs through GraphInfo, and the paper already has an exact vectorized simulator, which it uses as the referee. No experiment runs the design loop with that exact simulator, or with plain Monte Carlo, as the evaluator, so it is unclear whether the gains come from the learned model or from the search loop and its feedback. The speed-up is measured against the NDlib Monte Carlo simulator, not against the authors' own vectorized simulator. Theorem 1 does not close this gap: ϵ\epsilon is a total-variation bound on the joint next state of the whole network, which can be close to 1 on large graphs even when per-node errors are small, and the paper does not estimate it. The title and framing refer to "complex systems" and "network dynamics", but the method assumes known diffusion processes. A run of the loop with the exact simulator as evaluator, wall-clock comparisons against it, and a scope statement in the title or abstract would help addressing this.

  2. The 118/121 headline is confounded and reported without variance. The task reward counts activations within a fixed horizon, while the reverse-influence-sampling baselines (IMM, OPIM, SubSIM) optimize unbounded spread. Most critical node detection baselines ignore the outbreak location, which the designed programs exploit, and the only outbreak-aware baseline is the authors' own control. Each table cell comes from a single search with one base seed, on one network and outbreak. Since many wins are by small margins, it is not possible to tell which gains exceed run-to-run variability.

Questions for authors
  1. Clarification. What happens when the design loop uses the exact vectorized simulator (or Monte Carlo) as the evaluator instead of the Network World Model, for the same budget of rounds? If the returned programs and their scores are similar, the paper's contribution lies in the search loop rather than in the learned model. If they are worse, or the loop is much slower, this would substantially support the central claim.
  2. Clarification. On which networks were the world-model checkpoints used in the main experiments (Tables 1 and 13) trained, and with which optimizer, learning rate, and number of epochs? App. I.1 describes the BA-24 and SBM-24 suites, but the link to the end-to-end runs is not stated.
  3. Clarification. How do the RIS baselines perform when their objective is truncated at the same horizon HH used by the task reward? A comparison under a matched objective would show how much of the influence-maximization gains on large graphs remains.
  4. Clarification. Can you report the variance of the returned programs' scores over several independent searches (different seeds) for a few representative cells?
  5. Suggestion. Extending the framework to settings where the process form is unknown or misspecified, and the dynamics must be learned from observed cascades, would test the world-model framing more directly. This is future work and does not affect the rating.
9 comments
AuthorsOpenReview ↗

Response to Reviewer JPUz [1/3]

We thank the reviewer for the feedback and for recognizing the breadth of the evaluation and the transferability of the designed algorithms. We address each point below.

W1 & Q1: What the GNN learns

What happens when the design loop uses the exact vectorized simulator (or Monte Carlo) as the evaluator instead of the Network World Model, for the same budget of rounds?

The oracle simulator is a reference: The oracle is a vectorized simulator of the diffusion process with its true parameters. It exists only when every parameter of the process is known, so it cannot be used in real deployments, where the dynamics may be unknown. In our benchmarks, it plays the role of the real world: the final algorithms of our method, every baseline, and every algorithm-discovery system are scored on it for a fair comparison. Because it knows the exact dynamics, search with the oracle is the best case and essentially bounds what learned dynamics can achieve.

The oracle cannot be used in settings where no exact model exists. For example, the dynamics may be available only as observed cascades, as in our experiments on APS, Taoke, and Digg, where the Network World Model is trained directly on these cascades. Or a simulator may exist, but its dynamics are not given as explicit edge-level transmission probabilities, as in agent-based epidemic simulators, where spread emerges from individual behavior. In that case, no exact vectorized implementation can be built, and a candidate can only be scored by repeated simulation, so Monte Carlo is the fair comparison. Against Monte Carlo, the Network World Model is 4.0×4.0\times faster per rollout on Power Grid and 14.5×14.5\times faster on Digg (Figure 4), and breaks even after about five searches.

When the full model of the diffusion process is known, the oracle is fastest, because it uses the true parameters directly. The oracle and the Network World Model are both faster than Monte Carlo simulation. In the same design loop, the Network World Model finds algorithms within 0.13 percentage points of those found with the oracle on both networks (table below), and within a median of 0.05% over all 25 paired cells of our simulated tasks (at most 1.6%).

Setup: IM on Network Science and CND on Power Grid, 10% budget, IC. Only the evaluator changes. All three evaluators get score-only feedback (each candidate's score and its comparison with the current best), because a Monte Carlo search with diagnostics, credit, and probes would take too long. This also motivates our faster Network World Model. Given more time, we will soon add the full-feedback comparison for all three evaluators. The reported training cost includes data generation and model training time.

EvaluatorIM, Network Science (%) ↑\uparrowCND, Power Grid (%) ↓\downarrowTime per evaluation of 200 rollouts (s), NS / PGTraining cost, NS / PG
Oracle (true parameters)36.6219.160.22 / 3.54none
Monte Carlo (NDlib)36.5819.165.86 / 23.85none
Network World Model36.6119.161.62 / 6.7715 min / 3.3 h

W1: Transmission probabilities available to the programs

The true transmission probabilities are available to the generated programs through GraphInfo.

Learned edge weights: The Network World Model does not receive the true edge transmission probabilities. Its input weight is wuv=1w_{uv} = 1 on every edge, in both training and rollouts. The true probabilities are used only by the simulator that generates the training data and by the oracle that scores the final programs for a fair comparison. The model therefore observes only which nodes become infected, and learns the transmission probabilities from the network's structure and these observed infections.

We thank the reviewer for raising this point, which our original wording left ambiguous. The programs read the probabilities learned by the Network World Model, not the true ones: GraphInfo exposes the same probabilities the evaluator uses, so the learned edge probabilities are those from the Network World Model.

Some confusion may have arisen from ambiguous symbols and wording, which we have rewritten at lines 167-169, 189-192, 2225-2228, and Appendix C (lines 1356-1359) to clarify that transmission probabilities are learned by the model and are also used by GraphInfo.

AuthorsOpenReview ↗

Response to Reviewer JPUz [2/3]

W1: Theorem 1 on large graphs

Theorem 1 does not close this gap: ϵ\epsilon is a total-variation bound on the joint next state of the whole network, which can be close to 1 on large graphs even when per-node errors are small, and the paper does not estimate it.

ϵ\epsilon is a worst-case bound on the joint next state, and it can be loose on large graphs, where per-node errors add up across nodes in ϵ\epsilon, so it does not reflect how accurate the score itself is. Selection depends only on the score, which we measure. We replay algorithms designed on large graphs on both the Network World Model and the exact oracle simulator. We report the following:

  • Error: the Network World Model's score minus the oracle's, as % of the returned algorithm's true score.
  • Error spread: the largest minus the smallest error across the algorithms of one search.
  • Regret: the true score of the best algorithm the search tried minus that of the algorithm it returned.
Task (network)ModelAlgorithmsLargest error (%)Error spread in a search, median (%)Regret, max (%)
IM (Digg)IC800.04 to 0.780.200.04
IM (Digg)LT797.40 to 9.600.630.05
CND (PGP)IC404.08 to 4.241.370.59
CND (PGP)LT607.22 to 9.051.690.19

The largest error is at most 9.6%. Measured regret, even on large graphs, is much smaller than the theorem's bound. It is at most 0.59%, against up to 19.2% for twice the largest error. Thus, a small ϵ\epsilon guarantees a small score gap, but a large ϵ\epsilon does not imply a large one. The table shows that the score gap is small here, and the error spread, which is what selection depends on, is smaller still.

Q2: Checkpoints used in the main experiments

On which networks were the world-model checkpoints used in the main experiments (Tables 1 and 13) trained, and with which optimizer, learning rate, and number of epochs?

  • Networks: Each Network World Model is trained on the same network and diffusion process it is used on, separately for each task. BA-24 and SBM-24 are used only in the standalone study of Appendix I to test world model-specific performance on network topologies separate from those in our main tables.
  • Training: Adam, learning rate 10−310^{-3}, weight decay 5×10−45 \times 10^{-4}, batch size 32, up to 400 epochs with early stopping on validation Δ\DeltaF1 (patience 50).

We have added these details with further specifications on the Network World model in Appendix F, lines 1859-1968.

W2 & Q3: RIS baselines with a matched objective

How do the RIS baselines perform when their objective is truncated at the same horizon HH used by the task reward?

We replay every baseline method that uses RIS (IMM, OPIM, SubSim) to measure spread with no limiting horizon. These are compared against our designed algorithms, which were designed for the horizon H=10H = 10 score, so this setup favors the RIS baselines.

MethodNS 1%NS 5%NS 10%NS 20%Digg 1%Digg 5%Digg 10%Digg 20%NetHEPT 1%NetHEPT 5%NetHEPT 10%NetHEPT 20%
IMM (H=10H = 10)8.7 / 10.924.8 / 30.238.0 / 45.159.2 / 68.525.9 / 47.036.0 / 60.944.2 / 69.755.7 / 79.611.6 / 14.828.9 / 36.242.4 / 51.661.7 / 72.5
IMM (no horizon)8.7 / 10.924.8 / 30.238.0 / 45.159.2 / 68.526.1 / 56.236.0 / 65.344.2 / 71.955.7 / 80.211.6 / 15.028.9 / 36.342.4 / 51.661.7 / 72.5
OPIM (H=10H = 10)8.8 / 10.824.2 / 29.537.7 / 45.358.3 / 67.427.4 / 50.541.6 / 68.451.8 / 77.662.7 / 85.512.1 / 15.531.4 / 39.244.7 / 54.364.5 / 75.7
OPIM (no horizon)8.8 / 10.824.2 / 29.537.7 / 45.358.3 / 67.427.5 / 57.541.6 / 69.651.8 / 77.962.7 / 85.612.1 / 15.831.4 / 39.344.7 / 54.364.5 / 75.7
SubSIM (H=10H = 10)8.7 / 10.724.3 / 29.137.3 / 45.057.8 / 68.527.3 / 50.141.7 / 68.451.8 / 77.762.7 / 85.412.6 / 16.231.3 / 39.044.8 / 54.564.3 / 75.4
SubSIM (no horizon)8.7 / 10.724.3 / 29.137.3 / 45.057.8 / 68.527.4 / 57.441.7 / 69.751.8 / 78.062.7 / 85.512.6 / 16.531.3 / 39.144.8 / 54.564.3 / 75.4
Network World Model (H=10H = 10)8.9 / 11.325.3 / 30.939.1 / 46.860.1 / 69.330.4 / 56.845.8 / 73.256.5 / 83.169.4 / 93.512.9 / 16.933.0 / 41.447.8 / 58.168.5 / 80.0
Network World Model (no horizon)8.9 / 11.325.3 / 30.939.1 / 46.860.1 / 69.330.5 / 59.645.8 / 73.856.5 / 83.469.4 / 93.612.9 / 17.133.0 / 41.447.8 / 58.268.5 / 80.0

Our method remains ahead of the best RIS baselines in all settings. In 22 settings, it changes by at most 0.06 points because, under IC, the cascades are over by step 10. We added these results to Appendix J.3, lines 2188-2217.

AuthorsOpenReview ↗

Response to Reviewer JPUz [3/3]

W2 & Q4: Variance over independent searches

Can you report the variance of the returned programs' scores over several independent searches (different seeds) for a few representative cells?

We run the algorithm design loop with three independent searches using three random seeds on IM (Network Science) and CND (Power Grid), with a budget of 10% on IC. Given more time, we will soon repeat more runs with the three seeds and report variances for more graph topologies, budgets, and diffusion models.

TaskMean ±\pm std (Our method)Strongest baseline
IM, Network Science, spread (%) ↑\uparrow39.05±0.0239.05 \pm 0.0238.03 (IMM)
CND, Power Grid, infected (%) ↓\downarrow17.31±0.0217.31 \pm 0.0218.74 (Frontier)

Run-to-run variation is at the level of scoring noise: The returned programs differ by at most 0.04 points. The standard deviations, about 0.28 nodes (Network Science) and 0.99 nodes (Power Grid), are no larger than the referee's own standard error for one program (0.68 and 0.92 nodes). The margin over the strongest baseline, 1.02 points on IM and 1.43 points on CND, is about 59 and 72 standard deviations. These results are included in Appendix J.2, 2164-2186.

Q5: Dynamics learned from observed cascades

Extending the framework to settings where the process form is unknown or misspecified, and the dynamics must be learned from observed cascades, would test the world-model framing more directly.

The paper already includes testing this setting: cascade prediction. For cascade prediction, no simulator or process form is given, and the Network World Model learns the dynamics from logged real cascades on APS, Taoke, and Digg. This designed algorithm has the lowest MSLE on all three datasets: 0.078, 0.544, and 0.774, compared with 0.098, 0.716, and 0.834 for the strongest baseline.

We thank the reviewer for the detailed and constructive review, and we would be glad to clarify any remaining points.

Reviewer_tb3XOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how to improve the efficiency of automated algorithm design for complex networks. The authors propose an action-conditioned Network World Model (NWM) that learns network diffusion dynamics to predict the effects of different actions and serves as an evaluation environment for LLM-based algorithm search. Specifically, the LLM generates and refines executable algorithms, while the NWM provides simulation results, action-level credit estimates, and counterfactual feedback to guide the refinement process. The experiments cover eight network tasks and five diffusion models. The authors report that their method matches or outperforms the strongest baseline in 118 out of 121 experimental settings and achieves up to a 14.5× speedup in individual rollouts compared to conventional Monte Carlo simulation.

Critical strengths

1.The proposed framework is reasonable. It separates known intervention rules from learned diffusion dynamics and provides different types of feedback to guide algorithm design. The experiments cover several tasks and diffusion models, and the results are promising.

Critical weaknesses
  1. Figure 1 presents the overall framework but omits details of the Network World Model architecture. With known diffusion rules and ground-truth edge weights provided in some settings, it is unclear what the GNN learns and what differentiates this work from existing GNN simulators.

  2. NWM’s efficiency advantage is not well supported. Figure 4 only compares NWM with NDlib Monte Carlo. The vectorized Oracle evaluator (Appendix E.1) uses true diffusion parameters and requires no GNN training. Direct runtime comparison with this Oracle is needed to confirm speedup.

  3. The feedback mechanisms lack ablation validation. Rollout diagnostics, action credit and counterfactual probes are used, but their individual gains are unknown. Comparisons in Appendix G are against LLM baselines with inconsistent search strategies. Controlled ablations are required to demonstrate their utility.

Questions for authors
  1. Could the authors further clarify the architecture of the Network World Model? What does the GNN actually learn, and what is already defined by the structured transition head? In particular, when the diffusion rules and edge weights are known, what additional information does the GNN learn?

  2. Could the authors compare the runtime of NWM with the Vectorized Oracle Evaluator described in Appendix E.1? Since both support batched simulation, this comparison would help clarify the efficiency advantage of the learned model.

  3. Could the authors provide ablation studies to evaluate the contributions of rollout diagnostics, action-level credit, and counterfactual probes?

3 comments
AuthorsOpenReview ↗

Response to Reviewer tb3X [1/2]

We thank the reviewer for the feedback and address each point below.

W1 & Q1: Design loop with the oracle or Monte Carlo as evaluator

What does the GNN actually learn, and what is already defined by the structured transition head? In particular, when the diffusion rules and edge weights are known, what additional information does the GNN learn?

Learned edge weights: The Network World Model does not receive the true edge transmission probabilities. Its input weight is wuv=1w_{uv} = 1 on every edge, in both training and rollouts. The true probabilities are used only by the simulator that generates the training data and by the oracle that scores the final programs for a fair comparison. The model therefore observes only which nodes become infected, and learns the transmission probabilities from the network's structure and these observed infections.

Some confusion may have arisen from ambiguous symbols and wording, which we have rewritten at lines 167-169, 189-192, and 2225-2228 to clarify that the model learns the transmission probabilities, which are distinct from the true probabilities used by the oracle simulator.

NWM architecture: For clarity, we have added a detailed description of the architecture (Appendix F, lines 1859-1968). To summarize, each step applies the action to the network, runs a 3-layer GraphSAGE over node states, degree, and action features, and a head outputs each node's next state as one diffusion step, using a learned probability per incoming edge. Each model step corresponds to one step of the diffusion process; to score a candidate algorithm, we roll the model forward over the full horizon.

Difference from existing GNN diffusion models: Existing GNN diffusion models predict a fixed objective, such as the spread of a seed set, for one type of intervention (Xia et al., 2021; Ko et al., 2020; Ling et al., 2023). Our Network World Model is instead action-conditioned: it accepts any intervention at any step (seeding, node and edge removal, blocking, weight changes) and rolls the dynamics forward step by step. This is what enables all eight tasks. The step-level outcomes also give the agent rollout diagnostics and counterfactual probes, which help it score and revise algorithms better.

AuthorsOpenReview ↗

Response to Reviewer tb3X [2/2]

W2 & Q2: Runtime against the vectorized oracle

Could the authors compare the runtime of NWM with the Vectorized Oracle Evaluator described in Appendix E.1? Since both support batched simulation, this comparison would help clarify the efficiency advantage of the learned model.

The oracle simulator is a reference: The oracle is a vectorized simulator of the diffusion process with its true parameters. It exists only when every parameter of the process is known, so it cannot be used in real deployments, where the dynamics may be unknown. In our benchmarks, it plays the role of the real world: the final algorithms of our method, every baseline, and every algorithm-discovery system are scored on it for a fair comparison. Because it knows the exact dynamics, search with the oracle is the best case and essentially bounds what learned dynamics can achieve.

The oracle cannot be used in settings where no exact model exists. For example, the dynamics may be available only as observed cascades, as in our experiments on APS, Taoke, and Digg, where the Network World Model is trained directly on these cascades. Or a simulator may exist, but its dynamics are not given as explicit edge-level transmission probabilities, as in agent-based epidemic simulators, where spread emerges from individual behavior. In that case, no exact vectorized implementation can be built, and a candidate can only be scored by repeated simulation, so Monte Carlo is the fair comparison. Against Monte Carlo, the Network World Model is 4.0×4.0\times faster per rollout on Power Grid and 14.5×14.5\times faster on Digg (Figure 4), and breaks even after about five searches.

When the full model of the diffusion process is known, the oracle is fastest, because it uses the true parameters directly. The oracle and the Network World Model are both faster than Monte Carlo simulation. In the same design loop, the Network World Model finds algorithms within 0.13 percentage points of those found with the oracle on both networks (table below), and within a median of 0.05% over all 25 paired cells of our simulated tasks (at most 1.6%).

Setup: IM on Network Science and CND on Power Grid, 10% budget, IC. Only the evaluator changes. All three evaluators get score-only feedback (each candidate's score and its comparison with the current best), because a Monte Carlo search with diagnostics, credit, and probes would take too long. This also motivates our faster Network World Model. Given more time, we will soon add the full-feedback comparison for all three evaluators. The reported training cost includes data generation and model training time.

EvaluatorIM, Network Science (%) ↑\uparrowCND, Power Grid (%) ↓\downarrowTime per evaluation of 200 rollouts (s), NS / PGTraining cost, NS / PG
Oracle (true parameters)36.6219.160.22 / 3.54none
Monte Carlo (NDlib)36.5819.165.86 / 23.85none
Network World Model36.6119.161.62 / 6.7715 min / 3.3 h

W3 & Q3: Feedback ablation

Could the authors provide ablation studies to evaluate the contributions of rollout diagnostics, action-level credit, and counterfactual probes?

Feedback ablation setup: We add an ablation of the feedback components given to the agent at each round. Specifically, we start from score-only feedback and add rollout diagnostics (the per-step trace of how a plan unfolds, action-level credit, and more) and counterfactual probes (answers to what-if queries on alternative actions), separately and together. All other settings are fixed: IM on Network Science and CND on Power Grid, 10% budget, IC, 20 rounds. We have added this ablation and its analysis to Appendix H, lines 1985-2023.

VariantDiagnosticsProbesIM spread (%) ↑\uparrowCND infected (%) ↓\downarrow
Score onlyNoNo36.6119.16
Score + diagnosticsYesNo38.1918.20
Score + probesNoYes38.7217.56
Full (ours)YesYes39.0617.33
Strongest baseline38.03 (IMM)18.74 (Frontier)

Feedback ablation results: Without rollout diagnostics or counterfactual probes, the loop falls short of the strongest baselines, and adding each feedback component on its own is enough to outperform the strongest baseline.

Counterfactual probes contribute most: on their own, they recover 86-87% of the full loop's improvement over the score-only loop, against 52-64% for rollout diagnostics. The two components are complementary but overlap, so they partly convey the same information about where a plan loses spread.

We thank the reviewer again for the constructive feedback, and we hope these additions clarify the architecture, the role of the oracle, and the value of each feedback component.

AuthorsOpenReview ↗

Official Comment by Authors

We thank the reviewer for the detailed review, and we have addressed the concerns in our response above. As the discussion period is coming to a close, we would be grateful if you could take a look, and we are happy to clarify anything further.

Comments on the paper

2 comments
Area_Chair_qz3JOpenReview ↗

Please join the author-reviewer discussion

Dear all reviewers,

Our author and reviewer discussion will end soon in 18 November (AOE). If you have not yet joined the discussion, now it is the time to start it. We highly appreciate your time on discussing and exchanging thoughts to keep our ICLR standards. Many thanks!

Bests,
Your AC

AuthorsOpenReview ↗

Summary of Rebuttals and Discussion

We thank all reviewers for recognizing our core contributions: (i) a novel integrated algorithm-design environment that combines explicit network interventions, diffusion rollouts, and a coding agent (1i1y Strengths 2; tb3X Strengths), (ii) a broad evaluation of eight tasks and five diffusion models (JPUz Strengths 1; tb3X Strengths), and (iii) readable algorithms that transfer across networks (JPUz Strengths 2; 1i1y Strengths 1).

During the rebuttal, we addressed the following concerns raised by the reviewers:

  • Learned transmission probabilities (all reviewers): The Network World Model receives wuv=1w_{uv} = 1 on every edge and learns the transmission probabilities from observed transitions; generated programs read only these learned probabilities through GraphInfo. In the main experiments, the true probabilities are drawn at random and cannot be recovered from the network structure.
  • Learned model vs. exact oracle (all reviewers): The oracle uses the exact dynamics and exists only when the process is fully known. With full feedback, the algorithms designed with the Network World Model match those designed with the oracle within 0.07 points in every setting.
  • Runtime (tb3X, 1i1y): The Network World Model is 4.0×4.0\times to 14.5×14.5\times faster than Monte Carlo simulation per evaluation.
  • Feedback ablation (tb3X): Each feedback component (rollout diagnostics and counterfactual probes) improves performance, and the full loop outperforms the strongest baseline.
  • Variance and RIS horizon (JPUz): Standard deviations across three independent seeds are small, and the designed algorithms stay ahead of the RIS baselines (IMM, OPIM, SubSIM) on untruncated spread with no limiting horizon.

Reviewers JPUz and 1i1y raised their ratings after the discussion, with both stating that their concerns were resolved. The concerns of Reviewer tb3X (what the GNN learns, runtime, and feedback ablation) were properly addressed with the requested experiments and clarifications, and overlap with the concerns that JPUz and 1i1y considered resolved.

We thank the AC and reviewers for their time and hope the added evidence is helpful in forming the decision.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.