acceptodds
Under review as a conference paper at ICLR 2027

NetAgent: Agentic Wireless Base-Station Planning with Differentiable Propagation Optimization

Abstract

Reliable wireless service and quiet-zone protection depend on base-station placement. This is a constrained inverse-design problem in which sparse transmitter decisions are judged by the dense propagation fields they induce. Locations, power allocation, and station count interact: strengthening one region can illuminate another that should remain quiet. Electromagnetic simulation captures these effects but makes configuration search costly, while fast channel models miss local building geometry. A radio-map generator offers another route when transmitter locations are not prescribed: gradients of regional objectives shape a candidate field whose dominant peak identifies a deployable site. NetAgent trains a RadioDiff-based architecture without its original condition encoder and freezes the resulting radio-map prior. Short deterministic denoising allows joint optimization of spatial latents and source power weights; station coordinates are read from the optimized single-source fields. One shared multi-source optimization produces weight-ranked nested deployments, from which a constraint-first rule selects station count without separate searches. An agentic interface converts drawn or natural-language requests into validated regional constraints. On thirty OpenStreetMap-derived scenarios scored by an independent electromagnetic simulator, NetAgent Auto- reaches mean target attainment in s with stations on average, compared with for AutoBS and for a s exhaustive ray-tracing reference.

Primary area:
applications to robotics, autonomy, planning
Keywords:
differentiable propagation optimization, base-station placement, diffusion models, wireless network planning, LLM agents

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Ji1WOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents NetAgent, a differentiable planning framework for wireless base-station deployment. It optimizes spatial latents and power weights through a frozen RadioDiff-based radio-map prior to generate candidate station locations. A shared multi-source optimization then constructs nested Top-K deployments, from which a constraint-first rule selects the station count. The method is evaluated on 30 OpenStreetMap-derived scenarios using the independent WinProp propagation simulator and compared with AutoBS, a 3GPP TR 38.901 proxy, and Sionna Exhaustive. The results suggest that Auto-K offers a competitive trade-off between deployment quality and computational cost, although its mean attainment remains below that of the exhaustive reference.

Critical strengths

1.Meaningful constrained planning problem. The paper addresses a practical deployment problem involving strong-region coverage, weak-region suppression, signal-strength margins, and station-count selection. Using a differentiable generated propagation field to guide deployment optimization is a reasonable and potentially useful approach.

2.Independent simulation-based evaluation. The final deployments are evaluated using WinProp rather than solely through the learned propagation model. The experiments also compare fixed-K and variable-K planning and report both deployment quality and runtime, providing useful evidence about the practical trade-offs of the proposed framework.

Critical weaknesses

1.The independent contribution of the proposed Auto-K mechanism is not sufficiently isolated. The proposed Auto-K pipeline combines shared multi-source optimization, power-weight-based ranking, nested Top-K construction, and constraint-first selection. Table 6 reports an attainment of 0.5885 for Auto-K, compared with 0.5393 for the independently optimized Full variant, while reducing runtime from 29.84 s to 8.21 s. These results provide positive evidence for the overall pipeline. However, the current ablation does not distinguish the contributions of shared optimization, weight-based ranking, and nested-prefix construction. Consequently, while the results support the effectiveness of the overall Auto-K pipeline, they do not fully establish the independent contribution of its individual design choices or whether simpler shared-optimization and candidate-selection strategies could achieve similar gains.

2.The baseline comparison does not fully isolate the benefit of the proposed planning method from that of learned power allocation. The paper states that NetAgent learns per-source power weights, whereas the baselines use uniform weights, and acknowledges that this favors NetAgent in the main comparison. Consequently, the reported performance gains may reflect both the planning strategy and the additional flexibility of learned power allocation. The weight ablation provides some evidence for the benefit of learned weights, but it is limited to a ten-scenario subset and does not fully establish how much of the overall gain comes from this component. A more controlled comparison would help clarify the contribution of the proposed planning mechanism.

Questions for authors

1.Could the authors provide additional ablations to distinguish the effects of shared optimization and nested-prefix construction in Auto-K? In particular, how much of the reported gain over independently optimized fixed-K deployments is attributable to sharing the optimization across station counts, rather than to the candidate-selection rule itself?

2.For prefixes deemed feasible by the differentiable surrogate, how often do the corresponding deployments satisfy the original constraints under WinProp? Could the authors report the agreement between surrogate-based feasibility and simulator-based feasibility, together with the distribution of constraint violations for the selected Auto-K deployments? This would help clarify the reliability of the station-count selection rule without assuming that a discrepancy necessarily exists.

3.Could the authors clarify how the reported Auto-K attainment of 0.5885 relates to the three random seeds? The paper states that the first seed provides the reported deployment while also discussing results across three seeds. Please clarify which results are based on the first seed and which are aggregated across seeds, and specify the paired statistical test used to obtain the reported (p < 0.05) results.

4.Could the authors clarify the search-space comparison between NetAgent and Sionna Exhaustive? The implementation details describe 1 m-resolution coordinate extraction for NetAgent and an 8 m-spaced candidate grid for Sionna Exhaustive, whereas the reproducibility statement says that all methods share the same candidate grid. Clarifying the intended meaning of this statement and the candidate spaces used by the respective methods would make the comparison easier to interpret.

4 comments

1. Auto-K Mechanism Contributions

Additional real world measurement has been included in two 5 × 5 m obstacle layouts, each tested with one and two transmitters. Nine Strong and nine Weak points per method and setting yield 288 point records and 5,715 valid RSS readings. NetAgent uses the same 50-epoch budget as the main configuration. Two-source measurements are fused in linear power.

Table 6. Real world measurement across both layouts and station counts.

MethodA1 ScoreA2 ScoreB1 ScoreB2 ScoreMean Score ↑Mean Attainment ↑S (%)W (%)M (dB)C18 (%)
NetAgent4.3394.4433.6733.5413.9990.767580.633.315.0247.2
Sionna Exhaustive3.8964.0103.8522.8263.6460.765166.741.714.9338.9
AutoBS3.7933.8752.6062.8723.2870.661363.930.612.1538.9
TR 38.9011.9762.7483.8881.3032.4790.587452.819.49.0629.2

A1/A2 and B1/B2 denote layouts A/B with K=1/2. S, W, M, and C18 are mean components. C18 is the fraction of all 18 sampled points passing the service cutoff. Means weight the four settings equally.

NetAgent achieves the highest mean measured Score of 3.999 and leads three of four settings. At K=1, its A1/B1 mean Score is 4.006, compared with 3.874 for Sionna and 3.200 for AutoBS under identical transmitter settings. This single-source comparison supports a placement advantage without inter-transmitter allocation.

The revised ablations separate shared reuse from alternatives for ordering, candidate construction, and allocation. Full optimizes every station count independently. Auto-K constructs candidates from one joint maximum-count solution. Their revised comparison is given below.

Table 4. Station-count selection and source-selection ablations with a shared deployment region.

Method / variantMean KAttainment ↑Δ versus Auto-KMean time (s)
NetAgent Auto-K1.800.7083030.1
NetAgent Full2.320.7452+0.0369110.6
AutoBS3.070.4570-0.25130.62
Ordinal prefixN/A0.6630-0.045330.1
Best subsetN/A0.6660-0.0423936.0
Equal powerN/A0.6490-0.059330.1

Attainment summarizes the evaluation cohort. Δ is relative to Auto-K. N/A marks mean K values not defined for these controls.

Auto-K retains 95.0% of Full's Attainment with 72.8% less decision time and 22.4% fewer stations. Full's Attainment is 0.7452, compared with 0.7083 for Auto-K. The revised evidence supports a quality–computation trade-off rather than an Attainment gain over Full.

The three controls test different parts of the shared procedure. Ordinal prefix uses source-index order instead of weight order, retaining learned weights after normalization. Best subset evaluates all subsets of each cardinality from the shared solution, again renormalizing their weights. Equal power keeps the ranked candidate locations but assigns weights of 1/k1/k. Auto-K exceeds them by 0.0453, 0.0423, and 0.0593 Attainment, respectively. The broader subset search reports 936.0 s, versus 30.1 s for Auto-K.

For descending weight order π\pi, the main candidates are Ck={π1,…,πk}C_k=\{\pi_1,\ldots,\pi_k\}, with w^j(C)=wj/∑i∈Cwi\hat w_j(C)=w_j/\sum_{i\in C}w_i. Equation 7 selects the smallest k with violation Vk≤10−6V_k\leq10^{-6}. If none qualifies, (Vk,−Uk,k)(V_k,-U_k,k) is minimized lexicographically. Ordering constructs the candidates, normalization defines their powers, and the constraint-first rule selects K.

The candidate definitions and power rules are specified in Appendix G.3. The tested simpler ordering and broader subset-search alternatives do not match the complete pipeline's reported quality–cost combination. The controls characterize interacting choices rather than an additive causal decomposition. The matched-power planning comparison is addressed separately below.

3. Repeated-Run Aggregation

The revised Auto-K Attainment of 0.7083 aggregates 30 scenes and three runs per scene, giving 90 scene–run evaluations. Quantitative aggregates are distinguished from individual deployments shown as illustrations. The value quoted from the original submission is superseded by the rerun result.

Table 4. Station-count selection and source-selection ablations with a shared deployment region.

Method / variantMean KAttainment ↑Δ versus Auto-KMean time (s)
NetAgent Auto-K1.800.7083030.1
NetAgent Full2.320.7452+0.0369110.6
AutoBS3.070.4570-0.25130.62
Ordinal prefixN/A0.6630-0.045330.1
Best subsetN/A0.6660-0.0423936.0
Equal powerN/A0.6490-0.059330.1

Attainment summarizes the evaluation cohort. Δ is relative to Auto-K. N/A marks mean K values not defined for these controls.

Attainment is calculated before cohort averaging. For normalized components r=(S/ρs,W/ρw,M/Mmin,C/ρc)r=(S/\rho_s,W/\rho_w,M/M_{min},C/\rho_c), each deployment has A=14∑j=14clip[0,1](rj)A=\frac14\sum_{j=1}^{4}\mathrm{clip}_{[0,1]}(r_j). The table reports Aˉ=190∑n=190An\bar A=\frac1{90}\sum_{n=1}^{90}A_n. Clipping deployment-level components before averaging preserves the intended target-completion interpretation. An illustrated first-run deployment does not define this aggregate.

Section 5 and Appendix A.3 specify the common cohort and aggregation. The revised comparisons above report observed means under that protocol.

4. Common Search Resolution and Computational Scaling

All urban experiments were rerun on the common 1 m grid. Maps contain 256 × 256 pixels at 1 m per pixel. NetAgent extracts deployment coordinates on this grid, while Sionna Exhaustive searches 1 m-spaced coordinates in the permitted regions. The earlier 1 m versus 8 m mismatch is therefore removed in the revised evaluation.

Table 2. Urban fixed-count Scores with transmitter-specific deployment regions.

MethodK=1K=2K=3K=4K=5Matched mean
NetAgent3.1283.6923.9774.2264.3223.869
Sionna Exhaustive3.4474.008N/AN/AN/AN/A
TR 38.901 Proxy2.815N/AN/AN/AN/AN/A
AutoBS2.9173.3073.6113.9034.0263.553
Gain over AutoBS7.2%11.6%10.1%8.3%7.4%8.9%

Matched means cover all five counts. NetAgent is emphasized as the proposed method. N/A denotes an exhaustive search exceeding the 1,200 s budget, so its Score was not recorded.

The revised grid is aligned while the planning procedures remain distinct. NetAgent differentiates regional objectives through its frozen prior to update source representations jointly. Sionna enumerates candidate combinations. The recorded quality comparison retains Sionna's leading Scores at K=1 and K=2, while NetAgent exceeds AutoBS at every reported count.

Table 3. Fixed-K weight ablation and uniform-power AutoBS placement control.

Score / differenceK=2K=3K=4K=5Mean
NetAgent learned weights3.6923.9774.2264.3224.054
NetAgent uniform weights3.6113.8684.1664.2003.961
AutoBS uniform power3.3073.6113.9034.0263.712
Learned-weight increment+0.081+0.109+0.060+0.122+0.093
Uniform-power placement gain+0.304+0.257+0.263+0.174+0.250

Weight increment = NetAgent learned minus uniform. Placement gain = uniform-power NetAgent minus AutoBS. All conditions use the same 30 scenes and three runs per scene. Reported means and differences retain the manuscript's rounding.

Power allocation is also controlled on 30 scenes and three runs per scene, replacing the earlier ten-scene subset. Uniform-power NetAgent scores 3.961 versus 3.712 for uniform-power AutoBS, an approximately 6.7% gain. Learned weights then add 0.093 Score, or 2.35%. The placement advantage therefore remains under matched allocation, with a separate learned-weight increment.

Table 19. Candidate tuples in exhaustive joint search under equal candidate-set sizes.

Transmitters KTuplesRelative to K=1
1NN11
2N2N^2NN
3N3N^3N2N^2
4N4N^4N3N^3
5N5N^5N4N^4

This is symbolic search-space size. Unequal candidate sets give ∏iNi\prod_i N_i tuples.

Table 5. Decision times by geometry and station count.

Transmitter-specific deployment regions

MethodK=1K=2K=3K=4K=5
NetAgent14.7 s17.3 s20.1 s24.3 s29.2 s
AutoBS0.67 s0.71 s0.74 s0.76 s0.81 s
Sionna16.9 s62.7 s47.3 h†26.0 yr†1.52 × 10⁵ yr†
TR 38.901 Proxy102.0 s4.11 d†40.0 yr†1.80 × 10⁵ yr†9.92 × 10⁸ yr†

Shared deployment region

MethodK=1K=2K=3K=4K=5
NetAgent16.6 s17.9 s21.3 s24.7 s30.1 s
AutoBS0.57 s0.59 s0.61 s0.64 s0.68 s
Sionna354.3 s2.30 h†11.8 yr†5.77 × 10⁵ yr†2.81 × 10¹⁰ yr†
TR 38.901 Proxy50.5 min†3.78 yr†1.84 × 10⁵ yr†9.00 × 10⁹ yr†4.39 × 10¹⁴ yr†

Unmarked entries are recorded times. † Throughput-based exhaustive-search projections after the 1,200 s timeout, not observed completion times.

The product search space grows exponentially with source count. NetAgent completes in 14.7–30.1 s, while exhaustive multi-source searches exceed the budget and require marked throughput projections. The source-specific K=5 Sionna projection is 1.52 × 10⁵ years. This supports repeated planning within a minute-scale update cycle, with independent WinProp retained as the quality evaluator. The complete runtime table distinguishes recorded values from projections rather than treating projected searches as completed quality evaluations.

2. Independence Issue Between Winprop and Surrogates

WinProp serves as the independent electromagnetic evaluator, while the differentiable surrogate is used to guide deployment optimization. The two components have distinct roles. Deployment coordinates and power allocations are first generated by the optimizer and then evaluated independently in WinProp under the original regional requirements. The reported simulator-based results are determined by WinProp outputs, rather than by agreement with surrogate predictions. Accordingly, numerical agreement with the surrogate is not imposed as an evaluation criterion, and surrogate predictions do not override the external evaluation. This separation allows the resulting deployments to be assessed beyond the propagation model used during optimization.

Reviewer_DsrUOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The authors present NetAgent, a framework for optimizing wireless base-station deployments. The proposed approach introduces an agentic interface that translates natural-language instructions or user-drawn regions into validated, region-specific coverage constraints. It further leverages a pretrained diffusion-based radio propagation model to optimize base-station placement and transmit power, reducing the computational cost associated with expensive site-specific electromagnetic simulations.

Critical strengths

One strength of this paper is its formulation of wireless base-station deployment as a field-space constrained optimization problem, which allows different coverage objectives to be handled in a unified framework. The proposed Auto-K strategy is also a useful addition, as it automatically determines the number of base stations while optimizing their locations, reducing the need for manual configuration. Another advantage is the integration of an LLM-based interface, which provides a flexible and intuitive way for users to specify different coverage requirements across regions.

Critical weaknesses

One concern is that base-station deployment is typically an offline planning task without strict time constraints, so it is unclear how significant the reduction in simulation time is in practice. Second, the technical novelty seems somewhat limited, as several key components, such as differentiable propagation modeling and diffusion-based priors, are built upon existing work. The paper could better clarify its contributions beyond integrating these components. Finally, although the framework supports natural-language and drawing-based inputs, the LLM-based interface is not sufficiently evaluated. In particular, the paper provides limited discussion of prompt design, how ambiguous or conflicting instructions are handled, and how missing parameters are inferred. Experiments on the robustness of language parsing would also help strengthen this part of the work.

Questions for authors
  1. Could the authors clarify why reducing simulation time is particularly important in this setting?
  2. How does the LLM-based interface handle ambiguous or conflicting user instructions, as well as requests with missing parameters?
2 comments

1. Practical Runtime and Joint Planning Contributions

Additional real world measurement has been included in two 5 × 5 m obstacle layouts, each tested with one and two transmitters. Nine Strong and nine Weak points per method and setting yield 288 point records and 5,715 valid RSS readings. NetAgent uses the same 50-epoch budget as the main configuration. Two-source measurements are fused in linear power.

Table 6. Real world measurement across both layouts and station counts.

MethodA1 ScoreA2 ScoreB1 ScoreB2 ScoreMean Score ↑Mean Attainment ↑S (%)W (%)M (dB)C18 (%)
NetAgent4.3394.4433.6733.5413.9990.767580.633.315.0247.2
Sionna Exhaustive3.8964.0103.8522.8263.6460.765166.741.714.9338.9
AutoBS3.7933.8752.6062.8723.2870.661363.930.612.1538.9
TR 38.9011.9762.7483.8881.3032.4790.587452.819.49.0629.2

A1/A2 and B1/B2 denote layouts A/B with K=1/2. S, W, M, and C18 are mean components. C18 is the fraction of all 18 sampled points passing the service cutoff. Means weight the four settings equally.

NetAgent achieves the highest mean measured Score of 3.999 and leads three of four settings. At K=1, its A1/B1 mean Score is 4.006, compared with 3.874 for Sionna and 3.200 for AutoBS under identical transmitter settings. This single-source comparison supports a placement advantage without inter-transmitter allocation.

Repeated electromagnetic-situation and deployment updates matter in evolving 6G networks as transmitters, demands, protected regions, and available sites change. A useful solution must be obtained within the update cycle, beyond the initial offline design.

Table 19. Candidate tuples in exhaustive joint search under equal candidate-set sizes.

Transmitters KTuplesRelative to K=1
1NN11
2N2N^2NN
3N3N^3N2N^2
4N4N^4N3N^3
5N5N^5N4N^4

This is symbolic search-space size. Unequal candidate sets give ∏iNi\prod_i N_i tuples.

Table 5. Decision times by geometry and station count.

Transmitter-specific deployment regions

MethodK=1K=2K=3K=4K=5
NetAgent14.7 s17.3 s20.1 s24.3 s29.2 s
AutoBS0.67 s0.71 s0.74 s0.76 s0.81 s
Sionna16.9 s62.7 s47.3 h†26.0 yr†1.52 × 10⁵ yr†
TR 38.901 Proxy102.0 s4.11 d†40.0 yr†1.80 × 10⁵ yr†9.92 × 10⁸ yr†

Shared deployment region

MethodK=1K=2K=3K=4K=5
NetAgent16.6 s17.9 s21.3 s24.7 s30.1 s
AutoBS0.57 s0.59 s0.61 s0.64 s0.68 s
Sionna354.3 s2.30 h†11.8 yr†5.77 × 10⁵ yr†2.81 × 10¹⁰ yr†
TR 38.901 Proxy50.5 min†3.78 yr†1.84 × 10⁵ yr†9.00 × 10⁹ yr†4.39 × 10¹⁴ yr†

Unmarked entries are recorded times. † Throughput-based exhaustive-search projections after the 1,200 s timeout, not observed completion times.

NetAgent completes in 14.7–30.1 s, supporting minute-scale planning updates with independently evaluated quality. The source-specific K=5 Sionna projection is 1.52 × 10⁵ years, reflecting exponential multi-source search growth. Section 6.3 excludes startup and checkpoint loading, uses a 1,200 s search budget, and uses a 60 s throughput probe for marked projections. AutoBS is faster, while NetAgent supplies the stronger demonstrated deployment quality.

The contribution extends beyond component integration to joint inverse placement. Source-specific latent inputs and power logits are optimized together through a frozen diffusion radio-map prior, and coordinates are extracted from the resulting map peaks. Independent-source selection omits cross-source coordination, while conditional selection freezes earlier sites. Joint updates admit coordinated relocations excluded by these restrictions. Shared count selection then reuses one optimized solution across counts, connecting the formulation to practical planning efficiency.

2. Prompt Design and Handling Incomplete Instructions

The interface converts language and drawings into a typed task before optimization. The system prompt defines coordinates, region roles, shapes, source IDs, and the JSON schema. Runtime instructions specify allowed sources. Update prompts include the previous state so untouched fields are preserved. The loss planner generates bounded scene-specific parameters within a fixed objective family.

Material ambiguity and missing requirements require clarification. Named presets may fill omissions only when disclosed and accepted. Validated user settings override proposals. Identical Strong/Weak masks with disjoint qualifying intervals and required fractions summing above one identify a contradiction. The confirmed effective task is reused at execution, and edits invalidate prior approval.

Table 12. Corrected language evaluation.

Case groupStrict match (%)Exact match (%)Mean IoUWrong acceptance (%)
Clear100.0100.01.0000.0
Missing parameters0.00.00.861100.0
Ambiguous0.00.00.866100.0
Conflicting0.00.00.5540.0
Multi-turn100.0100.01.0000.0

Conflict rejection is scored separately from parse matching. All 60 conflict attempts were rejected. IoU measures mask agreement rather than the correctness of an execution decision.

Clear and multi-turn cases reach 100% recorded match, and all 60 tested conflicts are rejected. The complete table separates geometric matching from action correctness and preserves the outcomes for missing and ambiguous cases.

Table 13. Static-prompt evaluation under the corrected representation.

Case groupStrict match (%)Exact match (%)Mean IoUWrong acceptance (%)
Clear99.799.70.9990.0
Missing parameters0.00.00.861100.0
Ambiguous0.00.00.866100.0
Conflicting0.00.00.5540.0
Multi-turn100.0100.01.0000.0

Rates are percentages and IoU is unitless. This condition removes runtime deployment instructions while retaining the corrected representation.

Across 840 paired attempts, runtime augmentation preserves 838 outcomes and improves two repeated colloquial-input cases, increasing their IoU from approximately 0.7848 to 0.9999. This quantifies the tested prompt contrast.

Table 14. Strict match by expression form in the second benchmark.

ExpressionOriginalSynonymMixed languageColloquialTypo
Strict match (%)100.097.9100.0100.071.5

The second benchmark is reported separately from the corrected 840-attempt benchmark. Its sample counts and annotation records remain unspecified in the manuscript.

Table 15. Human-checked interaction cases.

CaseInputParsed taskExpected actionObserved outcome
Clear ellipseDeploy 3 BSs, ellipse3 main, 3 bright, 3 dark. Ellipse/rectangle. -50/-60 dBm, 90%, 15 dBExecuteMatch, IoU 1.000
Missing weakDeploy 3 BSs, strong regions. No Weak requirement3 main, 2 bright, 0 darkClarifyExecuted automatically. Incorrect action
Ambiguous sourcePut BS3 in the middle area3 main with BS3 at the default center, 3 bright, 3 darkClarifyExecuted. Incorrect action
ConflictStrong and Weak are the same circle3 main, 2 bright, 3 dark. Strong/Weak overlapRejectRejected
Multi-turnInitial BS3 is wrong. Feedback corrects it8 constraints. BS3 correctedUpdateMatch, IoU 1.000
Human-fragmented3 BSs. BS1 circle (59.4,34.5), r=20. BS2 8-gon (179.3,105.7), r=20.5, rotation 148.2°. BS3 rectangle (223.5,83.1), 36.7 × 26.2, rotation 111.2°. Bright 6-point polygon and rectangle. Dark circle, ellipse, and 5-gon. Margin 158 constraints comprising 3 main, 2 bright, 3 dark. -50/-60 dBm, 90%/90%, 15 dBExecuteMean IoU 0.990, minimum 0.971

Input and parsed fields retain the manuscript's abridged presentation. Main, bright, and dark denote deployment, Strong, and Weak regions. Expected actions and observed behavior are distinct.

The human-fragmented case reaches mean IoU 0.990 and minimum 0.971, while the corrected multi-turn case reaches 1.000. Expected actions and observed outcomes remain distinct. Appendices B and C therefore connect prompt design, parameter provenance, corrections, conflict handling, and expression changes to their corresponding evidence.

Reviewer_Uii2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents NetAgent, a diffusion-model-assisted framework for wireless base-station deployment planning under spatial coverage and suppression constraints. Instead of directly searching for base-station coordinates, this method optimizes the latent variables of a frozen RadioDiff-based model and extracts deployment locations from the resulting propagation fields. It also jointly optimizes power weights and introduces Auto-K, which derives nested deployment candidates from a shared optimization to automatically select the number of base stations. Experiments on 30 OpenStreetMap-derived scenarios, evaluated using an independent electromagnetic simulator, demonstrate competitive deployment quality and computational efficiency compared with several existing planning methods.

Critical strengths
  1. Interesting methodological perspective. The paper presents an interesting approach to wireless deployment planning by optimizing spatial latent variables of a frozen diffusion model rather than directly searching for base-station coordinates. Integrating this approach with joint power optimization and automatic station-count selection provides a meaningful methodological contribution.
  2. Promising quality-efficiency trade-off. The method achieves competitive deployment quality with substantially reduced computational cost compared with exhaustive search. Evaluation using an independent electromagnetic simulator, together with comparisons across different station counts, provides useful empirical support for its practical potential.
Critical weaknesses
  1. Insufficient isolation of the core methodological contributions. Although the proposed Auto-K achieves better performance than independently optimizing each station count, the current ablation does not adequately explain whether the improvement comes from shared optimization, weight-based ranking, or the construction of nested deployment candidates. Furthermore, a controlled comparison with a differentiable coordinate-space optimization baseline would better establish the specific benefits of latent-space optimization. Although differences in candidate-location resolution and power-allocation strategies are described, their effects on the reported performance gains are not sufficiently isolated.
  2. Limited evidence of constraint satisfaction and the benefits of automatic station-count selection. The reported weak-region suppression performance is relatively poor, particularly when multiple base stations are deployed. Although the paper acknowledges that some constraints may be infeasible, the results do not sufficiently distinguish optimization limitations from inherent infeasibility. In addition, Auto-K frequently selects a single base station, raising questions about its advantages over simpler station-count selection strategies.
Questions for authors
  1. Auto-K ablation. Could the authors further isolate the contributions of shared optimization, weight-based ranking, and nested candidate selection? In particular, how does Auto-K compare with simpler alternatives, such as always selecting K=1 or selecting K from independently optimized deployments under the same evaluation criterion?
  2. Baseline comparisons. Could the authors clarify how differences in candidate-location resolution and power-allocation strategies affect the comparisons? A controlled comparison with a differentiable coordinate-space optimization baseline under comparable evaluation settings would help to establish the specific advantages of latent-space optimization.
  3. Constraint satisfaction. Given the relatively low weak-region suppression performance in the fixed-K experiments, could the authors provide a more detailed analysis of constraint satisfaction for Auto-K, distinguishing feasible and infeasible scenarios? This would help clarify whether the remaining violations arise primarily from the optimization method or from conflicting deployment requirements.
3 comments

1. Auto-K Mechanisms and Adaptive Station Counts

Additional real world measurement has been included in two 5 × 5 m obstacle layouts, each tested with one and two transmitters. Nine Strong and nine Weak points per method and setting yield 288 point records and 5,715 valid RSS readings. NetAgent uses the same 50-epoch budget as the main configuration. Two-source measurements are fused in linear power.

Table 6. Real world measurement across both layouts and station counts.

MethodA1 ScoreA2 ScoreB1 ScoreB2 ScoreMean Score ↑Mean Attainment ↑S (%)W (%)M (dB)C18 (%)
NetAgent4.3394.4433.6733.5413.9990.767580.633.315.0247.2
Sionna Exhaustive3.8964.0103.8522.8263.6460.765166.741.714.9338.9
AutoBS3.7933.8752.6062.8723.2870.661363.930.612.1538.9
TR 38.9011.9762.7483.8881.3032.4790.587452.819.49.0629.2

A1/A2 and B1/B2 denote layouts A/B with K=1/2. S, W, M, and C18 are mean components. C18 is the fraction of all 18 sampled points passing the service cutoff. Means weight the four settings equally.

NetAgent achieves the highest mean measured Score of 3.999 and leads three of four settings. At K=1, its A1/B1 mean Score is 4.006, compared with 3.874 for Sionna and 3.200 for AutoBS under identical transmitter settings. This single-source comparison supports a placement advantage without inter-transmitter allocation.

Shared optimization is evaluated against Full, which independently optimizes every station count. Auto-K instead reuses one jointly optimized five-source solution. The rerun results below replace the earlier submission's values.

Table 4. Station-count selection and source-selection ablations with a shared deployment region.

Method / variantMean KAttainment ↑Δ versus Auto-KMean time (s)
NetAgent Auto-K1.800.7083030.1
NetAgent Full2.320.7452+0.0369110.6
AutoBS3.070.4570-0.25130.62
Ordinal prefixN/A0.6630-0.045330.1
Best subsetN/A0.6660-0.0423936.0
Equal powerN/A0.6490-0.059330.1

Attainment summarizes the evaluation cohort. Δ is relative to Auto-K. N/A marks mean K values not defined for these controls.

Auto-K retains 95.0% of Full's Attainment with 72.8% less decision time and 22.4% fewer stations. The revised comparison establishes a quality–cost trade-off. Full remains the independently optimized quality reference.

The three controls test different parts of the shared procedure. Ordinal prefix uses source-index order instead of weight order, retaining learned weights after normalization. Best subset evaluates all subsets of each cardinality from the shared solution, again renormalizing their weights. Equal power keeps the ranked candidate locations but assigns weights of 1/k1/k. Auto-K exceeds them by 0.0453, 0.0423, and 0.0593 Attainment, respectively. The broader subset search reports 936.0 s, versus 30.1 s for Auto-K.

For descending weight order π\pi, the main candidates are Ck={π1,…,πk}C_k=\{\pi_1,\ldots,\pi_k\}, with w^j(C)=wj/∑i∈Cwi\hat w_j(C)=w_j/\sum_{i\in C}w_i. Equation 7 selects the smallest k with violation Vk≤10−6V_k\leq10^{-6}. If none qualifies, (Vk,−Uk,k)(V_k,-U_k,k) is minimized lexicographically. Ordering constructs the candidates, normalization defines their powers, and the constraint-first rule selects K.

The reported mean K=1.80 also rules out a constant K=1 interpretation. For counts in 1–5, Section 6.2 gives Pr⁡(K>1)≥(E[K]−1)/4\Pr(K>1)\geq(\mathbb E[K]-1)/4, implying at least 20% multi-station deployments. The available independent-count comparison is Full. Taken together, these results support shared reuse and the evaluated ordering, candidate-generation, and allocation rules, while keeping their interacting effects distinct.

2. Controlled Grid and Power Comparisons

All urban methods were rerun on the common 1 m grid. NetAgent extracts coordinates from optimized 1 m radio maps, while Sionna Exhaustive evaluates 1 m-spaced candidates in the permitted regions. Thus, the revised comparison removes the previous coarse-grid mismatch. The remaining difference is how deployments are obtained.

Table 2. Urban fixed-count Scores with transmitter-specific deployment regions.

MethodK=1K=2K=3K=4K=5Matched mean
NetAgent3.1283.6923.9774.2264.3223.869
Sionna Exhaustive3.4474.008N/AN/AN/AN/A
TR 38.901 Proxy2.815N/AN/AN/AN/AN/A
AutoBS2.9173.3073.6113.9034.0263.553
Gain over AutoBS7.2%11.6%10.1%8.3%7.4%8.9%

Matched means cover all five counts. NetAgent is emphasized as the proposed method. N/A denotes an exhaustive search exceeding the 1,200 s budget, so its Score was not recorded.

NetAgent exceeds AutoBS at every K, with gains of 7.2–11.6%. Sionna has the highest recorded Scores at K=1 and K=2. Exhaustive Scores are left unavailable when the search exceeds its evaluation budget, and time projections are not substituted for quality results.

Power allocation is isolated through two matched comparisons on the same 30 scenes and three runs per scene. Uniform-power NetAgent versus uniform-power AutoBS tests the deployment-method difference. Learned-weight versus uniform-weight NetAgent measures the added allocation benefit.

Table 3. Fixed-K weight ablation and uniform-power AutoBS placement control.

Score / differenceK=2K=3K=4K=5Mean
NetAgent learned weights3.6923.9774.2264.3224.054
NetAgent uniform weights3.6113.8684.1664.2003.961
AutoBS uniform power3.3073.6113.9034.0263.712
Learned-weight increment+0.081+0.109+0.060+0.122+0.093
Uniform-power placement gain+0.304+0.257+0.263+0.174+0.250

Weight increment = NetAgent learned minus uniform. Placement gain = uniform-power NetAgent minus AutoBS. All conditions use the same 30 scenes and three runs per scene. Reported means and differences retain the manuscript's rounding.

Uniform-power NetAgent reaches 3.961 versus 3.712 for AutoBS. The manuscript-reported mean placement gain is approximately 0.250 Score, or 6.7%. Learned weights then add 0.093 Score, or 2.35%, raising the mean to 4.054. The deployment advantage therefore persists before learned allocation is introduced.

The methodological distinction is joint inverse planning through a frozen diffusion prior. Spatial latent inputs and source-power logits are optimized through the combined regional field, and map peaks determine coordinates. This permits coordinated changes across sources. Independent-source selection uses isolated fields, while conditional selection freezes earlier sites. Appendix D.4 defines these restrictions. The supported empirical comparison above concerns matched-resolution planning methods and matched power rules.

3. Weak Suppression and the Meaning of Constraint Violations

The revised paper distinguishes aggregate performance, target completion, and joint feasibility. Score is uncapped and retains above-target gains. Attainment clips each normalized component before averaging. Joint feasibility separately requires every applicable regional constraint to be satisfied.

Figure 4 explains the fixed-K service–protection trade-off. As K increases, NetAgent's Strong coverage rises from 65.4% to 94.0%, while Weak suppression falls from 39.9% to 10.5%. AutoBS suppression falls from 42.1% to 0.4%. Received powers add in the linear domain, and the total transmit-power budget grows with K. More sources can therefore improve service while raising received power in protected regions.

Auto-K evaluates each candidate against the same regional requirements rather than assuming that a larger count is better. The smallest prefix meeting the violation tolerance is selected. If no prefix qualifies, violation is minimized before utility and count are used. A selected deployment's residual violation identifies a shortfall for that candidate.

Table 1. Regional task specification and validation conditions.

FieldStored contentValidation condition
DeploymentSource ID and permitted maskNonempty free-space intersection
Service/protectionRegion masks and Strong/Weak roleValid geometry and map coordinates
RequirementsThreshold, required fraction, marginExplicit values or named preset
PrioritiesRegional objective prioritiesDefined before optimization
ConflictsOverlap and threshold relationsJoint requirements retained for evaluation

The task specification also identifies a concrete source of inherent infeasibility. For identical Strong and Weak masks with disjoint qualifying signal intervals, no point can satisfy both signal tests. Their successful fractions sum to at most one. Required fractions summing to more than one therefore certify a contradictory task, independently of the optimizer. Partial mask overlap alone does not establish this contradiction.

A residual violation and a contradictory task are different findings. The former describes an evaluated deployment, while the latter follows from incompatible requirements. The revised definitions make this distinction explicit, so neither a high Score nor low Weak suppression is used by itself to classify an entire scenario as feasible or inherently infeasible. Section 5.1 defines the metrics, Section 4.3 specifies the selection rule, and Table 1 records the task checks.

Comments on the paper

1 comment

We thank all reviewers for their constructive comments. Additional real world measurement and controlled experiments have been included to strengthen the evidence for deployment quality, clarify the contributions of individual mechanisms, and evaluate the language interface.

The expanded experiments include matched uniform-power comparisons to separate placement gains from learned allocation; Auto-K controls using ordinal prefixes, best subsets, and equal power; independent-source and conditional optimization controls; optimization-budget and regional-priority studies; and runtime and exhaustive-search complexity comparisons. Language evaluation additionally covers prompt design, expression variations, multi-turn interaction, conflict handling, and six human-checked cases. The urban comparisons were rerun using a common 1 m grid.

For real world measurement, two 5 × 5 m obstacle layouts were evaluated with one and two transmitters, producing four settings. Measurements were conducted at 5.8 GHz and nominal power of 10 dBm, with nine service-region and nine protected-region sampling points per method and setting. This yielded 288 point records and 5,715 valid RSS readings. Two-transmitter measurements were combined in linear power, and NetAgent used the same 50-epoch optimization budget as the main configuration. NetAgent achieved the highest mean measured Score of 3.999; the common-power single-transmitter comparison further supports a placement advantage without inter-transmitter power allocation. Separately, shared Auto-K retained 95.0% of Full’s Attainment with 72.8% less decision time.

To facilitate review, blue captions identify simulation-based experimental figures and tables, red captions identify real world measurement figures and tables, and green text highlights important experimental findings. Within tables, boldface, red numerical highlights, and shaded cells or rows emphasize key results and comparisons, with ranking conventions specified in the corresponding notes. These markings guide readers to the supporting evidence rather than indicating that every colored passage is newly added.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.