Fluid Antenna Networks Beyond Beamforming: An AI-Native Control Paradigm for 6G

Submitted to IEEE Wireless Communications Magazine, 2026 Fluid An tenna Net w orks Bey ond Beamforming: An AI-Nativ e Con trol P aradigm for 6G Ian F. Akyildiz, Life F ello w, IEEE and T uğçe Bilen, Mem b er, IEEE Abstract—Fluid Antenna Systems (F AS) introduce a new degree of freedom for wireless netw orks b y enabling the ph ysical an tenna p osition to adapt dynamically to c hanging radio con- ditions. While existing studies primarily emphasize physical- la yer gains, their broader implications for netw ork operation re- main largely unexplored. Once an tennas b ecome recongurable en tities, antenna p ositioning naturally b ecomes part of the net work con trol problem rather than a standalone optimization task. This article presents an AI-native persp ective on uid an tenna netw orks for future 6G systems. Instead of treating an tenna rep ositioning as an isolated operation, we consider a closed-lo op control arc hitecture in which an tenna adaptation is join tly managed with conv entional radio resource management (RRM) functions. Within this framework, real-time netw ork observ ations are translated into co ordinated antenna and re- source conguration decisions that resp ond to user mobility , trac demand, and ev olving in terference conditions. T o ad- dress the complexit y of m ulti-cell environmen ts, w e explore a m ulti-agent reinforcement learning (MARL) approac h that enables distributed and adaptive control across base stations. Illustrativ e results sho w that in telligent an tenna adaptation yields consistent performance gains, particularly at the cell edge, while also reducing inter-cell interference. These ndings suggest that the true p oten tial of uid antenna systems lies not only in recongurable hardware, but in intelligen t net work con trol architectures that can eectively exploit this additional spatial degree of freedom. Index T erms—Fluid antenna systems, 6G net works, AI- nativ e wireless netw orks, multi-agen t reinforcement learning, radio resource management I. Introduction Wireless communication systems ha v e steadily ev olv ed to w ard greater exibilit y and programmabilit y . Early cellular netw orks relied on largely static infrastructures, where an tennas, transmission strategies, and propagation en vironmen ts oered limited adaptabilit y during oper- ation. Ov er time, ho wev er, wireless technologies ha ve in tro duced increasing lev els of recongurability . Massive m ultiple-input m ultiple-output (MIMO) systems enabled programmable spatial pro cessing, while recongurable in- telligen t surfaces (RIS) allow ed direct manipulation of the propagation en vironment [1]. T ogether, these adv ances reect a broader shift: ph ysical components that w ere once static are b ecoming controllable and adaptive. Fluid an tenna systems (F AS) represent a further step in this evolution. Unlike conv en tional antennas with xed Ian F. Akyildiz is with T ruv a Inc., Alpharetta, GA 30022, USA (e-mail: ian@truv ainc.com). T uğçe Bilen is with the Department of Articial Intelligence and Data Engineering, Istanbul T ec hnical Univ ersit y , Istanbul, T urkey (e-mail: bilent@itu.edu.tr). p ositions and radiation characteristics, uid an tennas allo w their radiating elements to mo v e within a conned region [2]. By adapting their position or conguration to radio conditions, they can dynamically reshape the eectiv e channel exp erienced b y users. In con trast to MIMO and RIS, which optimize transmission ov er a giv en environmen t, F AS enables direct ph ysical control of the antenna itself, allowing the netw ork to inuence the c hannel realization. This in tro duces a new spatial degree of freedom, as illustrated in Fig. 1, extending adaptation b ey ond signal pro cessing to the physical antenna in terface. While F AS is often viewed as a physical-la yer enhance- men t that impro v es c hannel qualit y and exploits spatial div ersit y , its implications extend b eyond the physical la y er. In large-scale cellular systems, antenna congu- ration b ecomes in tert wined with user sc heduling, b eam managemen t, pow er allo cation, and inter-cell interference co ordination. Since wireless netw orks contin uously adapt to user mobilit y , trac demand, and channel v ariations, dynamically adjustable an tennas naturally in tro duce an additional con trol v ariable. A base station may reposition its an tenna, adjust b eamforming strategies, or adapt resource allo cation in resp onse to netw ork conditions, leading to a tight coupling betw een antenna adaptation and radio resource managemen t. This p ersp ectiv e suggests that F AS should b e view ed not merely as an antenna design paradigm, but as a new dimension of netw ork control. Antennas ev olv e from xed comm unication interfaces in to dynamic, state-dep endent net w ork assets. The k ey c hallenge is therefore to deter- mine ho w and when antenna congurations should be adapted under dynamic conditions. Articial intelligence (AI) pro vides a natural approac h to this problem, as learning-based metho ds can handle complex, time-v arying en vironmen ts where analytical modeling b ecomes dicult. Fluid an tenna net works exhibit exactly such characteris- tics, with large conguration spaces and p erformance that dep ends on user distribution, mobility , and in terference dynamics. Motiv ated b y these observ ations, this article explores an AI-nativ e p ersp ective on uid an tenna netw orks. Rather than fo cusing solely on antenna-lev el optimization, w e examine ho w uid an tennas can b e integrated into intelli- gen t netw ork con trol architectures that adapt to changing net w ork conditions. In particular, antenna reconguration is em b edded within a closed-lo op decision pro cess that join tly manages an tenna adaptation and con ven tional radio resource allocation. W e in tro duce an AI-driven Submitted to IEEE Wireless Communications Magazine, 2026 Fig. 1: (a) Evolution to w ard recongurable wireless netw orks, from static infrastructures to programmable wireless tec hnologies and uid an tenna systems (F AS). (b) Proposed AI-native closed-loop control arc hitecture for uid antenna net w orks, where netw ork observ ations driv e intelligen t an tenna adaptation and radio resource managemen t decisions. con trol framew ork for uid an tenna net works and discuss ho w learning-based approaches enable distributed and scalable decision-making across m ultiple base stations. As a represen tativ e example, w e consider a m ulti-agent reinforcemen t learning (MARL) approac h that allo ws base stations to co ordinate antenna p ositioning, beamforming, and resource allocation under dynamic conditions. W e fur- ther present illustrativ e simulation results demonstrating ho w adaptiv e antenna congurations can impro ve net work throughput, enhance cell-edge performance, and mitigate in ter-cell interference compared with con ven tional xed- an tenna systems. Finally , we discuss practical implemen- tation c hallenges and outline several researc h directions for uid an tenna netw orks in future 6G systems. I I. Recongurable Wireless Systems: F rom PHY Optimization to Net work Control Recongurable wireless tec hnologies hav e b een exten- siv ely studied to improv e the adaptability and p erfor- mance of comm unication systems. MIMO enables pro- grammable spatial pro cessing through large an tenna ar- ra ys [3], while RIS allows dynamic manipulation of the propagation en vironmen t via con trollable reections [4], [5]. These approaches signican tly enhance sp ectral e- ciency and cov erage. Ho wev er, they fundamentally operate under xed antenna infrastructures and do not treat an tenna p osition as a controllable v ariable. Building on this trend, F AS hav e recently emerged as a nov el paradigm that enables dynamic reconguration of an tenna p ositions, in tro ducing a new spatial degree of freedom for wireless comm unications. Existing stud- ies primarily fo cus on ph ysical-la yer asp ects, including c hannel gain optimization, spatial diversit y enhancement, and interference mitigation. F or example, relay-assisted and RIS-in tegrated F AS architectures hav e demonstrated notable improv emen ts in sp ectral eciency and reliability [6]–[9]. How ever, these works are largely limited to link- lev el or single-cell scenarios and do not explicitly address net w ork-level control and m ulti-cell in teractions. AI tec hniques ha v e also b een widely applied to radio resource managemen t (RRM) in wireless netw orks. Re- inforcemen t learning (RL) and MARL hav e b een used for user scheduling, p ow er control, and interference co- ordination in dynamic environmen ts [10]–[12]. These ap- proac hes pro vide scalable solutions for high-dimensional and time-v arying systems. Ho w ev er, they t ypically assume xed antenna congurations and do not consider an tenna recongurabilit y as part of the control space. T aken together, these researc h directions rev eal a clear disconnect in the literature. Recongurable wire- less technologies, including uid antenna systems, hav e b een predominan tly in v estigated from a physical-la y er p ersp ectiv e, focusing on channel enhancement, diversit y gains, and in terference mitigation, often in isolated or single-cell settings. In parallel, AI-driv en approaches hav e b een extensiv ely developed for net w ork-lev el optimization, particularly for radio resource management, but typically under the assumption of xed antenna infrastructures. As a result, the in terplay b etw een an tenna recon- gurabilit y and netw ork-level control decisions remains insucien tly explored, esp ecially in m ulti-cell and dy- namic en vironmen ts. While recent eorts hav e b egun to incorp orate learning-based tec hniques into recongurable systems, antenna conguration is still generally treated as a standalone optimization v ariable rather than an integral comp onen t of the netw ork control space. In contrast, this work adopts a unied con trol p ersp ec- tiv e in which an tenna conguration is explicitly embedded in to the netw ork decision pro cess. Sp ecically , antenna adaptation is jointly considered with conv entional radio resource management functions within a learning-driv en framew ork, enabling coordinated decision-making across Submitted to IEEE Wireless Communications Magazine, 2026 m ultiple cells. By treating antenna states as part of b oth the system state and the control action space, the prop osed approach bridges the gap betw een physical-la yer recongurabilit y and net w ork-lev el in telligence, position- ing uid antennas not merely as a link-level enhancemen t but as a fundamental control dimension for future 6G net w orks. I I I. Fluid Antenna Net works as a New Con trol Dimension Fluid antenna systems introduce a new degree of free- dom by enabling dynamic adaptation of antenna cong- urations. While often studied for channel enhancement and diversit y gains, their impact extends b eyond the ph ysical la y er. In m ulti-user and m ulti-cell en vironmen ts, an tenna conguration decisions interact with scheduling, b eamforming, and in terference co ordination, thereb y in- tro ducing an additional con trol dimension within the radio access net w ork. A uid antenna-enabled base station can b e viewed as a dynamic system whose op erational state evolv es o v er time. At any given time t , the antenna congu- ration of base station b can b e represented as A b ( t ) = ( p b ( t ) , θ b ( t ) , P b ( t ) ) , where p b ( t ) denotes the an tenna p o- sition within the allow able region, θ b ( t ) represen ts the orien tation or b eam direction, and P b ( t ) captures the radiation pattern or activ e conguration. These parame- ters collectively determine ho w the an tenna interacts with the surrounding propagation environmen t. This in terac- tion is particularly evident in the wireless channel. In con v entional systems with xed antennas, the c hannel primarily dep ends on user lo cation, propagation condi- tions, and m ultipath eects. In con trast, uid an tennas in tro duce an additional dep endency on the instan taneous an tenna conguration. Accordingly , the channel betw een user u and base station b at time t can b e expressed as h u , b ( t ) = f ( p b ( t ) , θ b ( t ) , lo cation u ( t ) , en vironmen t ) , where f ( · ) captures the combined inuence of antenna state, user p osition, and environmen tal characteristics. This form ulation highligh ts a key distinction: by adapting its an tenna conguration, the base station can actively shap e the eectiv e channel conditions exp erienced by its users. F rom a net work persp ective, antenna conguration b e- comes an additional control v ariable alongside scheduling, b eam selection, and p ow er allo cation. These decisions are inherently coupled. F or example, rep ositioning the an tenna may improv e the signal qualit y for a target user while simultaneously altering interference patterns across neighboring cells, thereby aecting user asso ciation and ov erall resource eciency . This coupling leads to a joint control problem in whic h antenna states and radio resource management decisions must b e optimized together. Solving this joint problem is challenging for several reasons. The antenna conguration space can b e large and potentially contin uous, particularly when multiple mo v ement and radiation options are av ailable. Net work conditions evolv e ov er time due to user mobility , trac uctuations, and dynamic in terference, while decisions at one base station inuence neighboring cells through in ter- cell coupling. As a result, uid antenna netw orks form high-dimensional and dynamic control systems, where con v entional rule-based or static optimization approaches are often insucien t. This complexity motiv ates the use of learning-based control mec hanisms that can adapt an- tenna congurations and resource allocation policies based on observ ed net w ork b ehavior. In the follo wing section, we build on this p ersp ectiv e and in tro duce a learning-driv en con trol arc hitecture for uid antenna netw orks. IV. AI-Native Con trol Architecture for Fluid An tenna Net w orks The previous section established that uid antenna systems introduce an additional control dimension into wireless netw orks. Once antenna congurations b ecome dynamically adjustable, the netw ork must determine ho w an tenna adaptation should interact with conv en tional radio resource management decisions. Addressing this c hallenge requires a control framework that con tinuously observ es netw ork conditions, interprets their implications, and selects appropriate an tenna and resource congura- tions in real time. T o enable this capabilit y , we consider an AI-native con trol architecture that integrates an tenna adaptation in to a closed-loop decision pro cess. Rather than treating an tenna conguration as an isolated optimization task, the prop osed framew ork em b eds an tenna con trol within a broader con text of net w ork state a wareness and adaptive decision-making. A. Netw ork State Representation Eectiv e con trol begins with constructing a compact yet expressiv e representation of the netw ork state. In uid an- tenna netw orks, this state must capture both conv entional net w ork information and the current antenna congura- tion, since antenna p ositioning directly aects channel conditions and in terference patterns. A represen tativ e state comprises c hannel measurements, user lo cations and trac demand, base-station antenna congurations, and in ter-cell in terference indicators. T ogether, these elements pro vide a structured description of the netw ork context, enabling the control mechanism to reason ab out the impact of dieren t congurations on system p erformance. Imp ortan tly , antenna conguration is not an external parameter but an in tegral component of the net w ork state, as it directly inuences the communication environmen t. B. Intelligen t Decision Engine Giv en the current netw ork state, the system must de- termine how to adapt. This functionality is performed b y an intelligen t decision engine that maps observed states to con trol actions. Due to the high-dimensional and dynamic nature of uid antenna netw orks, con ven tional rule-based or static optimization methods are insucien t to capture the complex interactions b etw een antenna congurations Submitted to IEEE Wireless Communications Magazine, 2026 Fig. 2: AI-nativ e control framework for uid antenna net works, illustrating how adaptiv e antenna congurations and radio resource management decisions are join tly optimized through a closed-lo op in teraction betw een the net w ork en vironmen t and learning-based control mechanisms. and net work b ehavior. Learning-based approac hes there- fore pro vide a natural solution, with reinforcement learn- ing oering an eective framework for sequential decision- making under dynamic conditions. By interacting with the netw ork environmen t and ob- serving p erformance outcomes, the decision engine can progressiv ely learn p olicies that asso ciate netw ork states with eective antenna and resource congurations. In m ulti-cell deploymen ts, this extends naturally to dis- tributed decision-making, where eac h base station op erates as an indep endent learning agent that adapts to lo cal observ ations while implicitly co ordinating through the shared wireless en vironment. C. Control A ctions Based on the con trol engine’s decisions, the system applies actions that join tly adapt antenna congurations and radio resources. These include b oth ph ysical an tenna adjustmen ts and conv en tional radio resource management op erations. A base station ma y rep osition its an tenna, adjust beam- forming directions, or select alternativ e radiation patterns, while simultaneously determining user scheduling and transmit-p o w er allo cation. These actions are inherently coupled: antenna rep ositioning aects whic h users can b e served eciently , while sc heduling decisions inuence the optimal an tenna conguration. T reating them inde- p enden tly leads to sub optimal p erformance, whereas joint adaptation enables more ecien t netw ork op eration. D. Closed-Lo op Control Model The interaction b etw een state, decision, and action forms a closed-lo op control system. After actions are applied, the resulting changes in c hannel conditions, inter- ference patterns, and user exp erience are reected in the up dated netw ork state, pro viding feedback for subsequent decisions. This feedbac k enables the system to contin u- ously rene its b eha vior and adapt to evolving netw ork conditions, including user mobility , trac uctuations, and dynamic in terference. Suc h a closed-loop architecture is particularly imp ortant in uid antenna netw orks, where the impact of decisions is highly context-dependent. The com bination of state-aw are decision-making and con tin- uous feedback enables scalable, adaptive control across large, dynamic wireless deplo yments. The next section builds on this arc hitecture and illus- trates its realization through a m ulti-agent reinforcement learning framew ork in multi-cell netw ork scenarios. V. Multi-Agent Reinforcemen t Learning Approach for Fluid An tenna Con trol The control arc hitecture describ ed in the previous sec- tion provides a general framew ork for integrating antenna adaptation in to net w ork decision pro cesses. The remaining c hallenge is how the decision engine can learn to select appropriate antenna and resource congurations under dynamic netw ork conditions. T o address this, we consider a multi-agen t reinforcement learning (MARL) approac h as a representativ e solution for distributed and adaptive con trol in uid antenna netw orks. Reinforcemen t learning is w ell-suited to such settings, in which decisions are made sequen tially under dynamic, partially observ able conditions. Each agen t observes its lo cal state s b ( t ) , selects an action a b ( t ) , and ev aluates the outcome through a reward signal r b ( t ) , where b denotes the base station (agen t) index. Ov er time, this Submitted to IEEE Wireless Communications Magazine, 2026 in teraction enables the system to learn policies π b that map net w ork conditions to eective control decisions, consisten t with the state–decision–action–feedbac k lo op in tro duced earlier. A. Learning F ramework In uid antenna netw orks, eac h base station acts as an indep endent learning agent b that adapts its antenna conguration and transmission parameters. A t each de- cision step t , the agent observ es its lo cal state s b ( t ) , including channel conditions, user demand, interference lev els, and antenna conguration. Based on this state, it selects an action a b ( t ) according to its policy π b , jointly determining an tenna rep ositioning and high-level radio resource allocation. The net work then evolv es to a new state s b ( t + 1 ) and pro duces a reward r b ( t ) reecting throughput, fairness, and in terference mitigation. Through rep eated interaction, eac h agent renes its policy π b under dynamic conditions. F or intuition, a typical state s b ( t ) may include c hannel qualit y indicators (e.g., RSRP/SINR), trac demand, and in ter-cell interference levels. An action a b ( t ) may corresp ond to selecting a new antenna p osition along with sc heduling and p ow er allo cation decisions. The reward r b ( t ) can b e dened as a weigh ted function of user rates and fairness (e.g., prop ortional fairness), penalized b y excessive in terference. This form ulation enables joint optimization of throughput, fairness, and interference. B. Distributed Multi-Cell Learning In practical deploymen ts, neigh b oring cells are coupled through interference and shared sp ectrum, making cen- tralized optimization dicult to scale. MARL pro vides a scalable alternative, where each base station up dates its p olicy π b based on lo cal transitions ( s b , a b , r b , s ′ b ) . Although decisions are made lo cally , agents interact through the shared wireless en vironmen t, enabling implicit co ordina- tion. F or example, impro ving the signal quality of a local user may alter interference exp erienced by neigh b oring cells. Through contin uous interaction, agents learn to balance lo cal p erformance with netw ork-wide interference dynamics, reducing the need for explicit co ordination. C. Learning W orkow and Algorithm The learning pro cess follows an iterative closed-lo op in teraction b etw een agen ts and the netw ork e n vironmen t, as summarized in Algorithm 1. At eac h iteration, each agen t observes s b ( t ) , selects a b ( t ) according to π b , and receiv es r b ( t ) and s b ( t + 1 ) . This feedback captures the impact of an tenna and resource decisions on netw ork per- formance, enabling the agent to ev aluate the eectiv eness of its actions under evolving netw ork conditions. Over time, agents learn p olicies that adapt to v ariations in mobilit y , trac, and in terference. Learning operates at the net w ork control lay er, while b eamforming is performed at the ph ysical lay er based on the selected antenna cong- uration, thereby preserving a hierarc hical structure. This can b e interpreted as a tw o-timescale pro cess: antenna rep ositioning adapts to slow er channel and interference trends, while b eamforming and scheduling op erate at faster timescales, allowing the system to balance long-term spatial adaptation with short-term c hannel tracking. T o illustrate, consider a single step. A t time t , a base station observes a state in which a cell-edge user exp eriences low SINR due to strong interference. The agen t selects an action that repositions the antenna a w a y from the dominant interference direction while sc heduling the user with adjusted transmission parameters. This reduces interference and improv es SINR, thereb y yielding a higher reward that reects increased throughput and fairness. Repeating this pro cess enables the agen t to pro- gressiv ely learn in terference-aw are antenna congurations that generalize across similar net work conditions. The next section ev aluates the eectiveness of this learning-based control approach through represen tative sim ulation results. Algorithm 1 MARL for Fluid Antenna Control 1: Initialize p olicy π b for each base station b 2: Initialize netw ork en vironment 3: for each time step t do 4: for each base station b do 5: Observ e lo cal state s b ( t ) 6: Select action a b ( t ) using policy π b 7: end for 8: Apply join t actions (an tenna conguration and high-lev el resource allo cation) 9: En vironment evolv es (channel and in terference update) 10: for each base station b do 11: Receiv e reward r b ( t ) 12: Observ e next state s b ( t + 1 ) 13: Up date p olicy π b using ( s b ( t ) , a b ( t ) , r b ( t ) , s b ( t + 1 )) 14: end for 15: end for VI. Illustrative P erformance Ev aluation T o illustrate the b enets of intelligen t uid antenna con trol, we consider a light w eigh t multi-cell simulation capturing user mobility , inter-cell in terference, and an- tenna reconguration. The ob jective is not exhaustiv e b enc hmarking, but to highlight how learning-based an- tenna control inuences p erformance under represen tativ e 6G conditions. A. Simulation Scenario W e consider a seven-cell hexagonal net work in whic h eac h base station employs a uid antenna that can rep osi- tion within a b ounded region. The channel mo del includes large-scale path loss and small-scale fading, while in ter-cell in terference arises from concurrent transmissions. Users are randomly distributed with mo derate mobility , lead- ing to time-v arying c hannel and in terference conditions. A t each decision interv al, base stations adapt antenna congurations and transmission strategies according to the selected control p olicy . The learning-based controller Submitted to IEEE Wireless Communications Magazine, 2026 20 40 60 80 100 User Density (a) 0.7 0.8 0.9 1 1.1 1.2 Normalized Throughput FAB SDAR RAS Proposed 20 40 60 80 100 User Density (b) 0.1 0.2 0.3 0.4 0.5 0.6 Cell-edge Throughput FAB SDAR RAS Proposed 20 40 60 80 100 User Density (c) 1.5 2 2.5 3 Spectral Efficiency (bps/Hz) FAB SDAR RAS Proposed Fig. 3: P erformance comparison under increasing user density . (a) Aggregate netw ork throughput v ersus user densit y . (b) Cell-edge throughput (5th percentile user rate), highligh ting p erformance for vulnerable users. (c) Sp ectral eciency . The learning-based uid antenna control yields mo derate gains in aggregate throughput but signicantly impro v es cell-edge p erformance under dense, interference-limited conditions. FAB SDAR RAS Proposed (a) 0 0.2 0.4 0.6 0.8 1 Jain's Index FAB SDAR RAS Proposed (b) 0 0.2 0.4 0.6 0.8 1 1.2 Inter-cell Interference Power 0 0.2 0.4 0.6 0.8 1 Throughput (c) 0 0.2 0.4 0.6 0.8 1 CDF FAB SDAR RAS Proposed Fig. 4: User-lev el and in terference-related performance comparison across con trol strategies. (a) Jain’s fairness index. (b) A verage inter-cell interference p ow er. (c) Cumulativ e distribution function (CDF) of user throughput across the net w ork. The prop osed metho d impro v es fairness while reducing in terference, resulting in a more balanced throughput distribution, particularly b eneting users in the low er-p erformance regime. observ es local state information (c hannel quality , interfer- ence, and trac demand) and selects joint actions com- prising an tenna p ositioning and high-level transmission decisions. The con trol architecture is hierarchical. The learning agen t operates at the net work control lay er, deter- mining an tenna congurations and high-lev el transmission parameters suc h as user selection and p ow er allo cation. Giv en the selected an tenna state, b eamforming w eights are computed by a conv entional ph ysical-la y er mo dule. This separation ensures that b eamforming op erates on top of a dynamically recongurable antenna structure, rather than replacing it. In this sense, uid an tenna control complemen ts b eamforming b y introducing an additional spatial con trol dimension. F or comparison, w e consider three baseline strategies: Fixed Antenna + Con ven tional Beamforming (F AB), whic h employs a static an tenna with standard b eam- forming; Signal-Driv en Antenna Rep ositioning (SDAR), whic h heuristically adapts an tenna p osition based on instan taneous signal strength; and Randomized Antenna Selection (RAS), whic h selects antenna states randomly without co ordination with transmission strategies. Perfor- mance is ev aluated using aggregate throughput, cell-edge throughput (5th p ercentile), sp ectral eciency , fairness (Jain’s index), inter-cell in terference, and user throughput distribution. B. Performance Results Fig. 3a shows aggregate throughput versus user den- sit y . The prop osed metho d ac hieves consistent improv e- men ts ov er baseline approaches, with gains b ecoming more pronounced in dense scenarios where interference coupling dominates system behavior. More imp ortantly , Fig. 3b highlights cell-edge throughput. The learning- based control provides substantial improv ements for users near cell b oundaries, where interference is strongest and con v entional b eamforming alone is insucien t. Notably , the gain at the cell edge is signicantly greater than the av erage improv ement, reac hing up to 50–70% higher throughput than baseline strategies under dense user conditions. This indicates that the primary b enet of uid an tenna con trol lies in protecting vulnerable users rather than only impro ving av erage p erformance. Fig. 3c presen ts sp ectral eciency , showing that join t adaptation of an tenna p osition and transmission strategy enables more eectiv e spatial reuse and in terference mitigation. A dditional insigh ts are pro vided in Fig. 4. The pro- p osed approach ac hieves higher fairness (Fig. 4a) while sim ultaneously reducing in ter-cell interference (Fig. 4b), indicating that p erformance gains are not obtained at the exp ense of user im balance. The throughput CDF in Fig. 4c further conrms that the impro vemen t is distributed across users, with notable gains in the low er tail of the Submitted to IEEE Wireless Communications Magazine, 2026 T ABLE I: Practical Challenges and Corresp onding Research Directions for Fluid An tenna Netw orks Challenge Key Issue Research Direction Control Overhead F requent antenna up dates increase con trol-plane signaling without prop ortional p erformance gains Hierarchical control architectures that decouple fast scheduling decisions from slow er antenna adaptation Hardware Reconguration Latency Physical antenna mo vemen t introduces dela ys that may exceed channel coherence time Latency-aw are con trol p olicies aligned with hard- ware constraints and time-scale separation Multi-Cell Co ordination Antenna adaptation decisions are inherently cou- pled across cells through interference Distributed coordination mechanisms and interference-a ware m ulti-agent learning Energy Consumption F requent antenna reconguration increases the operational energy footprint Energy-aw are con trol strategies that balance p er- formance gains with reconguration cost Integration with F uture RAN Architectures Antenna control requires tigh t in tegration with AI- native and softw are-dened netw ork frameworks Seamless in tegration with AI-nativ e RAN arc hi- tectures through standardized control in terfaces distribution. The observed gains are primarily driven b y the learn- ing agent’s abilit y to av oid in terference-sensitiv e antenna congurations. In dense deploymen ts, the controller learns to rep osition the antenna aw ay from spatial regions that create persistent inter-cell in terference while maintain- ing sucient signal strength for the intended user. F or example, in cell-edge conditions, the controller shifts the antenna aw ay from directions that strongly couple with neigh b oring-cell users, thereb y reducing cross-cell in terference while preserving the desired link quality . This leads to a signicant improv ement in eectiv e SINR for edge users, directly translating in to the observed gains in cell-edge throughput. Ov erall, this b ehavior reects a dynamic trade-o b etw een interference suppression and signal preserv ation that cannot b e achiev ed with xed- an tenna b eamforming alone. These results demonstrate that uid antenna control in tro duces a new degree of freedom that complements con v entional b eamforming, enabling more adaptive and in terference-a ware netw ork op eration in dynamic environ- men ts. VI I. Practical Implementation Challenges While F AS introduce new capabilities for wireless net- w orks, bringing these ideas into practical deploymen ts presen ts several important challenges. Unlik e purely al- gorithmic solutions, uid antenna con trol requires tight in teraction b etw een hardware reconguration, net work dy- namics, and intelligen t decision mechanisms. Understand- ing and addressing these aspects is essen tial for in tegrating uid an tenna systems into future 6G infrastructures. A. Control Ov erhead A k ey challenge is the signaling ov erhead induced b y an tenna adaptation, whic h relies on contin uous collec- tion of channel measurements, interference indicators, and user demand. Excessively frequent rep ositioning can burden the control plane without yielding prop ortional p erformance gains, esp ecially under slowly v arying chan- nels. This in tro duces a fundamental trade-o b etw een resp onsiv eness and eciency . Consequen tly , selecting ap- propriate control interv als becomes a system-level design problem, motiv ating hierarchical strategies that decouple fast resource allo cation from slow er antenna adaptation. B. Hardware Reconguration Latency A practical limitation arises from the non-instan taneous reconguration of uid antennas. Mec hanical or micro- electromec hanical implementations introduce delays b e- t w een control decisions and their execution, directly im- pacting the con trol loop. When adaptation is slo w er than c hannel dynamics, p erformance gains degrade. Therefore, con trol algorithms must explicitly account for hardw are resp onse times and av oid ov erly aggressive p olicies, un- derscoring the need for latency-a ware design. C. Multi-Cell Co ordination In m ulti-cell scenarios, an tenna congurations are in- trinsically coupled through interference. Rep ositioning that b enets one user may degrade performance in neigh b oring cells, leading to complex co ordination c hal- lenges. F ully cen tralized solutions incur high signaling o v erhead and require global kno wledge, while purely lo cal approac hes risk instability and sub optimality . Scalable op eration th us necessitates distributed learning and co- ordination mec hanisms for stable uid antenna control. D. Energy Consumption Dynamic antenna rep ositioning also introduces addi- tional energy costs. Dep ending on the implementation, mec hanical actuation or electronic switching can lead to non-negligible consumption, particularly under frequent adaptation. Energy-aw are control p olicies are therefore required to ensure that p erformance impro v emen ts justify op erational costs, a concern that b ecomes more pro- nounced in dense deplo yments. E. Integration with F uture RAN Architectures Fluid an tenna con trol m ust ultimately align with emerg- ing RAN paradigms. As systems ev olve to ward softw are- dened and AI-nativ e architectures, such as op en RAN and intelligen t controllers, an tenna adaptation can b e em b edded within unied optimization frameworks that join tly address resource allo cation, mobility , and inter- ference. Dening appropriate interfaces for seamless in- teraction with these comp onents is essential for practical deplo ymen t. Submitted to IEEE Wireless Communications Magazine, 2026 VI I I. Op en Research Directions Fluid an tenna netw orks op en a wide range of researc h opp ortunities across wireless system design, intelligen t con trol, and adaptive radio hardware. While learning- driv en antenna adaptation sho ws strong p otential, several c hallenges m ust be addressed before large-scale 6G deploy- men t becomes feasible. This section outlines k ey directions building on the practical issues discussed earlier. A. Hierarchical An tenna Control A promising direction is hierarchical con trol, where adaptation op erates across m ultiple timescales. An tenna rep ositioning t ypically evolv es slow er than scheduling or b eamforming, enabling a structure that combines fast lo cal decisions with slow er netw ork-level co ordination. Suc h designs impro v e stability and responsiveness while reducing signaling o verhead, directly addressing control o v erhead limitations. B. Joint Con trol of Programmable Radio Environmen ts F uture netw orks will include m ultiple programmable elemen ts, such as uid antennas and recongurable in- telligen t surfaces. Joint optimization of an tenna position- ing and environmen tal con trol signicantly expands the design space but also increases co ordination complexity . Ecien t mechanisms for jointly managing these compo- nen ts remain an op en challenge. C. Learning Under Limited Observ ability Learning-based control dep ends on accurate observ a- tions, yet practical systems op erate with partial, noisy , and dela yed information. Designing algorithms robust to limited observ ability is therefore critical. Com bining lo cal measuremen ts with prediction or co op erative information sharing oers a promising path forward. D. Scalable Multi-Agent Coordination Large-scale uid an tenna netw orks require co ordination among many distributed agents. Indep endent decisions ma y lead to instability or ineciency due to interference coupling. Scalable solutions based on distributed learning, co op erativ e p olicies, and in terference-aw are coordination are essen tial for stable op eration. E. Integration with AI-Nativ e Netw ork Architectures Fluid antenna control must align with emerging AI- nativ e RAN arc hitectures, where data-driven models and in telligen t control op erate across la yers. This requires standardized interfaces, ecient data pipelines, and ex- ible framew orks that integrate antenna adaptation with sc heduling, mobility managemen t, and orchestration func- tions. Understanding this interaction remains a key re- searc h direction. The relationship betw een these c hallenges and their cor- resp onding researc h directions is summarized in T able I. IX. Conclusion Fluid antenna systems introduce a new degree of free- dom by enabling dynamic adaptation of antenna cong- urations to changing radio conditions. Beyond physical- la y er gains, this work highlights their impact at the net w ork level, where an tenna conguration b ecomes part of the con trol problem. W e presented an AI-nativ e p ersp ective in which an- tenna adaptation is embedded into a closed-lo op frame- w ork alongside con ven tional radio resource management. Within this setting, learning-based approac hes, partic- ularly m ulti-agent reinforcemen t learning, enable dis- tributed and adaptive decision-making across multi-cell en vironmen ts. The results sho w that in telligent antenna rep ositioning improv es throughput, enhances cell-edge p erformance, and mitigates inter-cell in terference com- pared to static op eration. Lo oking ahead, uid antenna netw orks signal a shift in wireless system design, in whic h antennas ev olv e from passiv e in terfaces to con trollable net work elements. Real- izing this vision requires adv ances in adaptiv e hardware, scalable learning, and integrated control arc hitectures within AI-nativ e RANs. Therefore, uid antennas should b e viewed not only as a hardware capabilit y , but as a key enabler of more adaptiv e, ecien t, and intelligen t 6G netw orks. References [1] S. Kurma, K. Singh, A.-A. A. Boulogeorgos, T. A. T siftsis, and C.-P . Li, “Activ e ris-aided full-duplex cell-free massive mimo for mobile wireless netw orks of limited capacit y ,” IEEE T ransactions on V ehicular T echnology , vol. 75, no. 3, pp. 4110– 4128, 2026. [2] H. Hong, K.-K. W ong, C.-B. Chae, H. Xu, X. Guo, F. R. Ghadi, Y. Chen, Y. Xu, B. Liu, K.-F. T ong, and Y. Zhang, “A contemporary survey on uid antenna systems: F undamentals and netw orking persp ectives,” IEEE T ransactions on Net work Science and Engineering, vol. 13, pp. 2305–2328, 2026. [3] T. L. Marzetta, “Nonco op erative cellular wireless with unlimited num b ers of base station antennas,” IEEE T ransactions on Wireless Communications, vol. 9, no. 11, pp. 3590–3600, 2010. [4] E. Basar, M. Di Renzo, J. De Rosny , M. Debbah, M.-S. Alouini, and R. Zhang, “Wireless communications through recongurable in telligent surfaces,” IEEE Access, vol. 7, pp. 116 753–116 773, 2019. [5] Q. W u and R. Zhang, “T ow ards smart and recongurable envi- ronment: Intelligen t reecting surface aided wireless net w ork,” IEEE Communications Magazine, v ol. 58, no. 1, pp. 106–112, 2020. [6] C. W ang, G. Li, H. Zhang, K.-K. W ong, Z. Li, D. W. K. Ng, and C.-B. Chae, “Fluid antenna system lib erating multiuser mimo for isac via deep reinforcement learning,” IEEE T ransactions on Wireless Communications, vol. 23, no. 9, pp. 10 879–10 894, 2024. [7] R. Xu, Z. Y ang, Z. Zhang, M. Shikh-Bahaei, K. Huang, and D. Niyato, “Energy ecient uid antenna rela y (far)-assisted wireless communications,” IEEE Journal on Selected Areas in Communications, vol. 44, pp. 1160–1176, 2026. [8] J. Chen, Y. Xiao, Z. Peng, J. Zhu, X. Lei, C. Masouros, and K.- K. W ong, “Hybrid beamforming for ris-assisted m ultiuser uid antenna systems,” IEEE T ransactions on Wireless Communica- tions, vol. 25, pp. 2718–2732, 2026. [9] H. Xiao, X. Hu, K.-K. W ong, H. Hong, G. C. Alexandrop oulos, and C.-B. Chae, “Fluid recongurable intelligen t surfaces: Joint on-o selection and beamforming with discrete phase shifts,” IEEE Wireless Communications Letters, vol. 14, no. 10, pp. 3124–3128, 2025. Submitted to IEEE Wireless Communications Magazine, 2026 [10] N. C. Luong, D. T. Hoang, S. Gong, D. Niyato, P . W ang, Y.- C. Liang, and D. I. Kim, “Applications of deep reinforcemen t learning in comm unications and netw orking: A surv ey ,” IEEE Communications Surveys & T utorials, vol. 21, no. 4, pp. 3133– 3174, 2019. [11] K. Zhang, Z. Y ang, and T. Başar, Multi-Agen t Reinforcement Learning: A Selectiv e Ov erview of Theories and Algorithms. Cham: Springer International Publishing, 2021, pp. 321–384. [Online]. A v ailable: https://doi.org/10.1007/978- 3- 030- 60990- 0_12 [12] U. Challita, L. Dong, and W. Saad, “Proactive resource manage- ment for lte in unlicensed sp ectrum: A deep learning p ersp ec- tive,” IEEE T ransactions on Wireless Communications, v ol. 17, no. 7, pp. 4674–4689, 2018. Ian F. Akyildiz (Life F ellow, IEEE) received his B.S., M.S., and Ph.D. degrees in electrical and computer engineering from the Univer- sity of Erlangen–Nürnberg, Germany , in 1978, 1981, and 1984, respectively . F rom 1985 to 2020, he held the Ken Byers Chair Professor- ship at Georgia T ech, where he directed the Broadband Wireless Netw orking Laboratory . A visionary entrepreneur, he is the President of T ruv a Inc. and a key advisor to global institutions lik e TI I (Abu Dhabi) and Odine Labs (Istanbul). Since 2020, he has served as the founding Editor-in- Chief of the ITU Journal on F uture and Evolving T echnologies. His pioneering research spans 6G/7G systems, molecular comm unication, terahertz technology , and underw ater netw orking. As of Marc h 2026, he holds an H-index of 146 with ov er 155,000 citations. Dr. Akyildiz is an ACM F ellow and recipient of prestigious honors, including the Humboldt (Germany) and TÜBİT AK (Türkiye) A wards. T uğçe Bilen (Member, IEEE) received her B.Sc., M.Sc., and Ph.D. degrees in Computer Engineering from Istan bul T echnical Univ er- sity (ITU) in 2015, 2017, and 2022, resp ec- tively . She is currently an Assistant Professor in the Department of Articial In telligence and Data Engineering at ITU, where she previously served as a Researc h and T eaching Assistant. Her do ctoral research earned several prestigious honors, including the 2025 IEEE T urkey Section Ph.D. Thesis A ward, the 2023 T urkish Academ y of Sciences (TÜBA) First Prize in Science and T echnology , the 2023 Serhat Özy ar Y oung Scien tist Honorary A ward, and the 2022 ITU Best Ph.D. Thesis A ward. Her research fo cuses on 6G netw orks, Knowledge-Dened Networking (KDN), AI-driv en netw ork man agemen t, and digital twins, sp ecically integrating intelligen t systems in to future architectures. Dr. Bilen also serves as a reviewer for leading international journals.