ASNO-LLM: Large Language Model based Algorithm Selection Framework for 6G Network Service Optimization & Automation

ASNO-LLM: Large Language Model based Algorithm Selection Framework for 6G Network Service Optimization & Automation Anestis Dalgkitsis, Cyril Shih-Huan Hsu, Chrysa Papagianni, and Paola Grosso University of Amsterdam, LAB42, Science Park 900, 1098 XH Amsterdam, The Netherlands Email: {a.dalgkitsis, s.h.hsu, c.papagianni, p.grosso}@uva.nl Abstract—The integration of cloud computing into mobile networks has unlocked new possibilities through a unified and abstract infrastructure. However, this convergence introduced significant complexity, emphasizing the critical need for innovative service management and orchestration frameworks. Despite considerable efforts towards integrating automation to optimize network services, it remains clear that no universal algorithm can achieve optimal performance under all spatial and temporal conditions. To address this challenge, we introduce ASNO-LLM, a novel Algorithm Selection framework for Network Orchestration, designed specifically to address the complexity and automation challenges of next-generation 6G networks. ASNO-LLM serves as a platform capable of managing and dynamically selecting the most suitable optimization algorithm based on real-time performance metrics and user-defined preferences. Index Terms—Algorithm Selection, Service Optimization, Network Automation, Zero-touch Orchestration, Large Language Models. I. INTRODUCTION As the telecommunications landscape evolves towards the 6G era, efficient and autonomous network orchestration emerges as a foundational requirement to meet the evergrowing demands for flexibility, scalability, and performance in next-generation services. Yet, the complete automation of such multi-dimensional, dynamic systems remains a significant challenge. General-purpose optimization models quickly reach their limits, as no single algorithm can consistently deliver optimal results across all scenarios [1]. This phenomenon is widely recognized as performance complementarity. This limitation has a particularly strong presence in mobile and heterogeneous network environments, as the optimization environment and context varies a lot between policies. This is strongly reflected in the emergent spark of interest for automated algorithm selection techniques to identify which optimization algorithm should be used for automation in recent literature. Existing selection mechanisms typically rely on two main information sources: (i) problem instance features and (ii) the historical performance of algorithms. However, these methods approach selection with a ”black box” mentality, resulting in limited adaptability, high complexity, or degraded performance, that is particularly problematic in the lowlatency, high-reliability environments envisioned for 6G. The emergence of Large Language Models (LLMs) offers a promising new direction. Unlike traditional algorithm selection techniques, LLMs possess the unique ability to process and correlate structured and unstructured inputs, ranging from service requests and metadata to user intents and optimization goals. Recent advances in instruction following, chain-ofthought prompting, and functional reasoning allow LLMs to interpret code semantics and context-rich text with a level of abstraction and generalization previously unattainable. In this paper, we explore the integration of LLMs into the algorithm selection process for 6G network optimization and demonstrate its efficacy in real-world scenarios. AI and ML automation play a key role in managing network complexity by optimally allocating resources and ensuring seamless service continuity at any given time [2]. As a result, effective orchestration frameworks capable of perfectly curating the AI/ML automation based on the service KPIs play a major role in unlocking the full potential of the 6G network. Motivated by this potential, we introduce ASNO-LLM, a novel LLM-based Algorithm Selection framework for Network Orchestration. It dynamically selects the most appropriate optimization approach from a predefined set, based on the type of optimization required by the incoming service request, real-time infrastructure performance metrics and business-defined preferences provided as a combined textual prompt. The objective of our work is to create the basis for a universal network service optimization engine for the 6G architecture, that is able to employ the best suited optimization algorithm for any incoming service request. Our contributions are threefold by: •Allowing orchestrators to perform multi-model service optimization seamlessly, both pre-and-post-deployment. •Leveraging LLMs to select the optimal model from a set based on textual descriptions, KPIs, and Service Level Agreements (SLAs). •Proposing an algorithm selection module architecture, based on the SMO framework of the Desire6G European project [3], [4]. This work is part of the Desire6G Service Management and Orchestration (SMO) framework, called the Optimization Engine (OE) module. The Desire6G is a European Horizon project focused on proposing a new architecture for mobile networks through deep programmability, distributed intelligence, and intent-based orchestration for real-time, secure, and adaptive 6G services [3], [4]. The remainder of this paper is organized as follows: Section II presents a literature review of related optimization frameworks. Section III describes the problem formulation and system model. Section IV provides an in-depth analysis of the proposed solution architecture. Section V explains the experimental setup and evaluates the effectiveness of the proposed approach. Finally, Section VI concludes this work and discusses the future research directions. II. RELATED WORK Algorithm selection is not a new concept, it was first formalized by John R. Rice in 1976 [5] as the problem of selecting the most suitable algorithm for a given problem instance from a set of available algorithms [1]. In this context, several studies have explored the application of algorithm selection throughout the years. Specifically, traditional approaches include [6] where the authors use the Ridge Regression method to predict the runtime of an algorithm set on specific problem instances. Their solution computes a feature vector based on characteristics of the problem instance. Similarly, in [7] the authors devise a ranking strategy to determine the suitability of selection for a given problem instance. They designed a multi-criteria ranking based on multiple evaluation measures simultaneously, by using a Pareto-efficient ranking strategy based on a calculated performance matrix for each algorithm metric. Authors in [8], [9] propose similarity-based approaches to algorithm selection, relying on clustering and instance features. The latter work represents problem instances by constructing a feature vector that derives from static and dynamic features to quantify similarities. Algorithm Selection methods can be categorized into two distinct classes according to their feature representation [10]: Feature-free methods extract the problem instance representations themselves, whereas feature-based methods rely on manually calculated feature profiles [1]. Regarding featurebased methods, the authors in [11] emphasize the importance of algorithm feature representation in AS and argue that considering only problem instance features, can lead to performance degradation. It is important to note that the latest algorithm selection research, is heavily leaning in the direction of LLMs, as they are a great tool for feature extraction and basic reasoning abilities in from text. Specifically, in [1] Xingyu Wu et al., propose a novel algorithm selection framework, titled AS-LLM, that integrates algorithm features extracted by pretrained LLMs. They demonstrate that including algorithm representations significantly enhances selection accuracy compared to traditional methods relying only on problem features, especially when training data is limited. To evaluate algorithm selection in the context of 6G Service Management and Orchestration (SMO), we focus on the scenario of service partitioning in multi-site topologies. We follow the work of [12] to design a baseline decentralized network service partitioning scheme using cooperative deep multi-agent reinforcement learning, which has been shown to achieve faster convergence, improved partitioning optimality, and better scalability across different network topologies. Next, we perform a similar use of distributed reinforcement learning as the authors of [13], who demonstrated that a decentralized service function chain orchestration approach can significantly reduce service latency by 60.54% compared to centralized solutions, highlighting the importance of network service partitioning and embedding for latency-sensitive applications. In this work we combine the scenario of network service partitioning and algorithm selection, as we intend to study the use of LLMs in the context of 6G networking SMO. III. PROBLEM FORMULATION & SYSTEM MODEL In this section, we define the problem of algorithm selection within the context of 6G network service orchestration and formalize the main concepts by formulating the system model. A. Generic Service Optimization Problem Description During operation, multiple service requests arrive at the orchestrator with specific requirements and constraints, described in a structured textual form as a Network Service Description (NSD). In the case of optimization, we assume that a selection must be made from a predefined set of optimization algorithms or models, each one targeting different objectives, described through metadata. Instead of relying on fixed heuristics, keyword matching or traditional techniques that are unable to take into account the context of the optimization, we utilize an LLM that receives a prompt combining the request, algorithm descriptions, and business operational objectives. The most appropriate algorithm must then be selected from a pool to execute the optimization of the service request. This way one SMO component can act as a generic-purpose optimization framework, overcoming the single optimization algorithm limitations [1]. The selected algorithm then processes the request in a way that the optimization outcome reflects the intent expressed in the prompt. B. Algorithm Selection Problem Definition A set of optimization algorithms A, called the Model Pool (MP) can be defined as follows: A={A1, A2, . . . , An}.(1) Hence, the optimal algorithm selection for a given request can be expressed as: A∗∈ A.(2) We assume that for every optimization algorithm Anin the MP Athere is a structured textual description Dnof the capabilities and optimization objectives of each algorithm An. We define the NSD as abstracted structured textual data, denoted as S. The NSD is considered as input to the algorithm selection system. We define accordingly an abstracted textual description of the current network state N, and a text-based business operational objective defined by the user O. Hence, the algorithm selection problem can then be modeled as the result of combined textual representations of the service requirements, network state, objective and available optimization algorithm descriptions, that an LLM model can use to perform the selection: fLLM :S, N, O, D1, D2, . . . , Dn→A∗,(3) where Dnis the structured textual description of Anincluding code, Na textual representation of the current network state, and Oa generic objective prompt, a manual set of instructions in text format to provide more context regarding the needs of the client, e.g. ”minimize service latency”. C. Textual LLM input representation LLMs internally represent textual data through embeddings. We define Eas the embedding function that maps text inputs into a high-dimensional vector space Rd: E:T→Rd,(4) where Tdenotes the set of the textual input, and dthe dimensionality of the embedding. Based on that, an abstracted network status description N embedding can be defined as: E(N)∈Rd,(5) the NSD textual description embeddings as: E(S)∈Rd,(6) the textual description embeddings of the business operational objective defined by the user as: E(O)∈Rd,(7) whereas the algorithm textual description embeddings as: E(Di)∈Rd.(8) The final combined embedding that creates the prompt can be expressed as follows: X= [E(S); E(N); E(O); E(D1); E(D2); . . . . . . ;E(Dn)] ∈R(n+1)d.(9) IV. DESIGN & IMPLEMENTATION OF AN LLM-BASED ALGORITHM SELECTION FRAMEWORK FOR 6G NETWORKS In this section, we are transforming the formalized algorithm selection process as a framework in the context of 6G networks. A. Framework Objective Description The main objective of the ASNO-LLM framework is to select the appropriate algorithm from a set to perform network service optimization. Upon receiving a network service descriptor from the SMO, it examines the annotated metadata of the service graph, such as the SLA, resource requirements and hardware constraints. It combines this information with an abstracted description of the current network conditions and business operational objectives. At the final step, it has to decide and evoke an optimization algorithm from a set, that will perform the required optimization in a way that respects the given requirements. Figure 1. Illustration of the modular ASNO-LLM architecture and internal sub-modules. B. LLM-Based Module Architecture The internal architecture of the ASNO-LLM framework is represented as an autonomous module, depicted in Fig. 1. C. Main Components The main components that comprise the internal framework architecture are the following: •The Model Life Cycle Management is responsible for the optimization objective extraction from the incoming network service request, the corresponding prompt generation based on all available information, and finally the LLM-based model selection fLLM. •The Model Pool encapsulates a set of optimization algorithms Ain form of high level programming scripts Anwith an extensive textual description Dnof their functionality and evaluated performance. D. 6G SMO-layer Architecture Interfaces In order to enable communication with the other 6G architecture components, as defined by [3], [4], we define additional external interfaces with the following generalized modules from the recent bibliography: •The Orchestrator, as defined in recent EU projects, serves as the Service Management and Orchestration (SMO) component. It interfaces with both the Infrastructure Management layer and the other sub-modules. •The Monitoring Subsystem is responsible for transmitting live monitoring data via the Monitoring Bus. This data can be used to reconstruct the current state of the network and infrastructure internally for further computations. E. Algorithm Selection for Network Service Partitioning A network service can be defined as a sequence of interconnected network functions, forming a Service Function Chain (SFC) that must comply with specific SLAs. In multisite topologies, the SFC needs to be partitioned into multiple sub-chains, each of which is forwarded to the respective local Figure 2. Service partitioning optimization experimental scenario workflow. The optimization objective is applied by partitioning the service between a number of domains in a way that it respects the SLA. network infrastructure managers for local embedding. Due to limited visibility across domains, only abstracted information from each site is available to the partitioning algorithms. The overall network service partitioning process is illustrated in Fig. 2. F. Decision Process & Selection Probabilities As a decision process, the posterior probability of selecting each algorithm from the MP, given the combined prompt input, is described as follows: P(Ai|S, N, O, D1, D2, . . . , Dn)for each Ai∈ A.(10) Effectively, the LLM prediction equates to the optimal algorithm selection A∗for a given request, expressed as follows: A∗= arg max Ai∈A P(Ai|S, N, O, D1, D2, . . . , Dn).(11) G. Optimization Problem Formulation We equate the problem of algorithm selection, as described in equation III-B, with the main objective of network service optimization. The optimization objective of Aican vary, based on a measurable performance metric, such as cost, latency, bandwidth or a similar metric. We define M(Ai, N)as a quantifiable way of how Ai optimizes the network service request N. Based on that, the optimal algorithm can be mathematically expressed as: A∗= arg min Ai∈A M(Ai, N)(12) The LLM cannot calculate directly the M(Ai, N), in that case it internally estimates it through semantic pattern matching and knowledge inference based on the embeddings. We express this approximation as follows: arg max Ai∈A P(Ai|N, D1, . . . , Dn, C)≈arg min Ai∈A M(Ai, N) (13) V. PERFORMANCE EVALUATION In this performance evaluation, we assess the performance of the proposed algorithm selection framework by comparing it against various baseline methods and multiple LLM configurations and models. A. Setup, Equipment & Evaluation Environment This setup features a simulated 3-domain linear network topology designed to handle multiple SFC requests originating from a simulated access network. Service request generation and LLM algorithm selection were performed locally on an Apple M3 Pro device for local inference using the latest version of Ollama [14]. Context window is set to the default value in all models. B. Services In order to evaluate our framework, we generate NSDs based on 6G operator services, which capture diverse optimization requirements. We define the service requirements as presented in the Table I. Table I SERVICE TYPES Service Type Latency (ms) Throughput (Mbps) Reliability eMBB 1–10 100–1000 99.9% URLLC 0.5–5 10–100 99.999% mMTC 10–100 1–10 99% V2X 1–5 50–200 99.99% AR/VR 5–20 200–500 99.9% C. Baselines Specifically, the following private local LLM models were used for algorithm selection in our evaluation: •Llama3.2:1b: Lightweight local model to test selection under tight compute/latency and strict data-privacy constraints. •Llama3.2:3b: Stronger on-device baseline to study sensitivity to model size while keeping reproducibility and security. •qwen3:4b: Compact alternative balancing efficiency for mid-scale tasks. •qwen3:8b: Larger variant to examine scalability of reasoning depth. •gemma3:1b: Smallest release to study algorithm selection low-resource performance e.g. at the edge, with strong alignment and safety. •gemma3:4b: Medium-scale release to test trade-offs between efficiency and reasoning capacity. Similarly, the following online 3rd party LLM models were used for comparison: •GPT-4o: High-capability reference to serve as an approximate upper bound. •GPT-4o-mini: Cost/latency–efficient 3rd party option to evaluate budget-aware selection and trade-offs. •GPT-3.5-turbo: Widely used baseline for legacy comparisons quantify progress over prior 3rd party accessible models. Regarding the partitioning methods, the following were utilized from the related works section: D. Scenario & Experimental Results We begin by presenting the evaluation scenario of 100 iterations. Figure 3. (a) Partitioning admission success rate per service type. (b) Service Type Distribution. Fig. 3a presents the admission success rate per service type. Specifically, it shows the rate of successfully deployed partitioned services in all domains, without violating the SLA. Success rate varies by service type, we can see that V2X services have the lowest success rate as they require extremely low latency and most partition attempts violate the service SLA. Similarly, URLLC and AR/VR had a lower acceptance rate, approximately 80%, due to strict SLA requirements. On the other hand, 100% of eMBB and mMTC were partitioned and achieved their optimization and deployment goals successfully. Fig. 3b shows the service type distributions that were used in this work. E. LLM-based Algorithm Selection Performance Evaluation The following part of this section is dedicated to the performance analysis of LLM-based algorithm selection, comparing the baseline models and their ability to match the optimization context to the best method of the model pool to perform the optimization. We include both private local models running on consumer hardware and popular 3rd party services to cover both multiple use-cases for a diverse set of users. Different simulations of 100 iterations were executed for each model. First, Fig. 4a presents the partitioning admission success rate per selection LLM. Similar to Fig. 3a it compares the total number of successfully deployed partitioned services in all domains, without violating the SLA. It is clear that overall performance is high, as LLM models have the ability match patterns including the surrounding given context and environment of the problem, identifying the objective correctly compared to just keyword matching. Next, Fig. 4b shows the final SLA compliance of the successfully partitioned and deployed service per selection LLM. As we can see, smaller local models performed better at respecting the SLA compared to their larger counterparts. During this test, larger local models shown to overcomplicate their selection thought process and missing the main objective. In Fig. 4c, the final average E2E service latency after to the partitioning and deployment with optimization objective latency minimization shown per selection LLM. Similarly, Fig. 4d, the final average E2E service throughput after to the partitioning and deployment with optimization objective throughput maximization shown per selection LLM. Moreover, Fig. 4e compares the total LLM inference time between all local and online 3rd party models, to provide insightful information about usage in real world. For simple network settings, we can see that llama3.2:1b and gemma3:1b not only have a comparable performance in simple algorithm selection, but also have similar inference times in local hardware, running at a private setting compared to a 3rd part model, preserving that way the privacy of the operators and network providers. Finally, we combine the previous to provide a clear weightbased comparison between the selection methods for this algorithm selection study. The score is calculated as follows: P S = 0.3×Psuccess + 0.3×Scompliance + 0.2... ... ×1−L Lmax + 0.2×T Tmax ,(14) where Psuccess is the partitioning admission success rate, Scompliance is the SLA compliant partition, Lis the average service latency after deployment, Lmax is the maximum service latency, Tis the average service throughput after deployment, and Tmax is the maximum throughput of all simulation runs per algorithm. VI. CONCLUSIONS & FUTURE WORK In this paper we have introduced ASNO-LLM, an algorithm selection framework for 6G networks, that selects the most appropriate optimization method from a predefined set. The proposed framework leverages the semantic reasoning capabilities of LLMs to dynamically select the most suitable optimization algorithm, based on real-time network conditions, SLA requirements, and the NSD. The objective of our work is to create the basis for a universal network service optimization engine for the 6G architecture, that is able to employ the best suited optimization algorithm for any incoming service request. Future work will focus on enhancing the selection by incorporating closed-loop feedback, that evaluate the algorithm performance into the selection process and allow the framework to continuously learn and refine its decisions. Additionally, we plan to expand ASNO-LLM to support multiobjective optimization goals, to increase its applicability across optimization tasks for next-generation networks. ACKNOWLEDGMENT This work is supported by CIENA, the European Commission H2020 project DESIRE6G (101096466), and the Dutch National Growth Fund project FNS. REFERENCES [1] X. Wu, Y. Zhong, J. Wu, and K. Tan, “AS-LLM: When algorithm selection meets large language model,” 2024. [Online]. Available: https://openreview.net/forum?id=l7aD9VMQUq Figure 4. (a) Partitioning admission success rate per selection LLM, (b) SLA compliance per selection LLM, (c) Average E2E service latency posterior to the partitioning per selection LLM with optimization objective latency minimization, (d) Average E2E service throughput posterior to the partitioning per selection LLM with optimization objective throughput maximization, (e) Selection LLM total inference time, (f) Combined performance score. [2] C. Ssengonzi, O. P. Kogeda, and T. O. Olwal, “A survey of deep reinforcement learning application in 5g and beyond network slicing and virtualization,” Array, vol. 14, p. 100142, 2022. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S2590005622000133 [3] C. Papagianni et al., “Desire6g: Deep programmability & secure distributed intelligence for real-time e2e 6g networks,” European 6G Annual Journal (SNS), 2023. [Online]. Available: https://desire6g.eu/dissemination/publications/ [4] P. Szilagyi, P. Varga, A. Kovacs, C. Papagianni, and L. Toka, “Towards extreme network kpis with programmability in 6g,” in Proceedings of the 24th International Symposium on Theory, Algorithmic Foundations, and Protocol Design for Mobile Networks and Computing (MobiHoc ’23), 2023. [Online]. Available: https://dl.acm.org/doi/10.1145/3565287.3617610 [5] J. R. Rice, “The algorithm selection problem**this work was partially supported by the national science foundation through grant gp-32940x. this chapter was presented as the george e. forsythe memorial lecture at the computer science conference, february 19, 1975, washington, d. c.” ser. Advances in Computers, M. Rubinoff and M. C. Yovits, Eds. Elsevier, 1976, vol. 15, pp. 65–118. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0065245808605203 [6] L. Xu, F. Hutter, H. H. Hoos, and K. Leyton-Brown, “Satzilla: Portfolio-based algorithm selection for sat,” Journal of Artificial Intelligence Research, vol. 32, p. 565–606, Jul. 2008. [Online]. Available: http://dx.doi.org/10.1613/jair.2490 [7] T. Cunha, C. Soares, and A. C. de Carvalho, “A label ranking approach for selecting rankings of collaborative filtering algorithms,” in Proceedings of the 33rd Annual ACM Symposium on Applied Computing. ACM, 2018, pp. 1393–1395. [8] S. Kadioglu, Y. Malitsky, M. Sellmann, and K. Tierney, “ISAC – instance-specific algorithm configuration,” in Proceedings of the 19th European Conference on Artificial Intelligence (ECAI 2010). IOS Press, 2010, pp. 751–756. [9] R. Amadini, M. Gabbrielli, and J. Mauro, “SUNNY: a lazy portfolio approach for constraint solving,” Theory and Practice of Logic Programming, vol. 14, no. 4-5, pp. 509–524, 2014. [10] M. Alissa, K. Sim, and E. Hart, “Automated algorithm selection: from feature-based to feature-free approaches,” 2022. [Online]. Available: https://arxiv.org/abs/2203.13392 [11] A. Tornede, M. Wever, and E. H¨ ullermeier, “Extreme algorithm selection with dyadic feature representation,” in Discovery Science - 23rd International Conference, DS 2020, Thessaloniki, Greece, October 19–21, 2020, Proceedings, ser. Lecture Notes in Computer Science, vol. 12323. Springer, 2020, pp. 309–324. [Online]. Available: https://doi.org/10.1007/978-3-030-61527-7 21 [12] A. Pentelas, D. De Vleeschauwer, C.-Y. Chang, K. De Schepper, and P. Papadimitriou, “Deep multi-agent reinforcement learning with minimal cross-agent communication for sfc partitioning,” IEEE Access, vol. 11, pp. 40 384–40 398, 2023. [13] A. Dalgkitsis, L. Garrido, K. Ramantas, L. Alonso, and C. Verikoukis, “Schema: Service chain elastic management with distributed reinforcement learning,” in GLOBECOM 2021 - 2021 IEEE Global Communications Conference, 2021, pp. 1–6. [14] L. Hause and T. Safi, “ollamar: An r package for running large language models,” Journal of Open Source Software, vol. 10, no. 105, p. 7211, January 2025, oRCID: Lin Hause - 0000-00034590-7039; Tawab Safi - 0009-0000-5659-9890. [Online]. Available: https://joss.theoj.org/papers/10.21105/joss.07211