## Overview LinkedIn, the professional networking platform serving over 1 billion members and 69 million companies, developed a family of domain-adapted foundation models called EON (Economic Opportunity Network) to power GenAI experiences across their platform. This case study illustrates a sophisticated approach to building and deploying LLMs in production at massive scale, addressing the common enterprise challenges of cost, latency, domain specificity, and responsible AI compliance. The journey began in 2022 when LinkedIn started working on in-house models for personalized AI-assisted messaging. By 2023, they were leveraging GPT-4 for features like Premium profile writing suggestions and collaborative articles. However, proprietary models presented significant challenges around latency and cost at LinkedIn's scale. This motivated the development of domain-adapted models that could match or exceed proprietary model quality while being more cost-effective and better aligned with LinkedIn's specific professional domain. ## The EON Model Architecture and Training Approach The EON models represent a strategic middle ground between using off-the-shelf foundation models and building models entirely from scratch. LinkedIn evaluated multiple state-of-the-art foundation models including Llama, Mixtral, and Mistral as base architectures. The selection process involved measuring performance on reasoning, instruction following, and safety compliance using both open-source and LinkedIn-specific benchmarks. The training pipeline follows a two-step process. First, multi-task instruction tuning with reasoning traces adapts the base model to LinkedIn's professional domain. This step leverages data from the LinkedIn Economic Graph—a digital representation of the global economy and workforce—to create deeply personalized member experiences. The training data totaled approximately 200 million tokens of high-quality, diverse, and de-duplicated content. The second step involves preference and safety alignment using Reinforcement Learning with Human Feedback (RLHF) and Direct Preference Optimization (DPO). This ensures the models generate output aligned with member expectations while adhering to LinkedIn's trust, responsible AI, and fairness principles. A notable optimization they employed was prompt simplification strategies that standardized complex human-written prompts, achieving a 30% reduction in prompt size—a significant efficiency gain at production scale. ## Infrastructure and MLOps Pipeline LinkedIn built their training pipeline on top of an on-premise Kubernetes platform, designed as an end-to-end solution connecting data preprocessing, model training, offline inference, and comprehensive evaluation within a single integrated system. The modular architecture provides flexibility for testing different optimization techniques. For training optimization, teams can configure the pipeline to leverage various techniques including their in-house Liger Kernels, DeepSpeed ZeRO for distributed training, or Huggingface Accelerate. The inference stage similarly supports multiple options including vLLM for efficient serving. This modular approach allows rapid experimentation with state-of-the-art techniques as the GenAI landscape evolves. The evaluation infrastructure is particularly noteworthy. Model evaluation can be triggered on-demand or automatically after training runs, with results persisted in MLFlow and HDFS for deeper review. They developed an in-house evaluation platform that runs a comprehensive benchmark suite, and experiments with the best performance are published to a central company-wide leaderboard—similar in concept to the HuggingFace leaderboard but internal to LinkedIn. ## Evaluation Methodology LinkedIn employs a multi-faceted evaluation approach that combines open-source benchmarks with proprietary LinkedIn-specific assessments. They use the LM Evaluation Harness framework to measure performance on popular benchmarks such as ARC, MuSR, and IFEval, along with the Berkeley Function Calling Leaderboard (BFCL) for function calling capabilities. When reference labels are unavailable, they use GPT-4 variants as judges—a common but pragmatic approach in LLM evaluation. The team acknowledges that while these metrics don't completely replace human evaluation as raw scores, they provide reliable directional signals on model quality. Their evaluation covered three key task categories: job-fit-assessment (a long-text generation task requiring detailed explanations), formatted named entity recognition (NER), and function calling metrics. The job-fit-assessment task is particularly interesting as it requires not just generating categorical outputs but providing detailed reasoning about candidate suitability—a complex task requiring understanding of professional contexts. ## Cost Efficiency Analysis One of the most compelling production considerations is cost. LinkedIn's analysis showed the EON-8B model (domain-adapted from Llama 3.1-8B) to be 75x more cost-effective than GPT-4 and 6x more cost-effective than GPT-4o. These calculations were based on A100 GPU requirements to serve 1 QPS (query per second) of an interactive AI Premium experience, comparing production deployment of GPT instances on Azure versus on-premise deployment of EON-8B models. This dramatic cost differential illustrates why domain adaptation of smaller open-source models can be strategically superior to relying on proprietary models for high-volume production use cases, even when the proprietary models may have broader general capabilities. ## Safety and Responsible AI LinkedIn took a principled approach to safety and responsible AI. To adhere to their trust and fairness principles, they first instruction-tuned the model with synthetically-generated safe outputs for input prompts containing harmful content, then further aligned the model with preference data. Internal safety score measurements assess the out-of-box safety performance of GenAI models. Interestingly, they observed that multi-task instruction tuning initially led to a drop in safety benchmark performance—a common phenomenon where fine-tuning can degrade certain capabilities. However, the use of synthetic safety data and safety alignment steps helped recover performance. This highlights the importance of explicitly incorporating safety considerations throughout the training process rather than relying solely on the base model's safety properties. ## Production Application: Hiring Assistant The Hiring Assistant, launched in October 2024, serves as the primary production application showcasing EON models. This product automates recruiters' repetitive tasks by breaking down queries into multiple steps through an orchestration layer. A key component is an evaluation agent that assesses candidate suitability against AI-extracted job requirements. Nearly 90% of large language model calls in the Hiring Assistant flow come from this evaluation agent, which must parse extensive contexts including candidate profiles, resumes, recruiter notes, and job posts. This high-volume, context-rich use case demands both speed and accuracy—making it an ideal application for a cost-optimized, domain-adapted model. The production results are notable: the EON-8B model improved candidate-job-requirements matching accuracy by an absolute 4% over OpenAI's GPT-4o mini and 30% over Llama-3-8B-instruct, without additional fine-tuning on the specific matching task. This demonstrates the power of domain adaptation—the model's exposure to LinkedIn-specific professional data during training transferred to improved performance on related tasks. The EON model also helps the evaluation agent filter out biased and discriminatory job requirements, aligning with LinkedIn's Responsible AI principles. This is a practical application of safety alignment in a production context where the consequences of biased outputs could be significant. ## Key Learnings and Future Directions LinkedIn documented several important learnings from this initiative. Multi-task instruction tuning with high-quality, diverse instructions and reasoning traces achieves state-of-the-art performance on training tasks while improving domain-specific generalization. The combination of synthetically generated in-house instructions with publicly available open-source data improves instruction diversity and generalizability. Preference and safety alignment helps EON models generate output aligned with member expectations while adhering to governance requirements. Looking ahead, LinkedIn is enhancing EON models to support complex interactions with agents beyond single-task executions. They've improved the base Llama-3.1-8B-instruct model's function calling capabilities on multi-turn and multi-step execution, as measured by BFCLv2 and MMAU benchmarks. Their research roadmap includes enhancing planning and reasoning capabilities, designing efficient context representations, developing efficient storage and retrieval techniques, and dynamic goal identification and evaluation—all pointing toward more sophisticated agentic applications. ## Critical Assessment While LinkedIn presents impressive results, several aspects warrant balanced consideration. The cost comparisons between EON models and GPT-4 are compelling but may not account for all operational costs including infrastructure management, model maintenance, and engineering overhead. The claim of 75x cost efficiency is specifically measured for their on-premise deployment scenario and may vary for organizations with different infrastructure arrangements. The safety and bias mitigation claims, while positive, would benefit from more detailed third-party validation. Similarly, while the 4% improvement over GPT-4o mini on candidate matching is meaningful at scale, the comparison methodology (with prompts optimized by humans for each model) introduces some complexity in interpretation. Nevertheless, this case study represents a thoughtful, well-documented approach to deploying LLMs in production at enterprise scale, with particular attention to cost optimization, domain adaptation, evaluation rigor, and responsible AI considerations.