Slide 1: GenAI Large Telecom Model: The Future of Mobile Network Operations
GenAI Large Telecom Model: The Future of Mobile Network Operations May 12, 2025 11:00 am ET / May 13, 2025 0:00 am JT 1
Slide 2: Today's Speakers
Today’s Speakers Rajeev Koodli Shun Tamura Principal Fellow, AI Engineer, Research Institute of Research Institute of Advanced Technology, Advanced Technology, SoftBank Corp. SoftBank Corp. © SoftBank Corp. 2
Slide 3: Speaker Rajeev Koodli
Speaker Rajeev Koodli Principal Fellow Research Institute of Advanced Technology SoftBank Corp. © SoftBank Corp. 3
Slide 4: Technical Overview of LTM
Technical Overview of LTM 4
Slide 5: Introduction
Introduction © SoftBank Corp. 5
Slide 6: AI-RAN and LTM
AI-RAN and LTM AI-RAN LTM Machine AI Human AI AI for RAN Better gNB performance Better Operation by AI CAPEX Reduction OPEX Reduction © SoftBank Corp. 6
Slide 7: SoftBank AI-RAN: AITRAS
SoftBank AI-RAN: AITRAS : SoftBank Developed AITRAS Orchestrator Manage & control software Provide Edge AI Service Menu RAN L2/L3 Software Edge AI Apps Advanced radio signal processing software LLM Multimodal RAN L1 Software Robot LLM for ADS NVIDIA AI Enterprise*1 NVIDIA AI Aerial Built & implement
Slide 8: SoftBank's AI-for-RAN Use Cases Introduced at MWC25
SoftBank’s AI-for-RAN 3 Use Cases Introduced at MWC25 UL Channel MU-MIMO SRS Interpolation Scheduler Prediction SRS SRS © SoftBank Corp. 8
Slide 9: Human AI: Objectives
Human AI: Objectives Recommend Config Parameters for a new cell adjacent to cells X, Y, & Z Q&A, Summarization, Large Recommendation Fine-tune a Channel Estimation Telecom Network AI Ops AI Inferencing model R&D Engineer Model Business Units What is the root cause Applications of Packet Core Outage
Slide 10: What is LTM?
What is LTM ? Training a single foundation model on diverse data and fine-tuning it for each specific use case Data Use cases FT Text Chat Model Advantages FT Object Image Training Model Detection General Expertise Foundation Model FT Audio Audio Model Synthesis Transfer Learning Tabular FT Model In
Slide 11: Differences With Conventional Methods
Differences With Conventional Methods Conventional Methods Large Telecom Model Operation Simulation Text Operation Simulation Text data data data data data data Machine learning on Machine learning on Machine learning on individual datasets individual datasets individual datasets Trained on various
Slide 12: Journey
© SoftBank Corp. Journey = SoftBank 12
Slide 13: Context
Context ⚫ Started in summer 2024 ⚫ Single Modal models ⚫ Vast amount of heterogeneous network operations data (disparate sources, private to SoftBank) ⚫ Need for large enough infra for data processing and training loops © SoftBank Corp. 13
Slide 14: Data: Total of 4TB of SoftBank-Specific Data Corpus
Data: Total of 4TB of SoftBank-Specific Data Corpus 15min hourly daily 2024/05/07-2024/08/12 2023/08/01-2024/08/12 - RAN KPI data (1840GB ≈ 20GB a day) (1748GB) 2024/05/05-2024/08/12 2023/08/01-2024/08/12 - Core KPI data (27GB) (28GB) - - 2024/07/16-2024/08/12 Config data (573GB *zipped ≈ 500GB a da
Slide 15: Model
Model Dense Transformer Sparse Mixture of Experts Sample architecture Sample architecture ⚫ Dense Transformer and Sparse Mixture of Softmax Softmax Experts (SMoE) are popular architectures Linear Linear ⚫ SMoE allows for faster inference because of fewer active parameters Feed Forward Network FF FF
Slide 16: Challenges with Infrastructure
Challenges with Infrastructure Data Processing Training LLM training Fine tuning Full training Huge raw data ready data Computation resource required (Estimated, Unit: GPU Hour = 1 Hour spent on 1 DGX A100) Based on our approach with LoRA fine tuning Simulation Data size 1k cells 200k cells (All 5G
Slide 17: Options for Human AI (and Machine AI)
Options for Human AI (and Machine AI) In-context based Training based Prompting RAG Fine tuning Full training Cost: Lowest Scalability Task/Domain adaptability Cost: Highest Flexibility Dependency on Limited customizability Full Customizability retrieval quality Limited knowledge added Difficulty in
Slide 18: Supervised Fine Tuning
Supervised Fine Tuning Compute Efficient Sweet spot Compute Intensive PEFT (e.g. LoRA) PEFT with larger Continual Training with domain data domain data Pretraining from scratch • Add limited knowledge • Add more domain • Full adaptability • Full adaptability including • Can start with very small kno
Slide 19: End-End Pipeline
End-End Pipeline Data collection Data processing Training & Evaluation Containerization Serving as an API • Collect necessary data • Import into databases • Supervised • Model format conversion • Deployment on cloud without loss • SQL queries, ETL learning/reinforcement (ONNX, etc.) infrastructure •
Slide 20: Data Pipeline
Data Pipeline Slicing Strategy Instruction Human Review KPI Annotation Annotation Raw Data Sliced Data Pre-trained NLT Data Instruct Model Config RAN Config RAN Config Composite Raw Data Sliced Data NLT Data KPI/Config /Annotation NLT Data (plus other NLT data) Annotation RAN KPI RAN KPI for Trainin
Slide 21: Expert Annotation for Learning
Expert Annotation for Learning Vanilla Description Expert Annotation Absolute priority of carrier frequency used by cell Parameters used when reselection to another cell reselection procedure. Corresponds to parameter during RRC Idle. You can set the priority in 3GPP TS 36.304, Evolved Universal res
Slide 22: Training & Evaluation Pipeline
Training & Evaluation Pipeline Training Evaluation Pre-trained Composite Large Telecom Pre-trained Base Model KPI/Config Model Base Model /Annotation NLT Data Results (plus Other NLT data) Training Data Set Generated Answers Composite KPI/Config/ Eval Questions (Inference Annotation NLT Eval Expecte
Slide 23: Current Model Capabilities
Current Model Capabilities The current version 1 of the model is trained to perform natural language processing tasks (below). It is also finetuned for business applications (such as base station config management) ⚫ Open-ended Q&A ⚫ Trend analysis ⚫ Relation extraction ⚫ Summarization ⚫ Recommendat
Slide 24: General capabilities: What Do They Look Like
General capabilities: What Do They Look Like RAN KPI Relationship between RAN KPI parameters "question": "Describe the time-based trend of "question": "Identify the dependency pattern between `nr_setup_att` values." `dl_active_user_max` and `cu_max` values." "expected": "Increasing in morning, peaki
Slide 25: Lessons Learned
Lessons Learned Technical ⚫ Slicing the different data types (KPI, Config, Annotation) into chunks is crucial for training. Diverse data is important for general capabilities ⚫ Organizing different data types into Model context length can be challenging ⚫ Both data and model hyper-parameter tuning d
Slide 26: SoftBank
SoftBank
Slide 27: Speaker Shun Tamura
Speaker Shun Tamura AI Engineer Research Institute of Advanced Technology SoftBank Corp. © SoftBank Corp. 27
Slide 28: MNO Use Cases
MNO Use Cases 28
Slide 29: Specific UC Example: RAN Config Optimization
Specific UC Example: RAN Config Optimization Scenario ⚫ Handling congestions during events What LTM does ⚫ Optimize RAN config parameters of base stations within targeted area Simulated with NVIDIA Aerial Omniverse Digital Twin Input Output ⚫ Order for congestion handling (Intent) Fine Tuned ⚫ RAN c
Slide 30: Fine-Tuning Methodology for Specific Use Cases
Fine-Tuning Methodology for Specific Use Cases I want to optimize RAN configs You can update RAN configs Divide whole Japan into of this base station in this area. of the base station as follows: training areas Current RAN configs: Updated RAN configs: and evaluation areas Cell ID RAN RAN Cell ID RA
Slide 31: Evaluation Results
Evaluation Results 1. Easier 2. More difficult User specifies the parameter to update, Model infers both parameter to update model only infers its value and its value I want to optimize RAN configs of this base station I want to optimize RAN configs of this base station "BBB" in this area. What woul
Slide 32: Advanced: Full RAN Config Generation
Advanced: Full RAN Config Generation New base station deployment Construction of a new building （with RAN config generation） Simulated with NVIDIA Aerial Omniverse Digital Twin Simulated with NVIDIA Aerial Omniverse Digital Twin Scenario Input Output ⚫ Countermeasures for weak signal areas caused by
Slide 33: Evaluation Results: Full Generation
Evaluation Results: Full Generation Model infers the value of all the parameters I want to generate RAN configs of a new base station "CCC" in this area. What would be the best config value for each parameter? Current RAN configs: Cell ID RAN Config 1 RAN Config 2 AAA True 0 BBB False 1 CCC ??? ???
Slide 34: Examples of MNO Use Cases
Examples of MNO Use Cases Base Station Configurations Maintenance Sales Design Analysis Planning by experts by experts by experts Environ- Alerts mental Optimize and logs Customer Marketing data configurations feedback Action plan strategy Simulations Marketing On-site initiatives maintenance SHOP F
Slide 35: Lessons Learned
Lessons Learned LTM's applicability to very specific downstream tasks i.e., use cases Normally, using specific external tools is an option in such cases, but we proved Good possibility to solve these specific tasks with fine tuning Good reproducibility of real-world optimization Directly suggests th
Slide 36: Integration into Actual Operations
Integration into Actual Operations Additional Human training Large Fine-Tuning AI Base Serving Telecom Optimization of AI Agent Model base station configs (API-fication) Model Configuration logs, simulations, etc. AITRAS Orchestrator Various Maintenance and operations support proprietary data from S
Slide 37: Future Outlook
Future Outlook 38
Slide 38: Summary
Summary ⚫ A foundation model provides a unified approach to adopting GenAI across an organization without requiring specialized skill sets ⚫ A foundation model is also the basis for transfer learning to a variety of use cases in operations ⚫ Telcos have enormous amounts of invaluable multi-modal dat
Slide 39: Next Steps
Next Steps SoftBank will replace and optimize various operational use cases with AI LLM Agent Intent based instructions Tool Use Human AI Telco Operator Simulation provide as an API Retrieval Optimization Guardrail Other MNOs Database AI based signal processing Sync Machine AI Dec L2 L3 Mobile Users
Slide 40: SoftBank
SoftBank
Slide 41: Thank you!
Thank you! 45