Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AI2MMUM: AI-AI Oriented Multi-Modal Universal Model Leveraging Telecom Domain Large Model

Tianyu Jiao, Zhuoran Xiao, Yihang Huang, Chenhui Ye, Yijia Feng, Liyu Cai, Jiang Chang, Fangkun Liu, Yin Xu, Dazhi He, Yunfeng Guan, Wenjun Zhang

TL;DR

The paper addresses the challenge of building a scalable, task-aware AI for 6G that can process multi-modal wireless data and perform diverse air-interface tasks. It introduces AI2MMUM, a four-component architecture comprising a multi-modal radio feature extractor, a task-instruction module, a telecom-domain LLM backbone enhanced with LoRA, and lightweight task-specific heads, with frozen radio encoders bridged to the LLM via adapters. Using WAIR-D and DeepMIMO benchmarks, the approach achieves state-of-the-art results on five tasks (direct positioning, LOS/NLOS identification, MIMO precoding, beam selection, and path loss prediction), and ablations demonstrate the value of learnable prompts and LoRA in enabling cross-modal transfer. The work suggests that integrating radio and language capabilities yields robust, scalable, and transferable wireless intelligence with practical implications for unified wireless multi-modal systems.

Abstract

Designing a 6G-oriented universal model capable of processing multi-modal data and executing diverse air interface tasks has emerged as a common goal in future wireless systems. Building on our prior work in communication multi-modal alignment and telecom large language model (LLM), we propose a scalable, task-aware artificial intelligence-air interface multi-modal universal model (AI2MMUM), which flexibility and effectively perform various physical layer tasks according to subtle task instructions. The LLM backbone provides robust contextual comprehension and generalization capabilities, while a fine-tuning approach is adopted to incorporate domain-specific knowledge. To enhance task adaptability, task instructions consist of fixed task keywords and learnable, implicit prefix prompts. Frozen radio modality encoders extract universal representations and adapter layers subsequently bridge radio and language modalities. Moreover, lightweight task-specific heads are designed to directly output task objectives. Comprehensive evaluations demonstrate that AI2MMUM achieves SOTA performance across five representative physical environment/wireless channel-based downstream tasks using the WAIR-D and DeepMIMO datasets.

AI2MMUM: AI-AI Oriented Multi-Modal Universal Model Leveraging Telecom Domain Large Model

TL;DR

The paper addresses the challenge of building a scalable, task-aware AI for 6G that can process multi-modal wireless data and perform diverse air-interface tasks. It introduces AI2MMUM, a four-component architecture comprising a multi-modal radio feature extractor, a task-instruction module, a telecom-domain LLM backbone enhanced with LoRA, and lightweight task-specific heads, with frozen radio encoders bridged to the LLM via adapters. Using WAIR-D and DeepMIMO benchmarks, the approach achieves state-of-the-art results on five tasks (direct positioning, LOS/NLOS identification, MIMO precoding, beam selection, and path loss prediction), and ablations demonstrate the value of learnable prompts and LoRA in enabling cross-modal transfer. The work suggests that integrating radio and language capabilities yields robust, scalable, and transferable wireless intelligence with practical implications for unified wireless multi-modal systems.

Abstract

Designing a 6G-oriented universal model capable of processing multi-modal data and executing diverse air interface tasks has emerged as a common goal in future wireless systems. Building on our prior work in communication multi-modal alignment and telecom large language model (LLM), we propose a scalable, task-aware artificial intelligence-air interface multi-modal universal model (AI2MMUM), which flexibility and effectively perform various physical layer tasks according to subtle task instructions. The LLM backbone provides robust contextual comprehension and generalization capabilities, while a fine-tuning approach is adopted to incorporate domain-specific knowledge. To enhance task adaptability, task instructions consist of fixed task keywords and learnable, implicit prefix prompts. Frozen radio modality encoders extract universal representations and adapter layers subsequently bridge radio and language modalities. Moreover, lightweight task-specific heads are designed to directly output task objectives. Comprehensive evaluations demonstrate that AI2MMUM achieves SOTA performance across five representative physical environment/wireless channel-based downstream tasks using the WAIR-D and DeepMIMO datasets.
Paper Structure (12 sections, 7 equations, 5 figures, 1 table)

This paper contains 12 sections, 7 equations, 5 figures, 1 table.

Table of Contents

  1. Introduction
  2. Channel Model and Problem Statement
  3. Proposed Network Structure
  4. Multi-Modal Radio Feature Extraction Module
  5. Task Instruction Module
  6. Telecom LLM Backbone Enhanced with LoRA
  7. Task-Specific Heads
  8. Experimental Results and Analysis
  9. Dataset Generation and Model Settings
  10. Downstream Tasks and Benchmarks
  11. Results and Analysis
  12. Conclusion

Figures (5)

  • Figure 1: The 6G-oriented AI$^2$MMUM capable of processing wireless multi-modal data and performing various air interface tasks.
  • Figure 2: Network structure of the proposed 6G-oriented, scalable, and task-aware AI$^2$MMUM.
  • Figure 3: Framework for communication multi-modal alignment.
  • Figure 4: The performance of our proposed method and six benchmarks across the channel-based direct positioning, LOS/NLOS identification, and MIMO precoding tasks. Left: WAIR-D area #00032. Right: DeepMIMO O1 BS#12.
  • Figure 5: The performance of our proposed method and six benchmarks across the environment-based beam selection and path loss prediction tasks in WAIR-D area #00247.