PaperlayerView exact revision on arXiv

Source-linked AI summary

WARA: Toward Automated Wireless Optimization Research with Closed-Loop LLM Agents

Yuan Guo, Yilong Chen, Chao Hu, Xianghao Yu, Liang Hong, Jie Xu

arXiv:2608.14573v1cs.NIcs.AI

TL;DR

Automating end-to-end wireless optimization research requires coordinating models, algorithms, experiments, and claims rather than generating artifacts in isolation. WARA uses closed-loop artifact control and substantially outperforms one-shot LLM generation in manuscript validity.

  • Problem

    End-to-end automation of structured wireless resource-allocation optimization research remains insufficiently addressed despite its highly structured workflow.

  • Method

    WARA is a closed-loop multi-agent framework that constructs and validates research artifacts through staged modeling, experimentation, and deliverable generation.

  • Results

    68.5 versus 37.4: WARA improves average manuscript-level research-validity scores by 31.1 points over topic-paired one-shot LLM generation.

  • Takeaways & Limitations

    The results indicate that phase-structured artifact construction and executable feedback improve reliability in LLM-assisted wireless research.

  • Takeaways & Limitations

    Automatic checks cannot resolve contribution strength, baseline adequacy, or evidence sufficiency.

Abstract

from arXiv · show

Large language model (LLM) agents are increasingly capable of tool use, code execution, artifact inspection, and iterative revision, creating new opportunities for automating scientific and engineering research. To the best of our knowledge, this paper presents the first end-to-end autoresearch framework for the wireless domain, with a focus on wireless resource allocation optimization. We propose the Wireless AutoResearch Agent (WARA), a closed-loop multi-agent system for automated wireless optimization research. Given only an initial topic, WARA decomposes the workflow into three phases: research gap identification and problem proposal, wireless optimization modeling, algorithm design and experimentation, and research deliverable construction. Across these phases, WARA uses artifact-mediated control: upstream artifacts are consumed as inputs, structured outputs are stored for downstream use, and controller-managed gates validate consistency among models, algorithms, experiments, and claims. When validation fails, WARA repairs only the responsible artifact instead of restarting the whole workflow. We present a representative wireless resource allocation case study showing how WARA converts an initial topic into a complete research package with executable evidence and a synthesized technical manuscript. We further design a structured LLM-based ScoringAgent to evaluate manuscript-level research validity and optimization research maturity. Comparative results show that WARA substantially outperforms one-shot LLM generation and approaches the quality profile of recently accepted peer-reviewed technical papers. These results indicate that closed-loop artifact control is a promising path toward end-to-end LLM-assisted wireless optimization research. The source code is available at https://github.com/guoyuan-dotcom/WARA_CUHKSZ.

I. INTRODUCTION

WARA addresses the need for end-to-end closed-loop research automation in wireless resource allocation and optimization. It connects research problem identification, modeling, algorithm design, experimentation, and manuscript construction through artifact-mediated validation and targeted repair.

  • Motivation: LLM agents increasingly combine language reasoning with external tools, executable environments, reflection, and multi-agent coordination for scientific and engineering research.
  • Research gap: Recent autoresearch connects previously separate research phases into closed-loop workflows in which LLM agents perform algorithm design and experimentation through code generation and tool use.
  • Wireless motivation: Wireless resource allocation is central to characterizing performance limits and improving practical system performance under dynamic channel and network conditions.
  • Wireless motivation: Adaptive resource allocation is essential in highly dynamic wireless systems for improving communication efficiency, mitigating interference, and coordinating communication, sensing, and computing resources across space and time.
  • Research gap: Existing wireless LLM research includes optimization and network-control applications alongside wireless-domain agent frameworks and benchmarks, but does not provide the proposed end-to-end wireless autoresearch framework.
  • WARA framework: WARA uses upstream artifacts as inputs, structured downstream outputs, controller-managed consistency gates, and artifact-level repair when validation fails.
  • Evaluation: A representative case study produces executable experimental evidence and a synthesized manuscript, while a structured ScoringAgent evaluates manuscript-level research validity and optimization research maturity.

II. WARA FRAMEWORK · A. Overall Workflow

WARA is a closed-loop framework that constructs a complete wireless optimization research package from an initial topic through three coordinated phases. Its workflow uses persistent artifacts, role-specialized agents, gates, frozen contracts and handoffs, and controller-managed revision.

  • II. WARA FRAMEWORK: WARA automatically constructs a complete research package from an initial wireless research topic under a finite computational budget.
  • A. Overall Workflow: The workflow comprises three sequential phases: research gap identification and problem proposal; wireless optimization modeling, algorithm design, and experimentation; and research deliverable generation.
  • A. Overall Workflow: Phase one transforms a broad wireless topic into a concrete and technically grounded research problem.
  • A. Overall Workflow: Phase two produces validated numerical results and paper-ready figures through iterative execution, quick validation, and result verification.
  • A. Overall Workflow: Artifacts are persistent workspace files that store research objects generated or used by agents, execution tools, or validation steps for later checking and reuse.
  • A. Overall Workflow: After a gate passes, a contract freezes technical decisions such as assumptions, notation, variables, objectives, constraints, solution routes, solver requirements, experiment scope, and claim boundaries.
  • A. Overall Workflow: The controller coordinates agents, records outputs and dependencies, applies gates, freezes accepted contracts and handoffs, and routes failed outputs for revision.

B. Phase 1: Research Gap Identification and Problem Proposal … C. Phase 2: Wireless Optimization Modeling, Solving, and Experimentation

Phase 1 narrows an initial wireless topic into a specific, evidence-grounded optimization problem and freezes it as a handoff artifact. Phase 2 then transforms that problem into a system model, optimization formulation, solution route, executable code, and verified numerical evidence.

  • B. Phase 1: Research Gap Identification and Problem Proposal: Phase 1 jointly grounds the communication scenario, controllable resources, performance objectives, and supporting literature to define a wireless optimization problem.The phase progressively generates, evaluates, selects, and formalizes the research problem.
  • 1) Subphase 1.1: Identification of Wireless System, Variables, and Metrics:: Subphase 1.1 validates the topic against a wireless-domain taxonomy and identifies the scenario, tradeoffs, variables, objectives, constraints, and evaluation metrics.Its structured research frame filters unrelated systems, objectives, variables, and metrics.
  • 2) Subphase 1.2: Literature Search:: Subphase 1.2 uses the research frame to construct literature queries and retrieves evidence from multiple academic sources, expanding the search when coverage is insufficient.Sources include Crossref, Semantic Scholar, OpenAlex, arXiv, IEEE Xplore, and Google Scholar, alongside a local repository of seminal works.
  • 3) Subphase 1.3: Generation and Selection of Research Problems:: Subphase 1.3 generates multiple candidate wireless optimization problems and recommends one for controller evaluation against completeness, taxonomy consistency, coherence, and downstream suitability.The selected problem must be specific, optimization-oriented, technically grounded, and actionable; alternatives and selection rationale are retained for traceability.
  • 3) Subphase 1.3: Generation and Selection of Research Problems:: The wireless-domain taxonomy covers tasks, network settings, radio technologies, optimization variables, modeling assumptions, solver families, and performance metrics.Examples include ISAC, SWIPT, cell-free networks, RIS, beamforming, power allocation, imperfect CSI, SCA, SDR, sum rate, harvested power, and the Cramér–Rao bound.
  • 4) Subphase 1.4:Freezing and Exporting the Selected Problem:: Subphase 1.4 verifies topic-specific grounding, expands literature retrieval when necessary, and assembles a structured problem contract from the selected problem and supporting records.The frozen contract serves as the formal handoff artifact for Phase 2.
  • C. Phase 2: Wireless Optimization Modeling, Solving, and Experimentation: Phase 2 converts the frozen problem into a wireless system model and optimization formulation, derives a consistent solution route, implements executable experiment code, and generates verified numerical evidence.The process runs from model construction through evidence generation.

1) Subphase 2.1: Construction of System Model and Optimization Problem: … 5) Subphase 2.5: Result Verification and Promotion to Paper-Level Evidence:

Phase 2 converts a frozen wireless optimization problem into a validated model, solution route, algorithm, executable experiments, and paper-level evidence. Controller-managed checks enforce consistency, physical validity, implementation fidelity, and publication-ready results throughout this pipeline.

  • 1) Subphase 2.1: Construction of System Model and Optimization Problem:: FormulationAgent generates the wireless system model and optimization problem, including signal and channel models, variables, objectives, constraints, and performance metrics.The controller checks formulation completeness, physical meaningfulness, and consistency with the selected research problem.
  • 2) Subphase 2.2: Tractability Analysis and Solution-Route Selection:: TheoryAgent analyzes convexity, conic representability, nonconvex coupling, discrete variables, rank constraints, fractional structures, and uncertainty to select a solution route.Directly tractable problems use convex or conic solvers; others may use SDR, SCA, bisection, alternating optimization, robust approximation, mixed-integer optimization, or heuristics.
  • 3) Subphase 2.3: Algorithm Specification and Validation Principle:: TheoryAgent produces an implementable algorithm specification with variable updates, feasibility checks, stopping criteria, and baseline-comparison methodology.The controller checks consistency with the frozen optimization formulation and whether convergence, complexity, and optimality claims are supported.
  • 4) Subphase 2.4: Experiment Design and Code Execution:: ExperimentAgent designs comparisons, parameter sweeps, performance metrics, and figure objectives from the mathematical contract, algorithm contract, and claim map.The controller verifies that metrics, sweep variables, and benchmark methods match the frozen optimization problem.
  • 4) Subphase 2.4: Experiment Design and Code Execution:: A controller-owned Python validation harness checks interfaces, executes generated code, collects logs, generates result tables, verifies outputs, and tests fidelity to the algorithm specification.Solver-based implementations must invoke the designated solver rather than substituting a proxy implementation.
  • 5) Subphase 2.5: Result Verification and Promotion to Paper-Level Evidence:: ValidationAgent promotes executable experiments to verified numerical evidence by checking data provenance, benchmark consistency, metric validity, and figure readiness.Plotted curves must originate from saved result tables, compared methods must match the approved design, and metrics must be valid wireless-performance measures.
  • 5) Subphase 2.5: Result Verification and Promotion to Paper-Level Evidence:: A three-level expansion strategy progresses from scout sweeps to medium-scale stability checks and paper-scale sweeps for publication-quality figures.The levels increase parameter values, random seeds, grid density, and experimental scale as confidence in the intended trend grows.

D. Phase 3: Research Deliverable Generation … 4) Subphase 3.4: onstruction of the Full Manuscript Draft:

Phase 3 converts validated Phase 2 outputs into a complete research package through contract-controlled drafting, evidence-grounded results writing, technical assembly, and reference-validated manuscript construction. Controllers compile previews and verify consistency, completeness, traceability, and citation correctness before final deliverables are exported.

  • D. Phase 3: Research Deliverable Generation: Phase 3 transforms validated Phase 2 outputs into a complete research package, revises validated drafts, and exports manuscripts, figures, bibliographies, review reports, and repair records.The workflow is organized into Subphases 3.1–3.6.
  • 1) Subphase 3.1: Drafting Problem Formulation and Method Sections:: WritingAgent generates LaTeX formulation and method sections from the frozen mathematical and algorithm contracts, while the controller verifies notation, equations, constraints, and algorithm descriptions.A technical preview is compiled before verification.
  • 2) Subphase 3.2: Drafting Numerical Results and Supported Claims:: AnalysisAgent generates numerical-results text from verified figures and experiment records, mapping each numerical claim to supporting figures through a claim–figure traceability record.The controller checks that interpretations are evidence-grounded and remain within scope.
  • 3) Subphase 3.3: Assembly of Technical Sections:: WritingAgent assembles the abstract, keywords, conclusion, and core technical manuscript from validated sections while preserving consistency with the approved model, algorithm, evidence, and claims.The controller checks completeness and the high-level narrative.
  • 3) Subphase 3.3: Assembly of Technical Sections:: The controller’s repair routing supports targeted correction of manuscript-generation artifacts rather than restarting the full workflow.Table II is identified as covering controller repair routing in WARA.
  • 4) Subphase 3.4: onstruction of the Full Manuscript Draft:: LiteratureAgent validates Bib-TeX entries from the reference repository, after which WritingAgent writes the introduction and citations; the controller validates keys, venues, placement, and citation–claim consistency.IEEE and peer-reviewed sources are prioritized.

5) Subphase 3.5: Full Manuscript Review: … III. CASE STUDY

WARA reviews the complete manuscript against frozen research artifacts, repairs validated issues, and exports a final research package after passing a quality gate. Its gate–freezing–repair control layer preserves consistency across the workflow, while the case study specifies operational and experimental budgets.

  • 5) Subphase 3.5: Full Manuscript Review:: ReviewAgent evaluates manuscript consistency across equations, algorithms, figures, citations, abbreviations, and technical claims, assigning blocking, major, and minor issues to owners.The structured review report guides controller decisions about whether issues require repair.
  • 6) Subphase 3.6: Revision and Export of the Final Research Package:: After repairs, the final quality gate validates compilation, citation integrity, abbreviation consistency, figure quality, and claim support before WARA exports the research package.The package includes manuscript source, compiled PDF preview, bibliography, figures, review reports, repair history, and final validation records.
  • E. Gate–Freezing–Repair Mechanism: The gate–freezing–repair mechanism connects all phases by validating artifacts before reuse, freezing approved decisions, and routing failures to responsible agents.This control layer maintains consistency across formulation, algorithm design, experimentation, and manuscript generation.
  • 1) Gate:: Each gate evaluates candidate artifacts against active contracts, emits a machine-readable report, and blocks failed artifacts from downstream reuse.Passed artifacts are committed to the workspace, while reports include pass/fail status, detected issues, and repair instructions.
  • 2) Freezing:: WARA freezes four gate-approved artifacts—the problem handoff, mathematical contract, algorithm contract, and verified result package—for controlled downstream consumption.These artifacts define the selected problem, formulation, solution route, and numerical evidence; later agents cannot modify them without triggering the prescribed control process.
  • 3) Repair:: Repair targets only the failed artifact, routing formulation, algorithmic, experimental, evidence, and manuscript issues to their responsible agents while recording validation outcomes.Each cycle records the failed gate, artifact owner, repair instructions, revised artifact, and validation outcome.
  • III. CASE STUDY: The representative case study runs from an initial wireless topic to a complete research package under defined retrieval, repair, token, and experiment budgets.Two promoted sweeps use 12 and 9 x-axis points with 50 Monte Carlo seeds per point, producing 1050 paired simulation settings and 2100 method evaluations.

A. Research Gap Identification and Problem Proposal

Phase 1 selects a fronthaul-aware max-min power-control and user-centric-clustering direction for downlink cell-free massive MIMO. It frames joint AP–user association and power loading around the tension between useful signal strength, transmit power, and finite fronthaul capacity.

  • A. Research Gap Identification and Problem Proposal: WARA selects fronthaul-aware max-min power control and user-centric clustering for downlink cell-free massive MIMO.
  • A. Research Gap Identification and Problem Proposal: The research object is a downlink network with distributed APs, single-antenna users, finite AP transmit-power budgets, and finite per-AP fronthaul capacities.
  • A. Research Gap Identification and Problem Proposal: Activating an AP–user serving link can improve useful signal strength while consuming AP transmit power and fronthaul capacity.
  • A. Research Gap Identification and Problem Proposal: The selected mechanism jointly optimizes AP–user association and power loading to allocate fronthaul toward serving links with high fairness value.

B. Wireless Optimization Modeling, Algorithm Design, and Experimentation

WARA models downlink cell-free massive MIMO as a fronthaul-aware max-min admitted-rate problem, reformulates each trial target as a mixed-integer second-order-cone feasibility test, and solves it by scalar bisection. Experiments compare joint AP–user association and power loading against RZF-fixed across fronthaul capacity and user load.

  • Modeling: WARA formulates downlink cell-free massive MIMO as a fronthaul-aware max-min admitted spectral-efficiency problem with AP–user association, power loading, fronthaul, and SINR-dependent rate constraints.The binary variable x_mk denotes AP–user association, p_mk downlink power, r_k admitted spectral efficiency, and t the common fairness level.
  • Reformulation: The max-min formulation is tested at a trial target τ by setting r_k = t = τ, defining θ(τ) = 2τ − 1, and recovering physical powers as p_mk = u_mk^2.The target test uses solver-side amplitudes u_mk ≥ 0.
  • Reformulation: Each target test becomes a mixed-integer second-order-cone feasibility problem with association, service, fronthaul, AP-power, activation, and SINR constraints.The mixed-integer and nonconvex structure motivates the feasibility-based reformulation.
  • Experimentation: The proposed joint association and power-loading method achieves higher minimum admitted spectral efficiency t than the RZF-fixed benchmark across fronthaul-capacity and user-load sweeps.The advantage is attributed to jointly optimizing AP–user association and power loading under finite fronthaul and AP-power constraints while maintaining a higher fairness level as user load increases.

C. Phase 3: Research Deliverable Generation

Phase 3 converts validated Phase-2 outputs into a final research paper package. In the case study, WARA generates a four-page manuscript alongside verified supporting artifacts and quality-control reports.

  • Research deliverable generation: Phase 3 uses the selected problem direction, mathematical formulation, solution route, promoted figures, and verified references to generate the final paper package.The package includes the manuscript and associated validation artifacts.
  • Research deliverable generation: The final package contains a four-page manuscript, 13 references, two paper-ready figures, a post-revision quality-gate report, a contract-scope report, and an abbreviation check.The generated manuscript is shown in Fig. 2.

IV. COMPARATIVE EVALUATION RESULTS · A. ScoringAgent and Evaluation Criteria

WARA-generated manuscripts are evaluated with a structured LLM-based ScoringAgent for manuscript-level research validity and optimization research maturity. The agent applies fixed reviewer-style criteria, produces evidence-justified machine-readable reports, and aggregates dimension-level scores.

  • IV. COMPARATIVE EVALUATION RESULTS: The evaluation measures overall manuscript validity across problem definition, novelty, technical correctness, alignment, evidence, claims, writing, and references.These dimensions assess whether the manuscript is coherent, technically sound, evidence-supported, clearly communicated, and properly grounded.
  • IV. COMPARATIVE EVALUATION RESULTS: The evaluation separately examines the manuscript’s optimization research maturity because wireless optimization is a central focus.This optimization-focused assessment complements the broader manuscript-validity evaluation.
  • A. ScoringAgent and Evaluation Criteria: ScoringAgent receives each manuscript PDF and prompts an evaluator LLM to act as a strict wireless-communications reviewer.The fixed prompt requires predefined criteria, manuscript-visible evidence for each score, and a machine-readable JSON report.
  • A. ScoringAgent and Evaluation Criteria: Manuscript-level validity is assessed using eight research-validity dimensions and combined into an aggregate research-validity score.The dimensions cover problem definition, contribution positioning, system-model correctness, method–formulation alignment, evidence, claims, communication, and reference grounding.
  • A. ScoringAgent and Evaluation Criteria: For manuscript Pi, ScoringAgent assigns dimension-level scores that are used to compute the manuscript-level research-validity score.The scoring procedure operates at the individual-manuscript level before benchmark averages are reported.
  • A. ScoringAgent and Evaluation Criteria: Optimization maturity is represented by a six-dimensional normalized profile ui = [ui,1, ui,2, . . . , ui,6], with 0 ≤ ui,d ≤ 100.The overall optimization research-maturity score is computed from the normalized dimension scores.
  • A. ScoringAgent and Evaluation Criteria: For a benchmark set B containing N manuscripts, the evaluation reports average manuscript-level and optimization research-maturity scores.These averages summarize performance across the benchmark rather than a single manuscript.

B. Benchmark Sets and Comparative Results

WARA is evaluated against topic-paired one-shot GPT-5.5 manuscripts and accepted WCL papers using manuscript-validity and optimization-maturity benchmarks. It substantially outperforms one-shot generation through staged artifact control, executable validation, and iterative revision, while remaining weaker than accepted papers in several maturity dimensions.

  • Benchmark sets: The evaluation compares ten WARA manuscripts, ten topic-paired one-shot GPT-5.5 manuscripts, and ten recently accepted optimization-related WCL papers.The scoring agent uses Kimi K2.6 to reduce evaluator–generator coupling with the GPT-5.5 manuscript-generation backbone.
  • Manuscript-level validity: 68.5 average validity score for WARA versus 37.4 for the one-shot baseline, a 31.1-point improvement under the same topics and backbone.The largest gap is in evidence validity: the one-shot baseline receives 0.0, whereas WARA reaches 13.0 through executable validation and evidence recording.
  • Comparison with peer-reviewed papers: Compared with accepted WCL papers, WARA is closer in problem definition and novelty but weaker in evidence validity, claim support, and method maturity.Staged formulation and algorithm construction reduce mismatch between the stated optimization problem and the proposed solution.
  • Topic-level comparison: WARA outperforms the corresponding one-shot manuscript on all ten topics, indicating that its aggregate gain is not driven by only a few favorable examples.Each topic is repeatedly interpreted, checked, and revised through multiple agent stages rather than generated in a single forward pass.
  • Optimization research maturity: 69.7 average optimization research-maturity score for WARA versus 44.6 for the one-shot baseline.WARA fixes variables, objectives, constraints, and assumptions as a mathematical contract before deriving the solution route and experiment design.

V. CONCLUSION

WARA frames wireless optimization research as a chain of verifiable artifacts managed by closed-loop controls, targeted repair, and consistency gates. Its evaluation shows substantial gains over one-shot LLM generation, while experimental depth and evidence maturity remain the main gaps to accepted wireless papers.

  • Framework: WARA transforms an initial wireless topic into a research direction, optimization study, executable evidence, and final research deliverables.The framework treats the study as a chain of verifiable artifacts rather than a single generated manuscript.
  • Framework: Controller-managed gates, frozen interfaces, and targeted repair maintain consistency as the research package evolves.Repairs focus on the responsible artifact instead of restarting the complete workflow.
  • Case study: The case study produces a bounded direction, frozen mathematical contract, implementable solution route, and verified numerical evidence from a broad wireless topic.These components form a concrete and executable optimization study.
  • Evaluation: WARA improves substantially over one-shot LLM generation, especially in formulation–method alignment, evidence validity, claim support, and optimization research maturity.The comparative evaluation identifies these dimensions as the most visible gains.
  • Limitations and future direction: The remaining gap to accepted peer-reviewed wireless papers lies mainly in experimental depth and evidence maturity, requiring stronger experimental judgment.Human expertise could guide contribution strength, baseline adequacy, and evidence sufficiency when automatic checks cannot resolve them.
Loading 2608.14573v1…