PaperlayerView exact revision on arXiv

Source-linked AI summary

X-RIC: An Agentic RAN-Intelligent Controller Benchmark for Enhancing Reinforcement Learning

Adhwaa Al-Chaab, Ayman Younis, Dario Pompili

arXiv:2610.22269v1cs.AIcs.AR

TL;DR

O-RAN AI control faces demands for real-time adaptation and explainable decisions, while current approaches have limitations including poor sample efficiency and limited generalization. X-RIC combines reactive RL, agentic reasoning, and automated behavioral cloning, and introduces BARX for benchmarking agentic RAN control. In its network-slicing case study, X-RIC500 achieves 89.0% cost reduction against RL.

  • Problem

    O-RAN AI control requires real-time responsiveness, adaptation under uncertainty, and explainability, while RL approaches face poor sample efficiency and limited generalization.

    Source: §I · p. 1

  • Method

    X-RIC combines reactive RL with agentic reasoning and automated demonstration harvesting through dual-filtering behavioral cloning, and BARX benchmarks agentic RAN control.

    Source: §I · pp. 1–2

  • Results

    89.0% cost reduction is reported for X-RIC500 against the RL reference in total vRB and vCW allocation cost.

    Source: B. MDP Model · p. 5

  • Takeaways & Limitations

    The paper applies X-RIC to a three-network-slice case study and reports automated demonstration collection for sample-efficient BC training with production-grade reliability.

    Source: §V · p. 9

Abstract

from arXiv · show

The O-RAN ALLIANCE introduces architectural flexibility through disaggregation and AI-driven control. While the RAN Intelligent Controller (RIC) enables centralized optimization, current AI approaches lack reasoning capabilities and explainability. We propose eXplainable RIC (X-RIC), a novel framework integrating agentic AI into O-RAN architectures. X-RIC combines reactive Reinforcement Learning (RL) with strategic agentic reasoning through: (i) automated Behavioral Cloning (BC) using domain heuristics for sample-efficient training, (ii) a Telecom Language Model (TLM), a transformer-based reasoning module providing contextual abstraction and interpretability, and (iii) closed-loop co-adaptation within the Near-RT RIC control loop enabling continuous policy refinement. To ensure deployment readiness, we introduce BARX (Benchmarking Agentic Reasoning for RAN), a domain-specific evaluation framework assessing agentic control across latency, compliance, and determinism metrics. We validate X-RIC on SLA-aware network slicing in simulation and O-RAN-compliant testbed deployments. X-RIC achieves 89% operational cost reduction while maintaining SLA compliance, converging 75% faster with 50% fewer training steps and 46% lower variance.

I. INTRODUCTION

O-RAN’s AI-native control vision must meet real-time, adaptation, and explainability demands, while existing RL and language-model approaches retain limitations in sample efficiency, generalization, and closed-loop control. X-RIC combines reactive RL with agentic reasoning and automated behavioral cloning, alongside BARX for evaluating agentic RAN control.

I. INTRODUCTION

  • O-RAN AI control must address strict Near-RT latency, adaptation to dynamic conditions, and explainability, while RL also faces poor sample efficiency and limited generalization.The passages also identify interpretability as important for safety-critical operations.Source: §I · p. 1
  • X-RIC combines fast reactive learning with strategic reasoning to provide contextual abstraction, reflective learning, and decision traceability.Its architecture includes transformer-based natural-language rationales and closed-loop joint optimization.Source: §I · pp. 1–2
  • Existing O-RAN frameworks lack integrated AI reasoning and autonomous adaptation, while current LLM approaches lack continuous adaptation, closed-loop validation, or suitable latency guarantees.The related-work discussion also notes live-network exploration and scalability challenges for RL in dynamic, concurrent multi-slice settings.Source: §I · pp. 1–2
  • 89% cost reduction versus baselines is reported for the network-slicing case study under Near-RT RIC timing constraints.The introduction describes the case study as minimizing operational cost while meeting SLA requirements.Source: §I · p. 2
  • The paper introduces BARX for agentic RAN evaluation and dual-filtering BC with automated demonstration harvesting, using episode-level policy-impact gating and action-level critic validation.The authors characterize automated BC within O-RAN agentic control as a first, to their knowledge.Source: §I · p. 2

II. X-RIC FRAMEWORK DESIGN

X-RIC integrates agentic reasoning into Near-RT RIC operations and introduces BARX to benchmark agentic RAN control. Its TLM combines a frozen LLM with domain prompts and a trainable RAN-state encoder, while BARX evaluates control-relevant performance dimensions.

II. X-RIC FRAMEWORK DESIGN

  • X-RIC integrates agentic reasoning into Near-RT RIC operations and pairs it with dual-filtering demonstration collection for sample-efficient BC training.The framework design also introduces BARX for agentic RAN benchmarking.Source: §II · p. 2

A. TLM Module and Benchmarking

  • The TLM augments a frozen pretrained LLM with O-RAN prompts and a trainable encoder that maps telemetry into 256-dimensional embeddings aligned to the LLM representation space.Contrastive alignment on state-text pairs trains the encoder without full model fine-tuning.Source: A. TLM Module and Benchmarking · p. 2
  • BARX evaluates agentic reasoning across latency, compliance, determinism, and policy impact in closed-loop control settings.It measures response latency, validates replies against a Draft-7 JSON schema, quantifies consistency across valid responses, and assesses parsed control directives.Source: A. TLM Module and Benchmarking · pp. 2–3
  • BARX aggregates (µℓ, σℓ, ℓ95, E, C, and ∆¯R) into a normalized score from 0 to 100, with thresholds calibrated using foundation-model evaluation practices [27].Scores above 80 indicate deployment-grade performance; scores from 60 to 79 indicate marginal reliability requiring monitoring.Source: A. TLM Module and Benchmarking · p. 3

B. X-RIC Agent System Architecture

X-RIC pairs a critic-guided low-latency RL controller with a reasoning agent, then uses filtered demonstrations for policy updates. Its automated harvesting selects promising episodes and prunes actions the critic does not prefer.

B. X-RIC Agent System Architecture

  • A TLM monitoring agent recommends allocations every 10 seconds, while a BC-trained actor-critic remains the primary controller with latency below 10 ms.A critic selects the TLM action only when its evaluated reward exceeds that of the RL proposal; otherwise, the RL action proceeds.Source: B. X-RIC Agent System Architecture · p. 3

C. Automated Demonstration Collection for BC

  • Dual filtering automates demonstration harvesting by selecting episodes with positive policy impact, then retaining only state–action pairs the critic prefers to the current policy.The resulting filtered buffer supplies demonstrations for behavioral cloning updates.Source: C. Automated Demonstration Collection… · pp. 3–4
  • Episode-level filtering compares mean TLM and RL rewards to identify promising trajectories, smoothing noisy per-step rewards while avoiding comparisons across every step.Episodes pass when their reward difference exceeds a nonnegative margin.Source: C. Automated Demonstration Collection… · p. 4
  • The actor update balances return maximization with selective imitation, and the BC-loss weight is annealed as the policy approaches demonstrator performance.The annealing is intended to prevent the actor from being constrained below expert level.Source: C. Automated Demonstration Collection… · p. 4
  • X-RIC500 achieved 3× faster reward gain and 46% lower variance than RL, plus ~75% faster convergence and ~6× lower deployment cost than RL and TLM-RL.It reached 0.90 explained variance with 50% fewer training steps versus RL-only at ~0.87.Source: C. Automated Demonstration Collection… · p. 4, Figure 2 · p. 4, Figure 3 · p. 4

III. SLA GUARANTEE IN O-RAN USE CASES

The paper applies X-RIC to dynamic network slicing, where the objective is to minimize operating cost while meeting each slice’s SLA requirements.

III. SLA GUARANTEE IN O-RAN USE CASES

  • The case study applies X-RIC to a network-slicing setup,.The setup involves dynamic configuration and poses challenges for traditional RL methods.Source: §III · p. 4
  • The slicing problem’s objective is to minimize the total operating cost of maintaining multiple slices.Source: §III · p. 4
  • The cost objective is subject to each slice meeting its SLA-defined performance requirements.Source: §III · p. 4

A. Problem Description

The problem allocates virtual radio and compute resources across eMBB, URLLC, and mMTC slices to minimize operating cost subject to slice KPI thresholds. The resulting mixed-integer nonlinear problem is addressed with RL for near-real-time operation.

A. Problem Description

  • The network allocates discrete virtual Resource Blocks and continuous virtual Compute Workloads among slices to meet service requirements.The workload fraction represents CPU/GPU time-sharing, while resource blocks are discrete spectrum units.Source: A. Problem Description · p. 5
  • Slice-specific KPIs encode throughput, user-rate satisfaction, and URLLC latency requirements, with nonnegative KPI contributions ensured by max(0, ·).Source: A. Problem Description · p. 5
  • The operating-cost allocation problem is a mixed-integer nonlinear program because vRBs are integer-valued and the rate function is nonlinear; the paper uses RL to solve it near real time.Source: A. Problem Description · p. 5

B. MDP Model

The MDP models network control through states, actions, and a policy that maximizes discounted reward. Its reward balances allocation cost against SLA violations.

B. MDP Model

  • The control problem is formulated as an MDP with states, actions, a policy, and an objective to maximize discounted cumulative reward.The discount factor γ lies in [0, 1).Source: B. MDP Model · p. 5
  • 89.0% cost reduction is reported for X-RIC500 versus RL as the 0% reference; TLM-RL achieves 65.8%.Fig. 4 compares total vRB and vCW allocation cost reduction across approaches.Source: Figure 4 · p. 5, B. MDP Model · p. 5
  • The state includes slice arrival-rate and service-demand characteristics, while actions allocate vRBs and continuous vCW shares.Continuous vRB outputs are rounded to discrete allocations.Source: B. MDP Model · p. 5
  • The reward incentivizes cost minimization while penalizing SLA violations by η > 0.The equation specifies a penalty for positive KPI excess.Source: B. MDP Model · p. 5

C. TLM-based RL Approach Solution

The approach integrates TLM reasoning with DDPG through dual-filtered demonstrations, mixed replay training, and a continuous learning loop. The TLM evaluates proposed actions and can return corrected actions with explanations.

C. TLM-based RL Approach Solution

  • The approach combines TLM feedback with RL for dynamic resource allocation in O-RAN network slicing using dual-filtering BC.The solution is organized into five steps.Source: C. TLM-based RL Approach Solution · p. 5
  • DDPG is used for mixed-integer resource allocation, with slice KPIs, arrival rates, and service-demand characteristics as state inputs.Its off-policy learning supports exploration trajectories and expert demonstrations.Source: C. TLM-based RL Approach Solution · pp. 5–6
  • Dual filtering retains demonstrations that pass episode-level policy-impact gating and action-level critic validation.The filtered buffer supplies demonstrations for the BC loss.Source: C. TLM-based RL Approach Solution · p. 6
  • Training samples from both the exploration replay buffer and the filtered demonstration buffer.This combines self-exploration with TLM-guided demonstrations.Source: C. TLM-based RL Approach Solution · p. 6
  • The TLM evaluates proposed actions from encoded network states and returns a reasonability judgment, explanation, and corrected action when needed.For emergency URLLC congestion, it overrides an allocation to eMBB and mMTC with a latency-aware allocation.Source: C. TLM-based RL Approach Solution · p. 6
  • The closed loop generates and filters demonstrations, uses BC to guide policy learning, continues RL exploration, and updates guidance from performance feedback.The approach combines BC's sample efficiency with RL exploration.Source: C. TLM-based RL Approach Solution · p. 6

IV. PERFORMANCE EVALUATION

The supplied evaluation excerpts specify the workstation, episode duration, and dynamic traffic setting. They identify figures on emergency-traffic reasoning and BARX model results.

IV. PERFORMANCE EVALUATION

  • Experiments use an AMD Ryzen 9 9950X workstation with an RTX 5080 GPU.The RL agent is described as a BC-trained DDPG policy with vectorized state normalization.Source: §IV · p. 6
  • Episodes last 300 seconds and use sub-second control intervals under dynamic traffic and congestion scenarios.Source: §IV · p. 6
  • The evaluation excerpts identify Fig. 6 as LLM reasoning under emergency traffic and Table I as BARX results for LLM models.Source: Figure 6 · p. 6, Table I · p. 6

A. TLM Module Benchmarking

BARX evaluates fine-tuned open-source LLMs under multiple sampling regimes and strict latency and schema-compliance requirements. The supplied figures also identify measures of intervention stability, correction efficiency, latency, and BC's impact on RL.

A. TLM Module Benchmarking

  • Seven fine-tuned open-source LLMs are tested under three sampling regimes, with 100 JSON responses per model-regime combination.The study reports 2,100 total responses; TinyL-LaMA has the lowest latency at 342 ms but 100% parse errors.Source: A. TLM Module Benchmarking · p. 6
  • BARX assigns zero to models failing full schema compliance or the ℓ95 < 600 ms requirement; qualifying models are scored across latency, compliance, and policy-impact metrics.LLaMA3-8B is selected as the base reasoning module.Source: A. TLM Module Benchmarking · p. 6
  • Fig. 7 compares intervention strategies by coefficient of variation, performance gain per intervention, and latency distribution across five episodes.Lower coefficient of variation indicates greater stability.Source: Figure 7 · p. 7
  • Fig. 8 presents BC's impact on RL through mean reward, training-time breakdown, and coefficient of variation.The coefficient of variation indicates training stability.Source: Figure 8 · p. 7

B. Ablation of TLM Intervention Schedules

The ablation favors periodic TLM correction, while larger demonstration buffers improve reward and stability. In the over-the-air testbed, X-RIC500 provides larger SLA margins and lower resource use than the evaluated baselines.

B. Ablation of TLM Intervention Schedules

  • Periodic correction delivers the highest performance gain per correction (11.86), lowest coefficient of variation (15.2%), and median latency of 465.6 ms among six strategies.The study evaluates each strategy over five 300-second episodes using a pre-trained DDPG agent and deterministic TLM action predictions.Source: B. Ablation of TLM Intervention… · p. 7
  • 80.2% and 81.8% reward improvements over RL baseline for X-RIC300 and X-RIC500 accompany larger demonstration buffers, with higher-quality demonstrations reducing coefficient of variation.The passage reports moderate training-time increases and associates dual filtering with reduced learning instability and improved convergence.Source: B. Ablation of TLM Intervention… · pp. 7–8

C. Over-the-Air Testbed Implementation

  • The over-the-air evaluation uses an OpenAirInterface gNB, FlexRIC near-RT RIC, OAI 5G core, and two slices under controlled UDP congestion.Baselines include Equal Allocation, TLM-RL without behavioral cloning, and X-RIC initialized with different demonstration counts.Source: C. Over-the-Air Testbed Implementation · p. 8
  • 17 Mbps eMBB throughput and 5 ms URLLC latency give X-RIC500 margins of 42% above the 12.0 Mbps SLA and 50% below the 10.0 ms constraint.The passage reports minimal variance for both measures.Source: C. Over-the-Air Testbed Implementation · p. 8
  • 0.65 vCW usage (35% reduction) and 8.5–9.5 vRB blocks (25% reduction) let X-RIC500 satisfy SLAs with lower resource use.Equal Allocation uses vCW capacity Cw = 1.0 and 11.8–12.0 vRB blocks; TLM-RL and X-RIC300 show intermediate reductions.Source: C. Over-the-Air Testbed Implementation · pp. 8–9

V. CONCLUSION

The paper presents X-RIC as an agentic-AI framework for O-RAN, combining a hybrid control architecture, BARX evaluation, and dual-filtered behavioral cloning. Its SLA-aware network-slicing case study reports that X-RIC outperforms state-of-the-art baselines.

V. CONCLUSION

  • X-RIC integrates agentic AI into O-RAN with hybrid control, BARX evaluation, and dual-filtered automated demonstration collection, and outperforms state-of-the-art baselines in SLA-aware network slicing.The conclusion describes the case study as a three-network-slice system with dynamic configuration challenges for traditional RL within near-real-time RIC timescales.Source: §V · p. 9
Loading 2610.22269v1…