## Overview Ericsson, a leading telecommunications infrastructure provider, is addressing the challenge of deploying large language models in production telecom environments through their System Comprehension Lab. The research team, led by researcher Ali Mouk, is developing generative AI tools for both internal use and customer-facing applications. The core challenge they face is that while LLMs show promise for telecom infrastructure management, their inherent limitations—including hallucinations, lack of explainability, and unreliable logical reasoning—make them unsuitable for mission-critical telecom operations without significant architectural enhancements. The presentation outlines Ericsson's strategic positioning in the generative AI landscape. Rather than investing in pre-training foundation models from scratch (the domain of companies like Google, Meta, Anthropic, and OpenAI), Ericsson focuses on adapting existing pre-trained models to telecom-specific tasks and domains. This practical approach reflects a realistic assessment of resource allocation and leverages the broader AI community's investments in foundational research while focusing Ericsson's efforts on domain-specific customization and production deployment challenges. ## Technical Architecture and LLMOps Strategy Ericsson's approach to LLM deployment in production follows a three-phase adaptation strategy. The first phase involves selecting appropriate pre-trained models and adapting them to specific tasks such as question answering or code generation. This task-level adaptation is particularly important when building multi-agent frameworks, where different models may be optimized for different telecom-specific functions. The second phase focuses on domain adaptation using three primary techniques: prompt engineering, retrieval-augmented generation (RAG), and fine-tuning. The RAG implementation at Ericsson represents a sophisticated understanding of production deployment challenges. The team recognizes that while open-source solutions enable rapid prototyping of basic RAG systems, production-grade implementations face significant complexity. The "garbage in, garbage out" principle becomes critical when dealing with telecom documentation, where information spans multiple documents and requires understanding relationships across different sections. Ericsson deals with multimodal documents containing not just text but diagrams, figures, and meta-information embedded in document structure—such as PowerPoint presentations where the spatial arrangement of elements conveys semantic meaning about network flows and system architecture. Extracting and representing this meta-information for LLM consumption represents a significant LLMOps engineering challenge. Fine-tuning is applied selectively at Ericsson, typically using parameter-efficient methods like LoRA to adapt models to specific code repositories or telecom standards documentation. This reflects a pragmatic approach that recognizes the availability of pre-fine-tuned open-source models while acknowledging the need for final customization to proprietary Ericsson systems and documentation. ## Multimodal Challenges in Telecom The telecom domain presents unique multimodal challenges that extend far beyond the typical text-image-audio modalities that most commercial LLMs handle. Ericsson must process logs, videos, wireless sensing data, localization information, and standardization documentation—each requiring purpose-built encoding pipelines and translators. The vision is to bring all these modalities into a unified latent space that a central reasoning engine can process, but this requires significant infrastructure engineering beyond what standard multimodal models provide. The proposed architecture features a reasoning engine at its core (potentially consisting of multiple LLMs or agents) connected to various modality-specific encoders on one side and tools (SDKs, APIs, configuration file generators) on the other. This architecture aims to enable intent-driven Radio Access Network (RAN) management, where operators can request infrastructure enhancements using natural language, and the system automatically translates these requests into concrete configuration changes and optimizations. ## The Case for Symbolic Reasoning Integration Ericsson identifies four fundamental limitations of pure LLM approaches for telecom infrastructure management that motivate their exploration of symbolic reasoning integration: **Hallucination and Reliability**: The statistical, probabilistic nature of LLMs makes hallucinations inevitable. For telecom infrastructure decisions that affect network performance, security, and reliability, unreliable reasoning is unacceptable. Network configuration changes must be deterministic and verifiable. **Lack of True Logical Reasoning**: While chain-of-thought prompting represents the current state-of-the-art for reasoning in LLMs, Ericsson's assessment is that it merely adapts the probability distribution of outputs to follow logical-seeming patterns rather than performing genuine logical inference. This results in arithmetic and consistency errors that are problematic for telecom applications requiring precise calculations and parameter optimization. **Black Box Problem**: The neural network architecture underlying transformers makes them inherently opaque. Ericsson emphasizes that explainability is crucial in telecom contexts—if the system recommends changing infrastructure configuration or flags a security threat, operators need to understand the reasoning chain before taking action. The inability to explain decisions represents a significant barrier to production deployment and regulatory acceptance. **Out-of-Distribution Events**: While LLMs excel at language tasks because they train on massive text corpora, telecom infrastructure generates unique data distributions not well-represented in general pre-training data. More critically, telecom networks encounter "black swan" events—unexpected, rare scenarios that don't follow training distribution patterns. LLMs struggle with these out-of-distribution events, whereas reasoning systems that can apply general principles may handle novel situations more gracefully. ## Domain-Specific Challenges Beyond general LLM limitations, Ericsson identifies telecom-specific challenges that complicate production deployment. The diversity of modalities extends beyond what most LLM applications encounter, including radio sensing data and network telemetry that lack standard representation formats. The non-stationary nature of telecom data means that probability distributions in real network traffic don't match the distributions in pre-training data, even with domain-specific fine-tuning. The dynamic and diverse network environments create high uncertainty, making probabilistic predictions less reliable. To address these telecom-specific challenges, Ericsson proposes several complementary approaches. Digital twins enable running numerous scenarios that mimic real network behavior, generating synthetic training data and testing edge cases without risking production networks. However, the team acknowledges affordability concerns—training LLMs on extensive simulation data is computationally expensive. They propose architectural solutions like GPU-as-a-service within telecom infrastructure, where computing resources idle during low network optimization periods can be rented to external users (similar to AWS), creating a revenue stream that offsets AI infrastructure costs while maintaining capacity for peak optimization needs. ## Hybrid Neural-Symbolic Architecture The core technical innovation Ericsson is pursuing involves hybrid neural-symbolic AI architectures. This approach combines symbolic AI—the rule-based, expert system approach that dominated AI research before the neural network revolution—with modern transformer-based LLMs. Symbolic systems excel at explicit logical reasoning using knowledge bases and deterministic rules, providing the explainability and logical rigor that telecom applications demand. Neural networks excel at pattern recognition and scale efficiently through parallel training, which is why transformers have become the dominant architecture for large language models. The integration of these two paradigms happens through several mechanisms: **Symbolic Chain-of-Thought**: This extends standard chain-of-thought prompting by incorporating formal logical symbols and structures. Rather than simply asking the LLM to think step-by-step, the system teaches it to follow formal logical patterns (e.g., "P implies Q, not-Q implies not-P"). This constrains the reasoning process to follow valid logical inference rules. **Program-Aided Reasoning (PAL)**: This framework allows LLMs to delegate specific reasoning tasks to external rule-based systems or specialized solvers. When the LLM encounters a problem requiring precise calculation or logical inference, it generates code or queries to invoke these external tools, then incorporates their deterministic outputs into its response. This hybrid approach leverages LLMs' language understanding and planning capabilities while ensuring critical reasoning steps use reliable symbolic methods. **External Solver Integration**: Similar to PAL but more general, this involves treating symbolic reasoning engines as agents or tools within a multi-agent framework. The LLM orchestrates the overall problem-solving process but delegates formal reasoning, mathematical optimization, or rule-based decision-making to specialized solvers. ## Production Use Cases Ericsson outlines three primary production use cases for their hybrid neural-symbolic approach: **Network Optimization and Control**: LLM agents perform long-term planning for network optimization based on natural language intents from operators. The symbolic reasoning component ensures that optimization decisions respect hard constraints and follow logical rules about network behavior. Cross-layer reasoning represents a significant capability enhancement—traditionally, different OSI layers (physical, MAC, network) are optimized independently, but the hybrid system can reason across layers simultaneously. For example, when optimizing for energy efficiency, the system can adjust parameters across physical layer transmission power, MAC layer scheduling, and network layer routing in a coordinated fashion. Dynamic optimization enables on-the-fly infrastructure adaptation as conditions change, with the symbolic component ensuring changes don't violate stability or safety constraints. **Anomaly Detection and Security**: Log analysis benefits from combining pattern recognition (neural) with rule-based threat detection (symbolic). The LLM component can identify unusual patterns in log data and correlate events across different system components, while the symbolic component encodes known attack signatures and security rules. This hybrid approach provides improved threat detection because the rule-based component can extract meta-information that pure pattern recognition might miss, while the LLM can identify novel attack patterns that don't match existing rules but show suspicious characteristics. Adaptive security policies represent another application—operators can request policy changes in natural language, and the system translates these into formal security rules and implements them across the infrastructure, with the symbolic component verifying that new policies don't create conflicts or vulnerabilities. **AI-Native Wireless Systems Evolution**: Ericsson positions this work within a broader evolution toward 6G AI-native infrastructure. They describe a progression through intelligence layers: starting with AI components optimizing individual functions within single OSI layers, advancing to single AI systems managing entire protocol stacks within one layer, then reaching AI-native layers where unified AI handles all functions at one OSI layer, and ultimately achieving cross-layer intelligence that manages the entire network stack. This evolution parallels the progression from data-driven systems (using lookup tables and fixed rules, characteristic of 4G) through information-driven systems (basic AI adapting to applications, enabling 5G's differentiated services) to knowledge-driven systems (using reinforcement learning and transfer learning to generalize across scenarios) and finally to reasoning-driven systems (applying versatile logic and semantic understanding to enable capabilities like semantic communication that can exceed Shannon limits by transmitting compressed semantic tokens that receivers reconstruct using shared knowledge). ## Production Deployment Considerations While Ericsson presents an ambitious vision, the presentation acknowledges several unresolved challenges relevant to production deployment. The optimal architecture for telecom LLM applications remains an open question. The consensus appears to be that multi-agent frameworks offer the most flexibility because telecom represents not a single language or domain but multiple interconnected domains, each potentially requiring specialized models and reasoning approaches. The model collapse and hallucination problems, while potentially mitigated through symbolic reasoning integration, cannot be entirely eliminated. Ericsson's approach involves implementing guardrails through prompt engineering and fine-tuning, but acknowledges these are mitigation strategies rather than complete solutions. This honest assessment suggests they maintain realistic expectations about LLM capabilities in production. The affordability and scalability of the proposed architecture present practical deployment challenges. Running digital twins for extensive scenario testing, maintaining multiple specialized models in multi-agent frameworks, operating symbolic reasoning engines alongside LLMs, and processing diverse multimodal data streams all require substantial computational resources. The GPU-as-a-service concept represents one strategy for managing these costs, but the economic viability of AI-native infrastructure at scale remains to be proven. ## Strategic Context and Industry Positioning Ericsson's approach reflects a pragmatic industry position. By focusing on adaptation rather than foundation model development, they avoid competing directly with well-funded AI labs while concentrating resources on domain-specific challenges where their telecom expertise provides competitive advantage. The emphasis on explainability and deterministic reasoning aligns with regulatory and operational requirements in telecommunications that may not constrain consumer AI applications as strictly. The parallel evolution of telecom infrastructure and software development infrastructure that Ericsson describes—from hardware-centric 4G through virtualized and cloud-native 5G toward AI-native 6G—provides strategic justification for their research investments. Just as 5G's service-oriented architecture required cloud-native software infrastructure, 6G's envisioned AI-native capabilities will require the hybrid reasoning systems Ericsson is developing. The presentation references the Daniel Kahneman framework of System 1 (fast, intuitive) versus System 2 (slow, deliberate, logical) thinking, positioning current LLMs as System 1 and hybrid neural-symbolic approaches as moving toward System 2. This framing, also used by AI researchers like Andrej Karpathy, suggests Ericsson views symbolic reasoning integration not as a niche telecom requirement but as part of the broader evolution toward more capable AI systems. Their telecom use cases may serve as early proving grounds for hybrid architectures that eventually see broader adoption. ## Balanced Assessment Ericsson's presentation demonstrates sophisticated understanding of both LLM capabilities and limitations in production contexts. The acknowledgment that chain-of-thought prompting merely adapts probability distributions rather than performing true logical reasoning shows critical thinking about techniques often presented as solutions to reasoning limitations. The recognition that RAG complexity extends far beyond basic implementations reflects production deployment experience rather than just prototyping. However, several aspects warrant careful consideration. The symbolic reasoning approach, while theoretically sound, faces practical challenges that receive less attention in the presentation. Encoding telecom domain knowledge into formal rule systems requires extensive expert knowledge engineering—historically a major bottleneck for symbolic AI. The presentation doesn't detail how Ericsson plans to acquire, maintain, and update these knowledge bases as network technologies evolve. The integration points between neural and symbolic components may create new failure modes not present in either approach individually. The multimodal architecture, while addressing real telecom needs, introduces significant engineering complexity. Each new modality requires custom encoders and translation pipelines, creating maintenance burden and potential points of failure. The vision of unified latent spaces spanning diverse telecom-specific modalities represents a research challenge that may prove more difficult than the presentation suggests. The economic model of GPU-as-a-service to offset AI infrastructure costs faces practical hurdles including security concerns about external access to telecom infrastructure resources, complexity of workload scheduling to ensure GPU availability when network optimization needs arise, and uncertain market demand for GPU rental from telecom facilities with potentially higher latency than dedicated cloud providers. Overall, Ericsson's approach represents thoughtful engagement with real LLMOps challenges in a domain with stringent reliability and explainability requirements. The hybrid neural-symbolic direction addresses genuine limitations of pure LLM approaches, though significant engineering work remains to translate this research vision into production-ready systems. The presentation's value lies not in claiming solved problems but in clearly articulating the challenges of deploying LLMs in critical infrastructure and proposing architecturally sound directions for addressing them.