PaperlayerView exact revision on arXiv

Source-linked AI summary

6GAgentGym: Tool Use, Data Synthesis, and Agentic Learning for Network Management

Jiao Chen, Jianhua Tang, Xiaotong Yang, Zuohong Lv

arXiv:2603.29656v1cs.NIcs.AI

TL;DR

Existing benchmarks based on static questions or scripted episode replay do not support agents learning through closed-loop environmental feedback. 6GAgentGym combines a 42-tool interactive environment and learned Experiment Model with 6G-Forge, which synthesizes training trajectories from NS-3 seed data. The 8B model achieves comparable overall performance to GPT-5 on 6GAgentBench and exceeds it on long-horizon L3 tasks.

  • Problem

    Existing benchmarks based on static questions or scripted episode replay do not support agents learning through closed-loop environmental feedback.

    Source: §I · p. 1

  • Method

    6GAgentGym combines a 42-tool interactive environment and learned Experiment Model with 6G-Forge, which synthesizes training trajectories from NS-3 seed data.

    Source: §I · p. 2

  • Results

    The 8B model achieves comparable overall performance to GPT-5 on 6GAgentBench and exceeds it on long-horizon L3 tasks.

    Source: §I · p. 2, D. Result Analysis · p. 6

  • Takeaways & Limitations

    The framework provides a viable path toward autonomous, closed-loop network management.

    Source: §V · p. 7

    The Experiment Model omits full protocol-level transients, especially during handover and failure recovery, and the 42-tool set excludes radio-level operations such as beamforming and power control.

    Source: §V · p. 7

Abstract

from arXiv · show

Autonomous 6G network management requires agents that can execute tools, observe the resulting state changes, and adapt their decisions accordingly. Existing benchmarks based on static questions or scripted episode replay, however, do not support such closed-loop interaction, limiting agents to passive evaluation without the ability to learn from environmental feedback. This paper presents 6GAgentGym to provide closed-loop capability. The framework provides an interactive environment with 42 typed tools whose effect classification distinguishes read-only observation from state-mutating configuration, backed by a learned Experiment Model calibrated on NS-3 simulation data. 6G-Forge bootstraps closed-loop training trajectories from NS-3 seeds via iterative Self-Instruct generation with execution verification against the Experiment Model. Supervised fine-tuning on the resulting corpus followed by reinforcement learning with online closed-loop interaction enables an 8B open-source model to achieve comparable overall success rate to GPT-5 on the accompanying 6GAgentBench, with stronger performance on long-horizon tasks. Together, these components provide a viable path toward autonomous, closed-loop network management.

I. INTRODUCTION

6G network management spans complex, interdependent tasks that remain difficult to automate end to end, while existing approaches and benchmarks leave a gap in interactive, feedback-driven management. 6GAgentGym combines a closed-loop environment, synthetic training data, and agent training to address that gap.

I. INTRODUCTION

  • Manual optimization grows harder as technologies, band combinations, and network slices multiply, and configuration and change-management failures account for 45% of outages.The paper describes management spanning intent translation, slicing, trust policies, and multi-agent coordination.Source: §I · p. 1
  • Existing 6G benchmarks rely on static questions or scripted episodes, while prior tool-use synthesis lacks environmental feedback and network RL lacks semantic reasoning.The paper identifies an integrated environment where agents execute state-mutating tools, observe consequences, and learn from resulting trajectories as underexplored.Source: §I · p. 2
  • 6GAgentGym provides 42 effect-typed tools and an Experiment Model calibrated on NS-3 data to predict six-dimensional network state transitions.The tool effects distinguish Observation, Reasoning, and Configuration operations.Source: §I · p. 2
  • 6G-Forge bootstraps training trajectories from NS-3 seeds through iterative Self-Instruct generation, execution verification, and diversity filtering.Generation is constrained by typed tool signatures and effect labels, with verification against the Experiment Model.Source: §I · p. 2
  • Agentic SFT followed by online closed-loop RL enables an 8B model to match GPT-5 overall on 6GAgentBench and outperform it on long-horizon L3 tasks.The comparison is reported for the benchmark’s L1–L3 tiers.Source: §I · p. 2

A. LLM Agents for 6G Networks

Prior work advances LLM-based 6G agents, interactive evaluation, and network RL, but the approaches remain divided across static evaluation, scripted episodes, open-loop tools, or low-dimensional actions. The methodology overview presents an integrated framework for interactive network management.

A. LLM Agents for 6G Networks

  • Existing 6G agent studies include static QA, conversational control episodes, and closed-loop digital-twin or intent-translation work, but leave state-mutating tool execution with adaptive feedback unaddressed.6G-Bench evaluates multiple-choice questions; α3-Bench and α3-SecBench use episodes, while the cited digital-twin and intent-translation work addresses optimization and policy translation.Source: A. LLM Agents for 6G Networks · p. 2

B. Tool-Use Benchmarks and Data Synthesis

  • Interactive benchmarks ground evaluation in execution, and synthesis pipelines verify tool-use data, but existing tools remain open-loop rather than changing environmental state.The cited benchmark examples target web and software tasks, not network operations.Source: B. Tool-Use Benchmarks and Data… · p. 2

C. Reinforcement Learning for Network Management

  • Network RL optimizes fixed, low-dimensional actions, whereas LLM agents can interpret operator intent and compose multi-step diagnostic workflows.The paper presents these approaches as complementary paradigms at different levels.Source: C. Reinforcement Learning for Network… · pp. 2–3

III. METHODOLOGY

  • The framework overview combines an interactive environment, 6G-Forge trajectory synthesis, tiered L1–L3 evaluation, and an Agentic SFT plus RL training pipeline.The overview places the environment, synthesis, evaluation, and training components in one framework.Source: Figure 2 · p. 3, §III · p. 3

A. Agent Interaction Model

The agent repeatedly acts on current network state and observes updated results until intent verification or a step limit. The tool taxonomy distinguishes operations that mutate state, and NS-3 supplies simulation traces and ground truth.

A. Agent Interaction Model

  • At each step, the agent uses the current state and natural-language intent to select a typed tool, observes the returned result and updated state, then continues until verification or the step budget ends.The interaction is distinguished from static QA, scripted replay, and low-dimensional RL actions.Source: A. Agent Interaction Model · p. 3

1) Typed Tool System:

  • The 42 tools divide into read-only Observation, state-independent Reasoning, and state-mutating Configuration operations; configuration can invalidate prior observations and require re-observation.The tool set spans network management and UAV control.Source: 1) Typed Tool System · p. 3

2) NS-3 Simulation Backend:

  • NS-3 provides protocol-level simulation traces for six-dimensional state dynamics, seeds the Experiment Model and 6G-Forge, and validates final policies.The framework uses its NR module and FlowMonitor for per-flow metrics.Source: 2) NS-3 Simulation Backend · pp. 3–4, Figure 3 · p. 4

3) Experiment Model:

The Experiment Model is a compact LLM proxy for costly NS-3 interactions, trained on simulation-derived transitions. It predicts tool results and next states while preserving state for observation and reasoning calls and modeling configuration changes.

3) Experiment Model:

  • The Experiment Model approximates NS-3 dynamics as a compact learned proxy for repeated synthesis and RL interactions, rather than reproducing the simulator exactly.It operates over six network dimensions: slice type, latency, jitter, loss, throughput, and edge-compute load.Source: 3) Experiment Model · p. 4
  • At each step, the model receives current state, tool call and arguments, optional degradation, and interaction history, then produces reasoning, a predicted next state, and a tool result.Observation and Reasoning calls preserve state absent degradation; Configuration calls predict transition dynamics.Source: 3) Experiment Model · p. 4
  • The model is fine-tuned on NS-3 transition tuples with a joint objective for next-state prediction and tool-result generation, encouraging consistent transitions and faithful returns.The objective combines conditional log probabilities for the next state and tool result.Source: 3) Experiment Model · p. 4

C. 6G-Forge: Data Synthesis

6G-Forge addresses the cost and expertise barriers to trajectory collection by bootstrapping a corpus from NS-3 seeds through iterative LLM generation and execution verification. Seed annotation turns simulation traces into tool-use trajectories; Self-Instruct expansion generates additional tasks from those examples.

C. 6G-Forge: Data Synthesis

  • 6G-Forge bootstraps tool-use trajectories from a small set of NS-3 seeds to address costly simulations, expert-intensive authoring, and limited static-data diversity.It uses iterative LLM-driven generation with execution verification against Mθ.Source: C. 6G-Forge: Data Synthesis · p. 4

1) Step 1: Seed Annotation:

  • A teacher LLM identifies decision points in NS-3 traces and creates task descriptions and corresponding tool-call sequences.The seed pool is grounded in physical-layer dynamics and anchors subsequent generation.Source: 1) Step 1: Seed Annotation · p. 4

2) Step 2: Self-Instruct Expansion:

  • At each iteration, the teacher uses sampled seed trajectories and typed tool signatures to generate a new task and tool-call trajectory, sometimes adding a degradation scenario requiring recovery.The degradation scenario is specified with probability pdeg.Source: 2) Step 2: Self-Instruct Expansion · p. 4

3) Step 3: Execution Verification:

Candidate trajectories are checked through execution against the Experiment Model, while deduplication and iterative growth expand the verified corpus. The accompanying benchmark organizes tasks into three difficulty tiers and excludes tasks with high zero-shot success.

3) Step 3: Execution Verification:

  • Candidate trajectories execute step-by-step against Mθ; successful runs become golden traces, while errors prompt correction of the failed step.Both golden and error-recovery trajectories are retained as training data.Source: 3) Step 3: Execution Verification · p. 4

4) Step 4: Deduplication and Iterative Growth:

  • Verified trajectories are compared with the existing pool using ROUGE-L, near-duplicates are discarded, and the remaining trajectories are added across K iterations.The iterative process produces the synthetic dataset Dsyn.Source: 4) Step 4: Deduplication and Iterative… · p. 5

D. 6GAgentBench: Tiered Evaluation

  • 6GAgentBench defines L1 as at most 3 steps, L2 as 4–7 steps, and L3 as at least 8 steps, excluding tasks where any model exceeds 80% zero-shot success.The tiers cover elementary action chains, adaptive network workflows, and long-horizon multi-agent workflows under degradation, respectively.Source: D. 6GAgentBench: Tiered Evaluation · p. 5

E. Two-Stage Agentic Training

Training combines supervised fine-tuning on real and synthetic trajectories with reinforcement learning through online interaction against Mθ. The evaluation also includes frontier and non-LLM models under ReAct-style interaction loops.

1) Stage 1: Agentic SFT:

  • Agentic SFT fine-tunes the student model on the union of real NS-3 trajectories and synthetic trajectories from 6G-Forge.The combined dataset is Dsft = Dsyn ∪ Dreal.Source: 1) Stage 1: Agentic SFT · p. 5

2) Stage 2: Agentic RL:

  • Starting from the SFT checkpoint, agentic RL uses online tool interaction against Mθ and selects tasks where the SFT policy’s success rate is within a learnable range.The composite reward penalizes malformed tool calls and measures task correctness; optimization uses DAPO [31].Source: 2) Stage 2: Agentic RL · p. 5

1) Models:

  • The evaluation covers eight frontier models, three non-LLM baselines, and the fine-tuned 6GAgent-8B and 6GAgent-4B models under deterministic ReAct-style interaction.The evaluated frontier models span proprietary and open-source families.Source: 1) Models · p. 5

2) Training Data:

On 6GAgentBench, 6GAgent-8B (SFT+RL) matches GPT-5 overall and exceeds it on L3, while models generally decline from L1 to L3. The evaluation uses a held-out task protocol, and the training corpus combines NS-3 and synthetic trajectories.

2) Training Data:

  • The SFT corpus contains 3,000 real NS-3 trajectories and 50,000 synthetic trajectories from 15 Self-Instruct iterations, spanning five domains and a 30%/45%/25% L1/L2/L3 distribution.The synthetic set comprises 30,000 golden traces and 20,000 error-recovery augmented traces; approximately 8,000 frontier tasks are selected for RL.Source: 2) Training Data · p. 5

3) Holdout Protocol:

  • The holdout protocol separates evaluation scenarios from training seeds, removes tasks with ROUGE-L ≥0.7 similarity to training instances, and freezes Mθ before RL.The evaluation uses NS-3 seeds 51–80 rather than training seeds 1–50; 127 candidate tasks were removed by deduplication.Source: 3) Holdout Protocol · p. 5

4) Metrics:

  • Success Rate measures correct final answers through valid tool-use chains, while SPL penalizes unnecessarily long trajectories..Source: 4) Metrics · p. 5

B. Main Results

  • The three non-LLM baselines achieve 38–55% L1 performance but degrade sharply on L2 and L3.The baselines are Threshold-Rule, MAPE-K Heuristic, and adapted DRL-Slicing.Source: B. Main Results · p. 5
  • 50.1% overall SR for 6GAgent-8B (SFT+RL) is comparable to GPT-5’s 50.2%, and its L3 SR is 39.1% versus GPT-5’s 33.8%.The 8B model also exceeds Qwen3-VL-72B’s 36.8% overall SR; all models show sharp degradation from L1 to L3. RL adds +4.8% overall on top of SFT, with larger gains on L3 (+4.9%) than L1 (+3.5%).Source: B. Main Results · pp. 5–6, Table I · p. 6

C. Ablation Studies

The training-data ablations attribute gains to closed-loop state mutation, NS-3 real data, and agentic RL; result analysis compares difficulty tiers, RL algorithms, and path efficiency.

1) Training Data Composition:

  • Closed-loop trajectories add +4.8% over open-loop, while NS-3 real data adds +2.2% overall and +2.8% on L3.Closed-loop trajectories capture state mutations from Configuration tools; NS-3 data recovers physical-layer reasoning abstracted away by synthetic generation.Source: 1) Training Data Composition · p. 6
  • Agentic RL adds +4.8% over SFT, with larger gains on L3 (+4.9%) than L1 (+3.5%).Source: 1) Training Data Composition · p. 6

D. Result Analysis

  • 6GAgent-8B (+RL) reaches 39.1% on L3 versus GPT-5’s 33.8%, despite GPT-5 leading L1 at 72.4% versus 68.2%.The tier comparison suggests closed-loop RL is particularly beneficial for long-horizon tasks.Source: D. Result Analysis · p. 6
  • DAPO reaches 42.0% final SR versus GRPO’s 40.8% on the 4B model, converging faster through dynamic sampling of under-explored task types.Source: D. Result Analysis · p. 6
  • 6GAgent-8B (+RL) achieves the highest SPL at 42.8%, slightly above GPT-5’s 42.7%.Source: D. Result Analysis · p. 6

E. NS-3 Grounded Validation

On 120 held-out tasks replayed in NS-3, the Experiment Model closely tracks simulator success rates and preserves the RL advantage, while protocol-level transients remain a fidelity boundary.

E. NS-3 Grounded Validation

  • Validation replays each model’s tool-call trajectory on 120 held-out tasks directly in NS-3, measuring simulator states rather than Experiment Model predictions.Source: E. NS-3 Grounded Validation · p. 6
  • The Experiment Model–to–NS-3 success-rate gap averages 2.2% across models, with SR correlation r = 0.99 (p < 0.001); 6GAgent-8B’s RL advantage persists in NS-3 at 47.5% versus 42.5% (+5.0%).The tier gap grows from below 1% on L1 to around 4% on L3, where transient dynamics are more difficult to approximate.Source: E. NS-3 Grounded Validation · pp. 6–7

V. CONCLUSION

  • 6GAgentGym integrates a 42-tool interactive environment and learned Experiment Model, NS-3-seeded 6G-Forge synthesis, and 6GAgentBench evaluation.Agentic SFT followed by online closed-loop RL enables an 8B model to match GPT-5 overall and outperform it on long-horizon tasks.Source: §V · p. 7
  • The Experiment Model omits full protocol-level transients, especially during handover and failure recovery; the 42-tool set also excludes beamforming and power control.Source: §V · p. 7

APPENDIX A FULL TOOL CATALOG

The appendices describe the typed tool catalog and an interactive dashboard for inspecting NS-3 traces.

APPENDIX A FULL TOOL CATALOG

APPENDIX B INTERACTIVE VISUALIZATION DASHBOARD

  • The Streamlit dashboard reads NS-3 CSV traces and presents network metrics, decision-point types, and per-UAV latency views.Source: Appendix B · p. 8
Loading 2603.29656v1…