Paper deep dive

ACCoRD: Actor-Critic Conflict Resolution with Deep learning for O-RAN xApps

Cezary Adamczyk, Adrian Kliks

Year: 2026Venue: arXiv preprintArea: cs.MAType: PreprintEmbeddings: 27

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 95%

Last extracted: 7/8/2026, 6:53:34 PM

Summary

ACCoRD is an AI-driven conflict resolution method for O-RAN Near-RT RIC that uses an actor-critic ANN trained with PPO-Clip reinforcement learning to dynamically resolve control conflicts between xApps (like MRO and MLB), significantly reducing negative network events compared to static rule-based approaches.

Entities (8)

O-RAN · network-architecture · 99%ACCoRD · method · 98%Near-RT RIC · component · 97%PPO-Clip · algorithm · 96%MLB · xapp · 95%MRO · xapp · 95%Relative Penalty Score · metric · 94%Radio Link Failure (RLF) · network-event · 93%

Relation Signals (6)

ACCoRD → deployedin → Near-RT RIC

confidence 97% · resolve detected control conflicts in Near-Real Time RAN Intelligent Controller using a Conflict Resolution (CR) Agent

Artificial Neural Network (ANN) → trainedwith → PPO-Clip

confidence 96% · trained with a reinforcement learning algorithm PPO-Clip.

ACCoRD → uses → Artificial Neural Network (ANN)

confidence 95% · The implemented ANN analyzes data about the network and conflicting control decisions to infer optimal CR actions.

ACCoRD → resolvesconflictsbetween → MRO and MLB

confidence 94% · Their interplay causes indirect control conflicts. ACCoRD is evaluated against indirect control conflicts in this study.

Evaluation → usesmetric → Relative Penalty Score

confidence 93% · A novel evaluation metric, the Relative Penalty Score, is proposed to fairly rank methods across various scenarios.

ACCoRD → reduces → Radio Link Failure (RLF)

confidence 92% · significantly reducing negative network events caused by conflicting control decisions in medium and high traffic scenarios.

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Conflict Mitigation (ConMit) is a crucial part of intelligent network control in Open Radio Access Networks (O-RAN). In this paper, we propose a method named ACCoRD to resolve detected control conflicts in Near-Real Time RAN Intelligent Controller using a Conflict Resolution (CR) Agent with an Artificial Neural Network (ANN) trained with a reinforcement learning algorithm PPO-Clip. The implemented ANN analyzes data about the network and conflicting control decisions to infer optimal CR actions. The CR Agent gathers feedback from the network after each resolved conflict to assess its efficiency and adjust the ANN's weights during batch training. The evaluation of the proposed approach is based on simulation data. A new methodology for evaluating CR solutions is proposed. Results show that the proposed ANN-based method improves on the efficiency of rule-based approaches by significantly reducing negative network events caused by conflicting control decisions in medium and high traffic scenarios.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

26,835 characters extracted from source content.

Expand or collapse full text

ACCoRD: Actor-Critic Conflict Resolution with Deep learning for O-RAN xApps Cezary Adamczyk Institute of Radiocommunications Poznan University of Technology Pozna ́ n, Poland cezary.adamczyk@doctorate.put.poznan.pl Adrian Kliks Institute of Radiocommunications Poznan University of Technology Pozna ́ n, Poland Lule ̊ a University of Technology Lule ̊ a, Sweden adrian.kliks@put.poznan.pl Abstract—Conflict Mitigation (ConMit) is a crucial part of intelligent network control in Open Radio Access Networks (O- RAN). In this paper, we propose a method named ACCoRD to resolve detected control conflicts in Near-Real Time RAN Intelligent Controller using a Conflict Resolution (CR) Agent with an Artificial Neural Network (ANN) trained with a reinforcement learning algorithm PPO-Clip. The implemented ANN analyzes data about the network and conflicting control decisions to infer optimal CR actions. The CR Agent gathers feedback from the network after each resolved conflict to assess its efficiency and adjust the ANN’s weights during batch training. The evaluation of the proposed approach is based on simulation data. A new methodology for evaluating CR solutions is proposed. Results show that the proposed ANN-based method improves on the efficiency of rule-based approaches by significantly reducing negative network events caused by conflicting control decisions in medium and high traffic scenarios. Index Terms—conflict resolution, reinforcement learning, con- flict mitigation, Near-RT RIC, O-RAN, xApp. I. INTRODUCTION O NE of the promises of Open Radio Access Networks (O-RAN) is to enable multi-vendor RAN with robust network optimization solutions [1]. Mobile Network Operators (MNOs) that deploy O-RAN-compliant networks can mix and match hardware and software components from numerous vendors. Such heterogeneity of the RAN, if not managed properly, can lead to harmful interoperability issues. As stan- dards describing O-RAN mature over time and complexity of O-RAN deployments advances, Conflict Mitigation (ConMit) in O-RAN’s RAN Intelligent Controllers (RICs) becomes an increasingly important research topic. The O-RAN architecture is designed with native support for Artificial Intelligence (AI) and Machine Learning (ML) techniques. These methods are ideally suited for ConMit, where the execution of complex processing logic against large datasets can address challenges such as ensuring the reliability of conflict detection mechanisms and determining optimal Conflict Resolution (CR) logic. The complexity and This is the author’s version of an article that has been published in the proceedings of the 2026 IEEE INFOCOM conference. dynamic nature of O-RAN make efficient conflict mitigation using static, rule-based conventional approaches challenging. These AI/ML techniques are envisioned as enablers for robust ConMit measures, which can adapt to ever-changing network conditions and the evolving suite of deployed applications. AI/ML enables the switch from conflict resolution using static rules (e.g., fixed application priorities) to dynamic opti- mization. The expected approach involves an AI model acting as the core of the resolution logic. This model ingests a wide range of parameters describing network state and the detected conflict to decide on the optimal set of resolution actions. Such a system can benefit from ML approaches, like Reinforcement Learning (RL), allowing the AI model to adapt its operation and improve its decision-making over time based on the reward signal from the O-RAN environment. A. Related works Although it is a relatively new field of research, there is already a variety of work on ConMit in O-RAN. One of the first steps towards achieving a standardized ConMit approach was the introduction of the Conflict Mitigation Framework (CMF) by Adamczyk and Kliks [2], which defined archi- tectural extensions for the Near-Real Time (Near-RT) RAN Intelligent Controller (RIC) but relied on simple prioritization methods. Other works, such as QoS-Aware Conflict Mitigation (QACM) [3], propose rule-based methods focusing on QoS tar- gets but require non-standard inputs from xApps, making them impractical for multi-vendor deployments. Similarly, digital twin-based approaches like generalized Conflict Management scheme for Multi-Channel Power Control in O-RAN xApps (COMIX) [4] face challenges regarding computational delays and model accuracy. Alternative approaches like PACIFISTA [5] operate from the Service Management and Orchestration (SMO) level, resolving conflicts by disabling applications entirely, which removes both negative and positive impacts. Attempts at ML-powered solutions, such as xApp distillation [6], require complex operational efforts to compile multiple policies into a single agent, lacking ”plug & play” capability. Copyright © 2026 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. arXiv:2605.22306v1 [cs.MA] 21 May 2026 B. Problem statement and proposed solution To address the identified gap in robust CR methods for ConMit in the Near-RT RIC, we propose Actor-Critic Con- flict Resolution with Deep learning (ACCoRD) (Actor-Critic Conflict Resolution with Deep learning), a novel CR solution based on Artificial Neural Network (ANN), designed for deployment within the CMF. ACCoRD functions as the main decision-making entity within the CR Agent. By integrating an ANN, ACCoRD analyzes control conflicts by processing data regarding conflicting decisions and current network decisions to infer target resolution actions. While conflicts in O-RAN can be categorized into direct, indirect, and implicit types [7], ACCoRD is evaluated against indirect control conflicts in this study. However, its feature encoding design allows it to scale to other conflict types in the future. ACCoRD operates by intercepting conflict reports from the Conflict Detection (CD) Agent, collecting the necessary context information, and pre-processing inputs for its ANN. For each detected conflict, the ANN infers a set of resolution actions, which are applied to the conflicting control decisions. An actor-critic architecture trained via the Proximal Policy Optimization (PPO)-Clip algorithm is employed to enable ANN’s adaptation to dynamic network conditions. The operation of ACCoRD follows a specific data flow: 1) Control decision issuance: xApps submit their control decisions as E2 Control messages. 2) Conflict detection: The CD Agent analyzes the mes- sages. If a conflict is detected, a conflict report is generated and sent to the CR Agent. 3) Computation of policy logits: ACCoRD preprocesses conflict data with global features and feeds it through the ANN to compute CR action logits and a critic value. 4) Execution of conflict resolution actions: ACCoRD interprets the logits to determine resolution actions (e.g., modify, reject), applies them, and submits the post- resolution decisions. 5) Post-resolution observation: After resolving a conflict, ACCoRD triggers a post-resolution observation of the network to calculate the reward for RL training. To visualize these interactions, the detailed data flow for ACCoRD is presented in Figure 1. I. DESIGN OF THE ANN IN ACCORD Using AI/ML techniques, such as an ANN trained using one of the RL algorithms, ConMit can be performed without explicitly defining a specific resolution logic. The efficiency of an AI/ML-based solution relies heavily on its design: input feature selection, ANN architecture, utilized ML algorithm, and the reward function. A. ANN inputs The input architecture for ACCoRD’s ANN is designed to simultaneously process the aggregate environment state and the details of conflicting control decisions. The ANN accepts inputs divided into two primary categories: global features, which provide contextual information regarding the target cell and its active users, and control decision features, which encode the specific control decisions involved in the conflict. 1) Control decision features: Each inference step processes data describing up to N ConfDec conflicting control decisions. The ANN ingests these as a fixed-size, slot-based input tensor of dimensions (N ConfDec , 5). Each slot i corresponds to a deci- sion head and contains a feature vector x i = [e type ∥ v ′ param ]. • Parameter type embedding (e type ): a 4-dimensional one-hot vector identifying the parameter type (none, CIO, T, hysteresis). This ensures the ANN interprets pa- rameters based on learned categorical relationships rather than arbitrary numeric representations. • Parameter value (v ′ param ): the numeric value of the control parameter, normalized to the range [0, 1] using min-max scaling specific to the parameter type bounds. If a parameter type is absent from a conflict, its slot is filled with the none embedding and a zero value, and masked during processing. 2) Global features: Global features anchor the decision- making in the current network context. This vector consists of 11 floating-point values, explicitly normalized to ensure training stability: • Cell state: includes current values for hysteresis, CIO, and T (normalized by their maximum configurable values), and KPIs for availability (load < 100%), han- dover stability (non-ping-pong ratio), and total throughput (normalized by a constant of 100 Mbps). • User state: aggregated state of users connected to the target cell, including movement speed (normalized by 100 m/s), requested bitrate (normalized by 1 Gbps), satisfaction ratio, connection success rate, and throughput (normalized by 1 Mbps). B. ANN outputs The output layer defines the action space and provides auxiliary outputs for training. ACCoRD generates three sets of outputs: CR action logits, a validity mask, and a critic value. 1) CR action logits: The ANN utilizes a multi-head output structure with N ConfDec decision heads. Each head outputs a vector of N A = 4 logits, representing the preference for specific resolution actions. These logits are transformed via Softmax into a probability distribution P (a i ). The discrete action space A consists of: • NO_MODIFICATION: The decision is applied as is. • REJECTION_WITH_COOLDOWN: The decision is re- jected, and the source xApp is put on a cooldown. • INCREASE_1/DECREASE_1: The target parameter is incremented/decremented by a single logical step. The INCREASE_1 and DECREASE_1 actions map to the discrete values supported by the specific xApp (e.g., specific T values defined in 3GPP standards). For parameters with non-linear spacing, such as T, this method ensures that ACCoRD strictly adheres to valid configuration steps without requiring separate logic for each parameter type. Fig. 1. Detailed data flow for ACCoRD operation 2) Validity mask: Because the ANN utilizes a fixed number of decision heads (N ConfDec ) to process a variable number of conflicting decisions, not all output heads produce valid data for every inference step. To address this, the ANN outputs a validity mask—a boolean tensor of dimension N ConfDec . This mask corresponds to the input slots and identifies which decision heads represent real conflicting control decisions and which correspond to padding. During post-processing, this mask is applied to the output logits; invalid slots are masked with negative infinity (−∞). This mechanism ensures that their resulting probability after the Softmax operation is effectively zero. Consequently, the ANN’s policy is derived from valid slots only, preventing training on padding noise. 3) Critic value: The critic head outputs a scalar state-value estimate V (s), representing the expected cumulative reward, used to compute advantages during training. Mathematically, the critic approximates the expected return starting from state s following policy π θ : V (s) =E " ∞ X t=0 γ t r t | s 0 = s,π θ # (1) where γ is the discount factor and r t is the reward at time t. C. ANN structure The ANN implemented in ACCoRD is a feed-forward actor-critic network designed to handle a variable number of conflicting decisions while maintaining context awareness. Its internal architecture comprises three main functional blocks: 1) Control decision feature encoder: This component is a shared two-layer Multilayer Perceptron (MLP) with Rectified Linear Unit (ReLU) activation. It processes the normalized input feature (v ′ param ) for each conflicting decision indepen- dently. This encoding transforms the scalar input into a higher- dimensional latent representation, which allows the subsequent layers to process abstract features rather than raw values. Fig. 2. Structure of the actor-critic ANN implemented in ACCoRD 2) Control decision heads (actor): The network contains N ConfDec distinct control decision heads. To ensure that the resolution of one conflict considers the context of others, each head receives a composite input vector consisting of: • The encoded features of the specific control decision. • The parameter type embedding (e type ). • The vector of global features (cell and user state). • A mean-pooled summary of the encoded features from all other conflicting decisions. This context-aware design allows the head to output a conflict resolution action that is locally optimal for the specific deci- sion but globally consistent with the overall conflict scenario. 3) Critic: The critic is implemented as a three-layer MLP with ReLU activation and layer normalization applied after the first layer. Unlike the actor heads, which focus on individual decisions, the critic evaluates the state of the entire system. It takes the global features and the mean-pooled summary of all encoded control decision features as input to produce the scalar state-value estimate V (s). Figure 2 illustrates this structure. D. Training process Training utilizes the PPO-Clip algorithm, which balances implementation simplicity with reliable convergence. It was selected for ACCoRD because its strict bounding of policy updates provides good training stability, which is considered critical to prevent policy collapse in the highly non-stationary environment caused by the simultaneous operation of multiple uncoordinated xApps. The ”Clip” variant of PPO stabilizes Fig. 3. Data flow for ACCoRD’s RL training training by utilizing a clipped surrogate objective function L CLIP [8], defined as: L CLIP (θ) = ˆ E t h min(r t (θ) ˆ A t , clip(r t (θ), 1− ε, 1 + ε) ˆ A t ) i (2) where r t (θ) is the probability ratio between new and old policies, ˆ A t is the estimated advantage, and ε is the clipping hyperparameter. The flow of state, action, and reward data during the training phase is depicted in Fig. 3. Experience is gathered in 400 s intervals. At the end of each interval, observations are split into minibatches and used to update the policy over 3 PPO epochs. To ensure policy stability, Kullback-Leibler (KL) divergence is tracked using the approximate formula: D KL = ˆ E t log π θ old (a t |s t ) π θ (a t |s t ) (3) Consecutive steps with significantly high divergence trigger learning rate adjustments or early epoch termination. A 12- step pretraining of the critic is performed in each batch using Huber loss L δ to improve baseline estimations: L δ (y,f (x)) = ( 1 2 (y− f (x)) 2 for |y− f (x)|≤ δ, δ(|y− f (x)|− 1 2 δ) otherwise. (4) The PPO algorithm was configured with 3 epochs per update, a clipping parameter of 0.10, a target mean KL of 0.015, and an entropy bonus coefficient of 0.007. Starting learning rates were set to 10 −5 for the actor and 10 −4 for the critic. E. Reward calculation The reward function R network reflects the impact of CR de- cisions on global network performance. Metrics are observed in a rolling window of length T meas at detection time (t = 0) and post-resolution (t = T reward ). The reward is calculated as the weighted sum of changes in normalized metrics for Ping- Pong handovers (C P ), Radio Link Failure (RLF) (C RLF ), and Call Blockages (C CB ): R network =− w rew P ∆ ˆ C P + w rew RLF ∆ ˆ C RLF + w rew CB ∆ ˆ C CB w rew P + w rew RLF + w rew CB (5) Weights (w rew ) allow the operator to align the agent with specific intents. In this study, a balanced policy (w rew = 1.0 for all) is used. I. SIMULATION SCENARIO To evaluate the efficiency of the implemented CR Agent, a robust simulation of an O-RAN-compliant network is con- ducted. The simulation scenario comprises a hexagonal ar- rangement of eight 5G base stations (BSes). A single macro BS collocated with a micro BS is located in the center of the simulated area, with six evenly spaced micro BSes surrounding it. Each micro BS is configured with three 5G cells, each covering a 120-degree sector. The simulation area is defined as a circle with a radius of 450 m. To simulate varying levels of network load, User Equip- ments (UEs) are distributed within the simulation area accord- ing to three distinct deployment configurations: • Small: 15 UEs connected to the central macro BS and 6 UEs per micro BS, totaling 57 UEs. • Medium: 20 UEs per macro BS and 8 UEs per micro BS, totaling 76 UEs. • Large: 30 UEs per macro BS and 10 UEs per micro BS, totaling 100 UEs. Additionally, a single dedicated ’Reference UE’ is included in each run on top of these distributions to enable granular analysis of individual connectivity parameters during simula- tion. UEs exhibit diverse mobility (pedestrian and vehicular) and heterogeneous traffic profiles (low-bitrate voice, medium- bitrate data, high-bitrate data). The simulated network hosts two xApps: Mobility Ro- bustness Optimization (MRO) and Mobility Load Balanc- ing (MLB). MRO optimizes handover hysteresis and time- to-trigger (T) to minimize ping-pongs and RLFs. MLB controls Cell Individual Offset (CIO) to balance load. Their interplay causes indirect control conflicts. A. Parametrization of ACCoRD The operation of ACCoRD is configured by structural and operational parameters: • Maximum conflicting decisions (N ConfDec ): Set to 3 (CIO, T, Hysteresis). • Action space size (N A ): Set to 4. • Measurement window (T meas ): Set to 1.0 s to align with xApp processing intervals. • Cooldown duration (t CR ): Set to 10 s. • Training rollout (t training ): Experience is collected for 400 s before a policy update. IV. METHODOLOGY The evaluation is structured around various ConMit configu- rations to assess the impact of ACCoRD. The baseline config- uration is CMF inactive (No CM). The second configuration uses a rule-based Prioritization with Cooldown method, tested in two variants: prioritizing MRO and prioritizing MLB. The final configuration is ACCoRD. A. Evaluation procedure As the ANN is untrained upon initialization, it requires a substantial time of environment observation. To evaluate performance, ACCoRD is first trained during 10,000 seconds of simulation. Subsequently, a 1,000-second evaluation run is performed to assess the trained ANN’s capabilities. To ensure stable conditions, the initial 200 seconds of each run are discarded. Each configuration is tested in 36 different initial random UE placements. B. Evaluation metric A novel evaluation metric, the Relative Penalty Score, is proposed to fairly rank methods across various scenarios. First, a total penalty value P i is calculated for each simulation run by summing negative events weighted by their severity: P i = w pen P C i P + w pen RLF C i RLF + w pen CB C i CB (6) Weights are set to highlight the impact of service disruption: w pen P = 0.05, w pen RLF = 0.40, and w pen CB = 0.40. The metric is then normalized relative to the best-performing method (P best ) in the set: P i rel = P i P best · 100%(7) V. EVALUATION RESULTS A. Statistical results Table I summarizes the outcomes, showing the relative penalty metric averaged over 36 simulation runs. In the Small UE deployment, the Prioritize MRO method performs best (102.5%), with ACCoRD close behind at 104.1%. The rule-based method suffices for low traffic. How- ever, in the Medium UE deployment, ACCoRD demonstrates robustness, achieving a score of 101.4% and outperforming Prio. MRO by 2.8 p. The Large UE deployment shows TABLE I EVALUATION RESULTS BASED ON 36 SIMULATION RUNS RankAlgorithmAvg Penalty Avg Rel. Penalty (%) Delta to best (p) Small UE deployment configuration 1Prio. MRO292.13102.50.0 2ACCoRD297.73104.11.6 3ConMit disabled302.97106.33.8 4Prio. MLB330.43116.013.5 Medium UE deployment configuration 1ACCoRD402.84101.40.0 2Prio. MRO412.82104.22.8 3ConMit disabled424.45107.15.7 4Prio. MLB448.79113.211.8 Large UE deployment configuration 1ACCoRD549.87101.60.0 2Prio. MRO568.56105.64.0 3ConMit disabled583.45108.36.7 4Prio. MLB602.99111.910.3 even greater improvement, with ACCoRD scoring 101.6% and widening the gap to the next best method to 4.0 p. Prio. MLB consistently performs the worst. The efficiency of ACCoRD scales favorably with network load and conflict frequency, proving its value in challenging conditions. B. Case Study: CR policy derived by ACCoRD We examine a policy derived by ACCoRD in a medium deployment run where it achieved the lowest penalty. The agent learned a strategy governed by the following rules: • CIO stabilization: Conflicting decisions regarding CIO are always rejected, effectively stabilizing cell borders to prevent load oscillations. • Handover inertia: Decisions regarding T and Hys- teresis are either rejected or increased by single steps. Increasing these parameters creates ”inertia” in the sys- tem: higher T extends the time required to trigger a handover, while higher hysteresis requires a larger signal delta. This dampens the system’s reaction to transient channel fluctuations. • Overall rejection: In nearly 64% of conflicts, all con- flicting control decisions were rejected. Although the ”reject-all” strategy in nearly 64% of conflicts may appear conservative and could potentially block some intended xApp optimizations, it effectively acts as a noise filter that prioritizes network stability over aggressive, uncoor- dinated adjustments. This behavior is a direct consequence of the balanced reward policy used in this study, which heavily penalizes service disruptions. Tuning these reward weights allows MNOs to adjust the agent’s risk tolerance, potentially yielding more aggressive optimization profiles. Overall, this policy minimized ping-pong handovers and call blockages (CBs) significantly compared to baselines, while keeping RLF counts competitive, demonstrating ACCoRD’s ability to autonomously derive complex stabilization strategies. C. Computational complexity and inference latency To ensure practical viability within the Near-RT RIC, where the control loop operates between 10 ms and 1 s, the CR Agent must not introduce excessive delays. The inference time of the proposed ANN is bounded by the O(N ConfDec · L· d 2 ) com- putational complexity of the feed-forward MLP architecture, where L is the number of sequential layers in the inference path and d is the maximum layer width. Because the critic network is utilized exclusively during the training phase to compute advantages, it is omitted from the real-time execution path. Consequently, the runtime complexity depends solely on the shared feature encoder and the actor decision heads. Given the relatively low-dimensional state space and the lightweight architecture of ACCoRD (N ConfDec = 3, L = 4, d = 519), a single forward pass requires approximately 3.2 million multiply-accumulate operations. On standard CPU hardware, this executes in the microsecond to low-millisecond range. This negligible computational overhead ensures adherence to the latency requirements of the Near-RT control loop. VI. SUMMARY This paper introduced ACCoRD, an ANN-based CR method capable of adapting its policy to network conditions. AC- CoRD utilizes a novel input/output masking scheme to handle variable conflicts and learns through PPO-Clip to minimize a weighted penalty of negative network events. Simulations demonstrated that while simple prioritization works for low loads, ACCoRD provides unmatched performance in medium and high traffic scenarios. Future work will focus on: • New control targets: including control target type em- beddings (user, slice) to handle diverse conflict scopes. • Scalability: expanding the number of input heads to handle larger conflict sets without structural redesign. • Action space expansion: investigating other CR actions, including multi-step modifications or rejection without cooldown to fine-tune control. • Reward shaping and diverse KPIs: evaluating non- uniform reward weights (w rew ̸= 1.0) to reflect nuanced operator priorities, and analyzing the impact on additional metrics such as PRB utilization and spectral efficiency. • Handling rare events: adapting the training process to handle highly imbalanced datasets, where specific conflict instances may occur in less than 10% of operational time. • Real-world validation: transitioning from simulation to hardware-in-the-loop physical testbeds to verify the solution’s performance, generalization, and latency in complex, higher-dimensional over-the-air scenarios. REFERENCES [1] M. Polese, L. Bonati, S. D’Oro, S. Basagni, and T. Melodia, “Understand- ing O-RAN: Architecture, Interfaces, Algorithms, Security, and Research Challenges,” IEEE Communications Surveys & Tutorials, vol. 25, no. 2, p. 1376–1411, 2023. [2] C. Adamczyk and A. Kliks, “Conflict Mitigation Framework and Conflict Detection in O-RAN Near-RT RIC,” IEEE Communications Magazine, vol. 61, no. 12, p. 199–205, 2023. [3] A. Wadud, F. Golpayegani, and N. Afraz, “QACM: QoS-Aware xApp Conflict Mitigation in Open RAN,” IEEE Transactions on Green Com- munications and Networking, vol. 8, no. 3, p. 978–993, 2024. [4] A. E. Giannopoulos, S. T. Spantideas, G. Levis, A. S. Kalafatelis, and P. Trakadas, “COMIX: Generalized Conflict Management in O- RAN xApps—Architecture, Workflow, and a Power Control Case,” IEEE Access, vol. 13, p. 116 684–116 700, 2025. [5] P. B. d. Prever, S. D’Oro, L. Bonati, M. Polese, M. Tsampazi, H. Lehmann, and T. Melodia, “PACIFISTA: Conflict Evaluation and Management in Open RAN,” IEEE Transactions on Mobile Computing, p. 1–15, 2025. [6] H. Erdol, X. Wang, R. Piechocki, G. Oikonomou, and A. Parekh, “xApp Distillation: AI-based Conflict Mitigation in B5G O-RAN,” Computer Networks, vol. 274, p. 111848, 2026. [Online]. Available: https://w.sciencedirect.com/science/article/pii/S138912862500814X [7] O-RAN Working Group 3, Near-Real-time RAN Intelligent Controller Conflict Mitigation, O-RAN Alliance, October 2024, v01.00. Accessed: 2025-11-26. [Online]. Available: https://w.o-ran.org/specifications [8] J.Schulman,F.Wolski,P.Dhariwal,A.Radford,and O.Klimov,“ProximalPolicyOptimizationAlgorithms,”2017, doi:10.48550/arXiv.1707.06347. [Online]. Available: https://arxiv.org/ abs/1707.06347

Instant research discovery