System One Models for Wireless Decision-Making:Applications and Performance Evaluation

I. INTRODUCTION

• Wireless control repeatedly maps network states to bounded actions, motivating a native decision interface rather than indirect autoregressive generation.Prior approaches include optimization and task-specific learning, while foundation models aim to reuse representations across tasks.Source: §I · p. 1
• System-One models condition on a state, query, and admissible answer set to return probabilities over typed outcomes instead of free-form token sequences.The interface can support multiple decision schemas and confidence-aware execution; Jev is reported to use a parallel sampler and RLCD training.Source: §I · p. 1
• TypeSafe reports 70–500 ms end-to-end response times and 40–200× speedups on System-One-oriented workloads, while Deel reports quality gains on two production tasks.The TypeSafe figures are vendor-reported and workload dependent; Deel’s datasets, sample sizes, comparator models, and prompting configurations are undocumented, so its results are not independently reproducible from the reported information. Fig. 1 labels approved-answer matching and expense-categorization agreement; its reported figures are 70% to 97% and 50% to 86%, respectively.Source: §I · p. 2, Figure 1 · p. 2
• AnyJev offers an open-source route to typed decisions using pretrained LLMs without fine-tuning backbone weights.It supports zero-label debiasing and lightweight per-question calibration or closed-form decision heads.Source: §I · p. 2
• Wireless tasks fit this interface when actions are bounded, but continuous high-dimensional outputs and hard physical or protocol constraints call for other methods or hybrid control.The proposed hybrid can route routine high-confidence decisions through a fast path and escalate ambiguous, out-of-distribution, or high-risk cases.Source: §I · p. 2
• The article formalizes System-One models, maps suitable tasks across wireless domains, and evaluates receive-antenna selection and intent-conditioned RAN slicing against baselines and generative LLMs.The evaluation considers decision quality, latency, and end-to-end utility.Source: §I · p. 2

• The model conditions a distribution over admissible outcomes on the observed state, decision query, and finite answer space.The query specifies the decision criterion, while the answer space defines allowable outputs.Source: §II · p. 2
• For ordered outcomes, the predictive distribution yields a conditional expectation over the explicitly declared numerical levels, not an unrestricted continuous regression output.Binary decisions are the special case K = 2.Source: §II · pp. 2–3
• Strictly proper scoring rules such as cross-entropy and the Brier score suit distribution learning, while likelihood-based, calibration-oriented, reinforcement-learning, and hybrid procedures are alternative fitting strategies.The learning objective is distinct from the optimization procedure used to fit the model.Source: §II · p. 3
• Unlike a fixed-label classifier, a shared System-One model can address different tasks by changing the query and answer space without changing its parameters or training a task-specific output head.Accuracy and calibration on each new task still require evaluation.Source: §II · p. 3
• Unlike an autoregressive language model, a System-One interface directly exposes probabilities over declared decisions rather than deriving decisions from sequential token generation.Autoregressive models can perform classification, but the decision is derived through token generation rather than being their native output space.Source: §II · p. 3
• A normalized predictive distribution does not guarantee calibration, which must be assessed empirically for the data distribution and deployment regime.Probabilities may remain systematically over- or under-confident despite being normalized.Source: §II · p. 3

• Candidate applications span physical-layer control, radio resource management, mobility, and network operations.Source: §III · p. 3
• The discussion first identifies characteristics of suitable decision tasks, then organizes candidate applications using those criteria.Source: §III · p. 3
• The criteria also guide the selection of representative case studies in section IV.Source: §III · p. 3

A. Characteristics of Suitable Wireless Decisions

• Suitability depends less on protocol layer than on the decision's structure, timing, and risk profile.Source: A. Characteristics of Suitable… · p. 3

1) A bounded and explicit decision space:

• A finite set of meaningful alternatives lets the model return distributions over feasible outcomes and rejects out-of-schema actions by construction.Source: 1) A bounded and explicit decision space · p. 3
• High-dimensional continuous outputs, such as unconstrained precoding matrices or waveforms, are less direct targets unless discretized, factorized, or paired with a downstream optimizer.Source: 1) A bounded and explicit decision space · p. 3

2) Repeated state-conditioned decisions:

• Repeated state-conditioned decisions favor amortized inference, shifting computation to training so online actions can be produced quickly.One-off open-ended design, explanation, or long-horizon decomposition is better suited to optimization tools, simulators, or deliberative agents; System-One may serve as a bounded component.Source: 2) Repeated state-conditioned decisions · pp. 3–4

3) A compatible end-to-end control timescale:

• Control-loop suitability depends on the complete delay from state acquisition through preprocessing, inference, validation, signaling, and actuation.System-One models are initially more plausible for millisecond-to-second budgets than symbol-level processing; tighter loops require the complete implementation to meet the deadline.Source: 3) A compatible end-to-end control… · p. 4

4) Uncertainty that changes the downstream action:

• Probabilities are most useful when confidence changes whether a decision is executed, verified, deferred, or sent to a fallback.Low-confidence handling should be defined, and calibration under deployment shift matters alongside average accuracy.Source: 4) Uncertainty that changes the… · p. 4

5) Reusable structure across related decision schemas:

• Runtime-specified queries or answer sets can help when decision schemas vary but share structure, though schema flexibility does not establish zero-shot competence.Transfer to new queries, environments, and action sets must be measured explicitly.Source: 5) Reusable structure across related… · p. 4

6) Enforceable constraints and a safe fallback:

• Hard power, spectrum, reliability, fairness, and protocol requirements should be checked in a feasibility layer, with optimization, rules, or human review retained as fallback.Suitability is strengthened by explicit constraint checking and a credible fallback.Source: 6) Enforceable constraints and a safe… · p. 4

B. PHY-Level Decisions

• For link adaptation, a System-One model returns a distribution over admissible MCS choices, enabling conservative fallback, top-k verification, or abstention when the state is unfamiliar.The complete feature-construction, inference, and scheduling path must fit within the link-adaptation deadline.Source: B. PHY-Level Decisions · p. 4
• For finite beam codebooks, the model can choose one beam at high confidence or retain a small candidate subset when confidence is diffuse.The decision state may include partial CSI, received powers, position, blockage indicators, and prior beam measurements.Source: B. PHY-Level Decisions · p. 4
• Raw channel estimation, waveform generation, and unconstrained precoder design are not direct System-One tasks; another model or a decomposition must handle their continuous outputs.A System-One model can instead select a mode, codebook entry, or solver configuration while another component computes continuous parameters.Source: B. PHY-Level Decisions · pp. 4–5

C. MAC, RRM, and Mobility Decisions

• Jointly categorizing every user, resource block, beam, and power assignment is usually infeasible, so System-One should select a rule, group, profile, or candidate schedule before constrained allocation.The number of joint assignments grows combinatorially.Source: C. MAC, RRM, and Mobility Decisions · p. 5
• For RAN slicing, a typed decision model can choose among scheduling policies or bounded resource-sharing profiles for service intents, with abstention if none is predicted to satisfy them.Unrestricted per-resource-block power vectors should remain with a continuous controller.Source: C. MAC, RRM, and Mobility Decisions · p. 5
• For mobility, a distribution over candidate serving cells makes confidence explicit, allowing ambiguous cases to retain the current cell, trigger measurements, or invoke a slower policy.The decision objective accounts for handover failure, ping-pong, and outage costs.Source: C. MAC, RRM, and Mobility Decisions · p. 5

D. Network Management Decisions

• For network management, approved remediation actions can be bounded choices, with low confidence routed to observation or escalation rather than irreversible response.Open-ended, ambiguous, cross-domain intent interpretation favors LLM-based agents; System-One choices fit after objectives and admissible configurations are explicit.Source: D. Network Management Decisions · p. 5

IV. EXPERIMENTAL EVALUATION

• The evaluation covers receive-antenna selection and intent-conditioned RAN slicing in simulation, plus independently reported intent-driven edge-service orchestration.Together, these scenarios cover numerical selection, joint semantic and numerical decisions, and integration with network-service execution.Source: §IV · p. 5

A. Evaluation Methodology

• Methods are tested on the same task instances and scored by a common numerical evaluator; Jev returns categorical decisions while LLMs return an action identifier in JSON.API latency is measured at the client for successful requests and includes serving infrastructure and communication path; classical methods are timed locally.Source: A. Evaluation Methodology · p. 6

B. Scenario I: Receive-Antenna Selection

• The receive-antenna scenario selects two of eight candidate receive antennas for a two-transmit-antenna MIMO link across 30 channel–SNR instances.The setup uses 28 admissible antenna pairs per instance and compares against exhaustive search, norm-based, correlation-based, and random baselines.Source: B. Scenario I: Receive-Antenna Selection · p. 6
• 1.52 s was Jev's client-observed latency, approximately 8.5× faster than GPT-Sol's 12.86 s, while Jev's 20.19% mean relative capacity loss exceeded GPT-Sol's 0.00% and norm-based selection's 9.46%.The antenna-selection results illustrate a decision-quality and response-latency tradeoff.Source: Figure 3 · p. 6, Figure 2 · p. 6, B. Scenario I: Receive-Antenna Selection · p. 6
• The authors caution that the receive-antenna results cover only one task and should not be treated as a general limitation of the model class.They identify improved numerical reasoning and wireless-specific grounding as directions for further work.Source: B. Scenario I: Receive-Antenna Selection · p. 7

C. Scenario II: Intent-Conditioned RAN Slicing

• The benchmark allocates 100 resource blocks in increments of 20, yielding 21 admissible actions across 30 instances.The instances combine five service-intent profiles and six network-state variants.Source: C. Scenario II: Intent-Conditioned RAN… · p. 7
• Each AI method receives natural-language intent, traffic demand, channel conditions, and queue information; the simulator scores allocations using service satisfaction and applicable SLA penalties.Exhaustive enumeration is the full-information reference, while load-aware and random allocation are practical baselines; the load-aware policy does not interpret textual intent.Source: C. Scenario II: Intent-Conditioned RAN… · p. 7
• Utility regret, SLA violation severity, satisfaction of hard requirements, and latency are measured separately because the utility optimum need not meet every hard requirement.Overloaded instances test decisions when no admissible allocation satisfies all requirements.Source: C. Scenario II: Intent-Conditioned RAN… · p. 7
• Jev’s mean utility regret of 0.131 is comparable to Qwen (0.134) and GPT-Sol (0.145), while its median client-observed latency is 1.46 s versus 5.13 s for Qwen and 5.29 s for GPT-Sol.The learned decision models outperform random and load-aware baselines in decision quality.Source: C. Scenario II: Intent-Conditioned RAN… · p. 7

D. Scenario III: Intent-Driven Edge-Service Orchestration

• Jev translates natural-language intents into service contracts, while a numerical scheduler assigns tasks using current network and computing conditions.The evaluation combines live model responses, packet-level New Radio simulation, mobility, shared edge queues, and an image-reading service.Source: D. Scenario III: Intent-Driven… · p. 7
• The edge results illustrate combining bounded intent interpretation with numerical scheduling, while showing why application utility must be assessed alongside decision latency.These results are independently reported from [29], not part of the paper’s simulations.Source: D. Scenario III: Intent-Driven… · p. 7
• In cloud orchestration, Jev’s median decision latency is 374.6 ms versus 482.9 ms for DeepSeek and 984.5 ms for Gemini, and Jev is faster in the other two settings.Fig. 6 reports medians with 25th–75th percentile whiskers; each experiment uses separate measurements.Source: D. Scenario III: Intent-Driven… · p. 7, Figure 6 · p. 8
• Cloud completion is 97.00% with Jev versus 93.51% with DeepSeek and 88.65% with Gemini; local-model completion is 96.62% with Jev versus 91.84% with Qwen.For image service, Jev completes 459 requests correctly and on time, compared with 463 for DeepSeek and 435 for Qwen; workloads must be compared within each panel.Source: D. Scenario III: Intent-Driven… · p. 7, Figure 7 · p. 8

A. Wireless-Native Decision Foundation Models

• A wireless-native System-One foundation model would be pretrained across many bounded control problems, aiming to reuse wireless structure across PHY, RRM, mobility, and management tasks.The proposed goal is adaptation to a new query or answer schema with little task-specific data.Source: A. Wireless-Native Decision Foundation… · p. 8

V. OPEN CHALLENGES AND RESEARCH DIRECTIONS

• Incompatible modalities and timescales, inconsistent action semantics, scarce failure examples, negative transfer, and missing representative decision–outcome datasets are major obstacles to that reuse.The paper frames these capabilities as not yet established for realistic wireless deployments.Source: A. Wireless-Native Decision Foundation… · p. 8, §V · p. 8

B. Generalization and Calibrated Uncertainty

• Wireless distributions vary with geography, frequency, hardware, mobility, traffic, and network configuration, so random train–test splits do not establish deployment robustness.The paper calls for held-out cells, environments, devices, and decision schemas, plus out-of-distribution detection and risk–coverage analysis.Source: B. Generalization and Calibrated… · p. 8

C. Constraint Satisfaction and Closed-Loop Safety

• A normalized decision distribution does not guarantee feasibility, stability, or acceptable long-term behavior; deterministic checks or safety shields should enforce hard constraints.Hysteresis, state history, optimizers, or digital twins are proposed for temporal effects, but the resulting hybrid system must be verified as a whole.Source: C. Constraint Satisfaction and… · pp. 8–9

D. Efficient Deployment and Reproducible Evaluation

• Practical performance includes end-to-end latency and energy across measurement, state construction, inference, validation, signaling, and actuation.The paper calls for matched-compute and information-budget benchmarks reporting tail latency, calibration, constraint violations, robustness, and escalation costs.Source: D. Efficient Deployment and… · p. 9

VI. CONCLUSION

• In receive-antenna selection Jev is faster but does not match stronger numerical and LLM-based baselines in decision quality; in intent-conditioned RAN slicing its quality is comparable to evaluated LLMs with more than threefold lower latency.The paper concludes that value depends strongly on the task.Source: §VI · p. 9
• The paper positions System-One models as a complementary inference layer for repeated bounded decisions, with uncertain, constrained, or computationally demanding cases handled by more capable reasoning models.Realizing this hybrid vision requires wireless-native training, reliable calibration under distribution shift, explicit constraint enforcement, and reproducible end-to-end evaluation.Source: §VI · p. 9