1. Bibliographic Information

1.1. Title

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

1.2. Authors

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

1.3. Journal/Conference

The paper is published as a preprint on arXiv (arxiv:2505.10003) as of 2025-05-15T06:32:59.000Z. While not yet peer-reviewed and officially published in a journal or conference proceedings at the time of this analysis, the authors' affiliations (implied by the Fellow, IEEE for Wenjun Zhang) suggest a strong connection to the Institute of Electrical and Electronics Engineers (IEEE), a globally recognized professional organization for advancing technology. Papers from these authors often target high-impact journals and conferences in wireless communications and AI.

1.4. Publication Year

2025 (Based on the UTC publication date: 2025-05-15T06:32:59.000Z)

1.5. Abstract

The paper proposes AI2MMUM (Artificial Intelligence-Air Interface Multi-Modal Universal Model), a 6G-oriented universal model designed to process diverse multi-modal data and execute various physical layer tasks in wireless systems. Building on previous work in communication multi-modal alignment and telecom large language models (LLMs), AI2MMUM is scalable and task-aware, flexibly performing tasks based on subtle instructions. Its core features include an LLM backbone for robust contextual comprehension and generalization, enhanced by fine-tuning with domain-specific knowledge using Low-Rank Adaptation (LoRA). Task instructions combine fixed keywords with learnable, implicit prefix prompts for adaptability. Frozen radio modality encoders extract universal representations, with adapter layers bridging radio and language modalities. Lightweight task-specific heads directly output task objectives. Comprehensive evaluations demonstrate AI2MMUM achieves SOTA (State-of-the-Art) performance across five representative physical environment/wireless channel-based downstream tasks using the WAIR-D and DeepMIMO datasets.

1.6. Original Source Link

Official Source: https://arxiv.org/abs/2505.10003 PDF Link: https://arxiv.org/pdf/2505.10003v1.pdf Publication Status: Preprint on arXiv.

2. Executive Summary

2.1. Background & Motivation

The vision for 6G (sixth-generation wireless communication) networks is to integrate pervasive intelligence and natively support artificial intelligence (AI). Traditional wireless AI models are typically task-specific, meaning they are designed to solve only one type of problem (e.g., beamforming). This leads to several issues:

1. Limited Transferability: A model trained for one task cannot easily be adapted or used for another, even if the tasks are related.

2. Increased System Complexity: As the number of AI applications in wireless systems grows, managing many small, specialized models becomes exponentially complex.

3. Model Management Challenges: Deploying, updating, and maintaining a large collection of diverse models is inefficient and resource-intensive.

   The core problem the paper aims to solve is the lack of a universal model in wireless communication that can process diverse multi-modal data (like vision, maps, location, wireless channels, radar) and perform a wide range of air interface tasks (e.g., positioning, beamforming) with high accuracy and flexibility. This is crucial for emerging 6G technologies such as Integrated Sensing and Communication (ISAC), vision-aided communication, and Vehicle-to-Everything (V2X), which inherently involve multiple data types and complex interactions.

The paper's entry point is to leverage the success of Large Language Models (LLMs) in natural language processing, which have demonstrated remarkable contextual comprehension and generalization capabilities across diverse tasks. The innovative idea is to adapt LLM principles to the wireless domain, creating an AI-air interface multi-modal universal model (AI2MMUM) that can understand task instructions and process various wireless modalities.

2.2. Main Contributions / Findings

The paper's primary contributions are:

1. Proposed AI2MMUM Framework: The authors introduce a novel, scalable, and task-aware AI2MMUM framework specifically designed for 6G environments. This framework integrates multi-modal radio feature extraction, task instructions, an LLM backbone, and task-specific heads.

2. Task Instruction Module Design: A flexible task instruction module is proposed, combining fixed task keywords with learnable, implicit prefix prompts. This design enhances AI2MMUM's transferability and improves its ability to adapt to diverse tasks by providing discriminative information.

3. Integration of Pre-trained Multi-modal Encoders: The framework leverages robust multi-modal radio encoders (specifically EPNN and CFENN from their prior work), which are pre-trained on large-scale datasets to extract universal, task-agnostic features from physical environment and wireless channel data. These encoders are frozen, and adapter layers are used to bridge the radio and language embedding spaces, promoting cross-modal knowledge transfer.

4. Telecom LLM Backbone with LoRA: The LLM backbone, derived from a telecom LLM (retrained from LLaMA2-7B with a telecommunication corpus), provides strong contextual comprehension. Low-Rank Adaptation (LoRA) is employed for fine-tuning, enabling efficient incorporation of wireless domain-specific knowledge while preserving the LLM's original language capabilities and significantly reducing the number of trainable parameters.

5. Lightweight Task-Specific Heads: The model utilizes lightweight task-specific heads that directly output task objectives from a single predicted token, enhancing efficiency and simplifying the external network structure.

6. Comprehensive Evaluation and SOTA Performance: AI2MMUM is comprehensively evaluated across five representative physical environment/wireless channel-based downstream tasks (direct positioning, LOS/NLOS identification, MIMO precoding, beam selection, and path loss prediction). The model consistently achieves SOTA performance compared to traditional non-LLM methods and models lacking its innovative design, demonstrating the compatibility and synergy between radio and language knowledge.

   These findings address the critical need for a unified AI model in 6G by offering a flexible, efficient, and high-performing solution that can generalize across multiple wireless tasks and data modalities, reducing system complexity and improving AI management.

3. Prerequisite Knowledge & Related Work

3.1. Foundational Concepts

To fully understand the AI2MMUM paper, a beginner should be familiar with the following core concepts:

• 6G (Sixth-Generation Wireless Communication): The next generation of wireless communication technology, envisioned to succeed 5G. 6G aims for even higher data rates, lower latency, massive connectivity, and pervasive intelligence, often incorporating AI natively into the network design. Key features include Integrated Sensing and Communication (ISAC), vision-aided communication, and Vehicle-to-Everything (V2X).
• Artificial Intelligence (AI): A broad field of computer science that enables machines to perform tasks that typically require human intelligence, such as learning, problem-solving, perception, and language understanding.
• Multi-modal Data: Data that comes from multiple different modalities or sources. For example, in wireless communication, this could include:
  • Vision: Images or video data (e.g., from cameras).
  • Maps/Location: Geographical information, GPS coordinates.
  • Wireless Channels (CSI): Information about how radio signals propagate through the environment.
  • Radar/LiDAR: Sensor data providing distance and environmental mapping.
  • Language: Textual instructions or descriptions.
• Universal Model (or Foundation Model): A large, pre-trained AI model designed to perform a wide range of tasks across different domains or modalities with high accuracy, leveraging extensive parameters and vast datasets for knowledge integration, reasoning, and generalization. The goal is to avoid designing a new model for every single task.
• Large Language Model (LLM): A type of AI model, typically based on the Transformer architecture, trained on massive amounts of text data. LLMs are capable of understanding, generating, and processing human language, performing tasks like translation, summarization, and question-answering. They excel at contextual comprehension and generalization. Examples include GPT-3, LLaMA, and BERT.
• AI-air interface: Refers to the application of AI directly to the air interface, which is the radio link between a user device (like a smartphone) and the base station. This involves AI managing and optimizing various physical layer tasks.
• Physical Layer Tasks: The lowest layer in the OSI (Open Systems Interconnection) model, responsible for the actual transmission and reception of raw data bits over a physical medium. In wireless communication, physical layer tasks include:
  • Positioning: Determining the precise location of a user equipment (UE).
  • LOS/NLOS Identification: Distinguishing between Line-of-Sight (direct path between transmitter and receiver) and Non-Line-of-Sight (path obstructed by obstacles) conditions, which significantly impacts signal quality.
  • MIMO Precoding: In Massive MIMO systems, it's a signal processing technique at the transmitter to optimize the signal transmitted from multiple antennas to multiple receivers, improving signal quality and capacity.
  • Beam Selection: Choosing the optimal narrow beam from a set of available beams (e.g., from a DFT codebook) to direct energy towards a specific user, especially critical in millimeter wave (mmWave) communications.
  • Path Loss Prediction: Estimating the signal power reduction as it propagates from a transmitter to a receiver, crucial for network planning and coverage analysis.
• Massive Multiple-Input Multiple-Output (MIMO): A key technology in modern wireless communication where both the transmitter (e.g., Base Station - BS) and receiver (e.g., User Equipment - UE) are equipped with multiple antennas. Massive MIMO uses a very large number of antennas at the BS to serve multiple UEs simultaneously, improving spectral efficiency and reliability.
• Orthogonal Frequency Division Multiplexing (OFDM): A digital modulation scheme used in 5G, Wi-Fi, and LTE. It divides a single wideband channel into many narrower orthogonal subcarrier frequencies, improving robustness against frequency-selective fading and inter-symbol interference.
• Channel State Information (CSI): A crucial piece of information in wireless communication that describes the characteristics of the wireless channel between a transmitter and a receiver. It includes parameters like angle of arrival (AoA), time delay, and amplitude attenuation. CSI is essential for precoding, beamforming, and other adaptive transmission techniques.
• Low-Rank Adaptation (LoRA): A parameter-efficient fine-tuning technique for large AI models. Instead of fine-tuning all parameters of a pre-trained model, LoRA injects small, trainable matrices (adapters) into specific layers (e.g., self-attention query/key matrices). These adapter matrices are low-rank, meaning they have significantly fewer parameters than the original matrices. This drastically reduces the number of parameters that need to be trained, making fine-tuning faster and less memory-intensive, especially useful for adapting LLMs to new domains while preserving original knowledge.
• Contrastive Learning: A self-supervised learning paradigm where a model learns representations by pushing similar (positive) samples closer together in an embedding space and dissimilar (negative) samples farther apart. It's often used for pre-training encoders without explicit labels, extracting universal features.
• Transformer Architecture: A neural network architecture introduced in 2017, foundational to modern LLMs. It relies heavily on self-attention mechanisms to process sequential data (like text or feature sequences), allowing the model to weigh the importance of different parts of the input sequence when processing each element. It consists of multiple stacked Transformer blocks, each containing self-attention and feedforward networks.
• Self-Attention Mechanism: A core component of the Transformer architecture. It allows the model to dynamically weigh the importance of different input tokens (or features) when processing a particular token. It computes Query (Q), Key (K), and Value (V) matrices from the input. The attention score is calculated by the dot product of $Q$ and $K$, scaled by $\sqrt{d_k}$ (where $d_k$ is the dimension of $Q$ and $K$), and then passed through a softmax function to get attention weights. These weights are then multiplied by $V$ to get the attended output. The standard formula is: $ \mathrm{Attention}(Q, K, V) = \mathrm{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right)V $ Where:
  • $Q$: Query matrix, derived from the input embedding.
  • $K$: Key matrix, derived from the input embedding.
  • $V$: Value matrix, derived from the input embedding.
  • $Q K^T$: Dot product of Query and Key, measuring similarity.
  • $\sqrt{d_k}$: Scaling factor, where $d_k$ is the dimension of the key vectors, used to prevent the dot product from becoming too large and pushing the softmax into saturated regions.
  • $\mathrm{softmax}$: Normalization function that converts raw scores into probabilities.

3.2. Previous Works

The paper builds upon and differentiates itself from several prior attempts and visions for large AI models in wireless communications:

• Visions for 6G and Big AI Models [1, 2]:
  • [1] Chen et al. discuss the opportunities, challenges, and research directions for "Big AI models for 6G wireless networks."
  • [2] Bariah et al. explore "Large generative AI models for telecom: The next big thing?"
  • Relevance: These papers establish the high-level vision and motivation for AI2MMUM – the need for universal, large AI models in 6G to handle complexity and enable pervasive intelligence. AI2MMUM aims to provide a concrete architectural and training methodology to realize this vision.
• Wireless-centric Foundation Models [3]:
  • [3] Xu et al. proposed Large Multi-Modal Models (LMMs) as "universal foundation models for AI-native wireless systems," integrating capabilities like multi-modal data fusion, grounding, and instructibility.
  • Relevance: This work is a direct conceptual precursor to AI2MMUM. Both recognize the need for LMMs in wireless. AI2MMUM provides a detailed model structure, specific training methods (LoRA), and empirical validation across diverse physical layer tasks, moving beyond preliminary visions to a practical implementation.
• Cross-modal Fusion in V2X and Networking [4, 5, 6]:
  • [4] Cao et al. introduced MAPLM, a benchmark incorporating 2D images, 3D LiDAR point clouds, and map contexts into LLMs for map and traffic scene understanding in V2X scenarios.
  • [5] Guan et al. developed Talk2Radar for bridging natural language with 4D mmWave radar for 3D visual grounding in autonomous driving.
  • [6] Wu et al. created NetLLM to adapt LLMs for processing multi-modal networking data and generating task-specific answers.
  • Relevance: These works demonstrate the growing trend of integrating LLMs with diverse modalities (vision, LiDAR, radar) in related domains like V2X and networking. They show the potential of LLMs for multi-modal understanding. AI2MMUM extends this idea specifically to the air interface and core physical layer tasks, focusing on wireless channels and physical environment data, which are distinct from generic V2X or networking data.
• Prior Work on Communication Multi-modal Alignment [7, 8]:
  • [7] Jiao et al. (the current paper's authors) worked on "6G-oriented CSI-based multi-modal pre-training and downstream task adaptation paradigm."
  • [8] Jiao et al. (the current paper's authors) also addressed "the curse of scenario and task generalization in AI-6G: A multi-modal paradigm," proposing EPNN and CFENN.
  • Relevance: These are the authors' foundational works that AI2MMUM directly builds upon. They provided the pre-trained Environment Perception Neural Network (EPNN) and Channel Feature Extraction Neural Network (CFENN) modules, which are crucial components for AI2MMUM's multi-modal radio feature extraction. AI2MMUM integrates these pre-trained encoders into an LLM-centric universal model, adding the task instruction module, LoRA-enhanced LLM backbone, and task-specific heads.

3.3. Technological Evolution

The field of wireless communication AI has evolved from:

1. Task-Specific AI Models: Early AI in wireless involved designing small, highly specialized neural networks for individual tasks like channel estimation or modulation classification. These were efficient for their specific purpose but lacked versatility.

2. Model-Driven vs. Data-Driven: A shift towards data-driven AI where models learn directly from massive wireless data, reducing reliance on complex mathematical models.

3. Large-Scale Pre-training: Inspired by computer vision and NLP, the idea of pre-training large models on vast datasets to learn general representations, then fine-tuning them for specific tasks.

4. Multi-modal Integration: With 6G applications like ISAC and V2X, the need to integrate different modalities (e.g., vision, CSI, LiDAR) became apparent, leading to research in multi-modal alignment and fusion.

5. Foundation Models/Universal Models: The latest evolution, driven by the success of LLMs, is the push for powerful general-purpose AI models that can handle diverse tasks and modalities, acting as a single intelligence for complex systems.

   This paper's work fits squarely into the fifth stage of this evolution, proposing a concrete architecture for a universal AI model (AI2MMUM) that combines multi-modal processing with LLM capabilities, explicitly addressing 6G requirements.

3.4. Differentiation Analysis

Compared to the main methods and visions in related work, AI2MMUM offers several key differentiators:

• Systematic Model Structure for AI-Air Interface: While [3] proposed a conceptual LMM for AI-native wireless systems, AI2MMUM provides a detailed and systematic model architecture (Fig. 2) specifically tailored for the AI-air interface, outlining how multi-modal radio features, task instructions, and an LLM backbone are integrated.

• Explicit Task Instruction Design: AI2MMUM introduces a novel task instruction module that combines fixed task keywords with learnable prefix prompts. This explicit mechanism for instructibility allows the model to dynamically adapt its behavior to specific task requirements, which is more advanced than generic multi-modal data fusion approaches.

• Leveraging Pre-trained Wireless Modality Encoders: Instead of training multi-modal encoders from scratch for each new application, AI2MMUM effectively utilizes robust, pre-trained EPNN and CFENN from prior work [7, 8]. These frozen encoders provide universal radio modality representations, significantly reducing the need for extensive labeled data and enhancing generalization. Adapter layers then efficiently bridge these representations to the LLM's embedding space.

• LoRA-Enhanced Telecom LLM for Domain Adaptation: The paper adopts LoRA to fine-tune a telecom LLM (retrained from LLaMA2-7B). This is a cost-effective and flexible approach to inject wireless domain-specific knowledge into a general-purpose LLM while preserving its vast language understanding, a crucial aspect for practical deployment in resource-constrained 6G networks. This is more specific to wireless physical layer tasks than general networking LLMs like NetLLM [6].

• Direct Output with Task-Specific Heads: AI2MMUM employs lightweight task-specific heads for direct output of task objectives from a single predicted token. This contrasts with traditional LLM approaches that might involve multiple iterations of token generation and de-tokenization, improving prediction accuracy and computational efficiency for numerical or structured wireless tasks.

• Comprehensive Experimental Validation: The paper provides extensive empirical evidence across five distinct physical layer tasks and two datasets, demonstrating SOTA performance and validating the necessity of each proposed module through detailed ablation studies. This goes beyond conceptual proposals or domain-adjacent applications by focusing directly on core air interface problems.

  In essence, AI2MMUM differentiates itself by offering a more concrete, integrated, and empirically validated framework for building an AI-air interface multi-modal universal model, specifically addressing the unique challenges and opportunities of 6G wireless systems.

4. Methodology

4.1. Principles

The core idea behind AI2MMUM is to develop a single, versatile AI model that can understand human-like instructions and process various types of wireless communication data (multi-modal data) to perform diverse tasks at the physical layer. The theoretical basis is rooted in leveraging the contextual comprehension and generalization capabilities of Large Language Models (LLMs), which are excellent at processing sequential data and following instructions. By adapting an LLM backbone to integrate wireless modalities and task-specific guidance, the model can learn to map complex wireless scenarios to desired outputs for tasks like positioning or beamforming, reducing the need for separate, specialized AI models for each task. The intuition is that just as LLMs can understand subtle nuances in language, a similarly designed model could understand subtle task instructions and extract relevant features from wireless data to perform specific wireless operations.

4.2. Core Methodology In-depth (Layer by Layer)

The proposed AI2MMUM is a 6G-oriented, scalable, and task-aware model composed of four primary components: a multi-modal radio feature extraction module, a task instruction module, a telecom LLM backbone enhanced with LoRA, and task-specific heads. The overall architecture is depicted in Figure 2.

The process begins by taking wireless Channel State Information (CSI) and a task instruction as inputs. Let's first define the wireless channel model used, as this forms the basis for one of the key input modalities.

The paper considers a Massive Multiple-Input Multiple-Output (MIMO) system operating in Orthogonal Frequency Division Multiplexing (OFDM) mode with $N_c$ subcarriers. The Base Station (BS) is equipped with $N_t$ antennas arranged in a Uniform Linear Array (ULA), and the User Equipment (UE) has a single antenna. The wireless channel between the BS and the UE for a given frequency $f$ can be written as:

$ \mathbf{h}(f) = \sum_{i=1}^{N_{\mathrm{path}}} \alpha_i e^{-j2\pi f \tau_i} \mathbf{a}(\theta_i) $

Where:

• $f$: The carrier frequency.

• $N_{\mathrm{path}}$: The number of propagation paths, representing different routes a signal takes from transmitter to receiver.

• $\alpha_i$: The amplitude attenuation (strength reduction) of the $i$-th path.

• $\tau_i$: The time delay of the $i$-th path, indicating how long it takes for the signal to travel along that path.

• $\theta_i$: The Angle of Arrival (AoA) of the $i$-th path, specifying the direction from which the signal arrives at the receiver.

• $\mathbf{a}(\theta_i)$: The array vector for the $i$-th path, which describes how the signal phase and amplitude vary across the ULA antennas due to the AoA. It is defined as:

  $ \mathbf{a}(\theta_i) = [1, e^{-j\beta \cos \theta_i}, \cdot \cdot \cdot, e^{-j\beta (N_t - 1) \cos \theta_i}]^T $

Where:

• $\beta = 2\pi df / c$: A constant factor.

• $d$: The antenna spacing between adjacent elements in the ULA.

• $c$: The speed of light.

• $N_t$: The number of antennas at the BS.

  Consequently, the CSI matrix $\mathbf{H} \in \mathbb{C}^{N_t \times N_c}$ for all subcarriers is defined as:

$ \mathbf{H} = [\mathbf{h}(f_1), \mathbf{h}(f_2), \cdots, \mathbf{h}(f_{N_c})] $

Where:

• $\{ f_i \mid i = 1, 2, \cdot \cdot \cdot, N_c \}$ is the set of subcarrier frequencies used in OFDM.

• Each column $\mathbf{h}(f_i)$ is the channel vector for a specific subcarrier $f_i$.

  With this understanding of CSI, the AI2MMUM processes the wireless CSI data and task instructions as follows:

4.2.1. Multi-Modal Radio Feature Extraction Module

This module is responsible for robustly extracting informative features from wireless data, specifically physical environment data and wireless channel data.

1. Pre-trained Encoders (EPNN and CFENN): The module uses Environment Perception Neural Network (EPNN) and Channel Feature Extraction Neural Network (CFENN). These networks were pre-trained in the authors' prior work [8] on extensive datasets (2.25M modality sample pairs from 9,000 WAIR-D areas) using contrastive learning.

   • EPNN: Processes physical environment modality data (e.g., area maps, BS positions, UE positions).

   • CFENN: Processes wireless channel modality data (e.g., CSI).

   • Contrastive Learning: A technique where the models learn by maximizing the similarity of related environment-channel pairs and minimizing the similarity of unrelated pairs in an embedding space. This process enables EPNN and CFENN to extract universal modality representations that generalize across different scenarios.

   • Frozen Encoders: Once pre-trained, EPNN and CFENN are frozen, meaning their parameters are not updated during AI2MMUM training. They act as fixed feature extractors. The framework for this communication multi-modal alignment is illustrated in Figure 3.

     该图像是一个示意图，展示了无线信道和物理环境如何通过 CFNN 和 EPNN 处理并输出结果。上方的无线信道被映射到 CFNN，而下方的物理环境则通过 EPNN 进行处理，最终生成的输出数据呈现于右侧的矩阵中。

   Fig. 3. Framework for communication multi-modal alignment.

   The output of these encoders, representing the wireless channel $\mathbf{H}$ (or physical environment data $P$), is transformed into a token embedding vector $\mathbf{E_H}$ (or $\mathbf{E_P}$):

   $ \mathbf{E_H} = f_{\mathbf{H}}(\mathbf{H}; \Theta_{\mathbf{H}}) $

   Where:

   • $\mathbf{E_H}$: The token embedding vector for the wireless channel data.
   • $f_{\mathbf{H}}$: The function performed by the CFENN (or EPNN for environment data).
   • $\mathbf{H}$: The input wireless channel data (or physical environment data $P$).
   • $\Theta_{\mathbf{H}}$: The neural network parameters of the CFENN (or EPNN), which are frozen. The EPNN and CFENN generate universal modality representations with a dimensionality of 128.

2. Adapter Layers: These are small neural network layers (linear layers) used to bridge the dimensional mismatch between the 128-dimensional output of the radio modality encoders and the 4096-dimensional input required by the LLM backbone.

   • Purpose: They facilitate seamless cross-modal knowledge transfer and integration between the radio and language modalities.
   • Flexibility: When new modalities (like radar or LiDAR) are introduced, only new adapter layers need to be trained, while the existing modality encoders remain frozen, minimizing computational overhead.

4.2.2. Task Instruction Module

This module provides AI2MMUM with discriminative information to steer it towards processing wireless data for specific tasks.

1. Language-based Instructions: Task instructions are provided as human-friendly text.

2. Tokenization and Embedding: The text instruction $L$ is first tokenized (converted into vocabulary indices) and then mapped to high-dimensional token embeddings $\mathbf{E_L}$:

   $ \mathbf{E_L} = f_{\mathrm{L}}(\mathrm{L}; \Theta_{\mathrm{L}}) $

   Where:

   • $\mathbf{E_L}$: The token embedding vector for the task instruction.
   • $f_{\mathrm{L}}$: The function performed by the tokenizer and embedding layer for language.
   • $\mathrm{L}$: The input task instruction text.
   • $\Theta_{\mathrm{L}}$: The parameters of the tokenizer and embedding layer for language. The token embedding dimension for language is 4096.

3. Fixed Task Keywords and Learnable Prefix Prompts: To construct optimal, task-specific prompts, the module integrates:

   • Fixed Task Keyword Embeddings: Consistent keywords (e.g., "position", "LOS status", "precoding") that explicitly define the task.
   • Learnable Prefix Prompts: Trainable embeddings (consisting of multiple tokens) that implicitly encode task instructions. These are learned during training to optimize task performance. This design enhances AI2MMUM's transferability and aligns with human cognition.

4.2.3. Telecom LLM Backbone Enhanced with LoRA

This is the central processing unit of AI2MMUM.

1. Concatenation and Positional Embeddings: The token embeddings from the radio modality ($\mathbf{E_H}$ or $\mathbf{E_P}$) and the task instruction ($\mathbf{E_L}$) are concatenated: $\mathrm{Concat}(\mathbf{E_H}, \mathbf{E_L})$. Positional embeddings are then added to this combined sequence of tokens.

   • Purpose of Positional Embeddings: To provide sequential information, as Transformer architectures process tokens in parallel without inherent knowledge of their order. This is crucial for accurate multi-modal context handling.

     $ \mathbf{E_B} = f_{\mathrm{B}}(\mathrm{Concat}(\mathbf{E_H}, \mathbf{E_L}); \Theta_{\mathrm{B}}) $

   Where:

   • $\mathbf{E_B}$: The feature output from the LLM backbone after processing the concatenated embeddings.
   • $f_{\mathrm{B}}$: The function performed by the LLM backbone.
   • $\Theta_{\mathrm{B}}$: The parameters of the LLM backbone.

2. LLM Backbone Structure: The backbone comprises multiple stacked Transformer blocks. Each block typically includes:

   • Self-Attention Mechanisms: These allow the model to weigh the importance of different tokens (both language and radio features) within the combined sequence when processing each token, facilitating multi-modal context comprehension.
   • Feedforward Networks: These process the output of the self-attention layer independently for each token position, adding non-linearity and further feature transformation.
   • Pre-training: The LLM backbone is derived from a telecom LLM, which was retrained from the LLaMA2-7B model using a telecommunication corpus. This pre-training provides robust language knowledge and generalization capabilities.

3. Low-Rank Adaptation (LoRA) for Fine-tuning: To incorporate wireless domain-specific knowledge efficiently without altering the large pre-trained LLM backbone extensively, LoRA is employed.

   • Mechanism: For each pre-trained weight matrix $W_0$ (e.g., in the query and key matrices of the self-attention mechanism) of dimension $a \times b$, LoRA introduces two low-rank matrices: $A \in \mathbb{R}^{a \times r}$ and $B \in \mathbb{R}^{r \times b}$, where $r \ll \min\{a, b\}$. The update to the original weight matrix is approximated as $W = W_0 + AB$.

   • Training: During adaptation, $W_0$ remains frozen. Only the parameters in matrices $A$ and $B$ are trained. This significantly reduces the number of trainable parameters (from $a \times b$ to $a \times r + r \times b$), making fine-tuning much more efficient.

   • Scalability: Multiple LoRAs can share the same backbone, allowing rapid modality switching and high scalability when adapting to different wireless modalities or tasks.

     The architecture showing LoRA integration is part of Figure 2:

     该图像是图表，展示了6G导向的可扩展任务感知AI2MMUM的网络结构。图中包括多模态无线电特征提取、通信LLM主干和任务特定头部等模块，涉及位置、LOS状态等任务指令以及任务关键描述。

   Fig. 2. Network structure of the proposed 6G-oriented, scalable, and task-aware AI2MMUM

4.2.4. Task-Specific Heads

These lightweight modules are responsible for transforming the LLM backbone's output into the final task objective.

1. Single-Pass Approach: Instead of generating multiple language tokens and then de-embedding them, LLMs typically use the last token of the output sequence from a single forward pass as the predicted result for that iteration. This focuses the backbone on the original multi-modal input, minimizing output uncertainty and reducing inference time.

2. Structure: Task-specific heads consist of a single linear layer.

   • Purpose: To transform the 4096-dimensional task-related feature token (the last token embedding from the LLM backbone) into the desired downstream task objective.

   • API Encapsulation: In practical applications, these heads can be encapsulated within diverse Application Programming Interfaces (APIs), allowing AI2MMUM to identify and invoke the appropriate API based on task instructions and function calls.

     $ \mathrm{T} = f_{\mathrm{T}}(\mathbf{E_B}; \boldsymbol{\Theta}_{\mathrm{T}}) $

   Where:

   • $\mathrm{T}$: The final sub-task objective (e.g., UE position, LOS status, precoding matrix).

   • $f_{\mathrm{T}}$: The function performed by the task-specific head.

   • $\mathbf{E_B}$: The output feature from the LLM backbone.

   • $\boldsymbol{\Theta}_{\mathrm{T}}$: The neural network parameters of the task-specific head.

     This structured design allows AI2MMUM to flexibly and effectively perform various physical layer tasks by intelligently interpreting task instructions and extracting relevant features from diverse wireless modalities.

5. Experimental Setup

5.1. Datasets

The experiments utilize two distinct datasets:

1. WAIR-D (Wireless AI Research Dataset) [9]:

   • Source: A real-world dataset for wireless AI research.
   • Characteristics: It comprises 10,000 real-world areas of varying sizes. The authors' prior work on communication multi-modal alignment used 2.25 million modality sample pairs from 9,000 WAIR-D areas (numbered #01001 to #10000). These samples included physical environment data (area maps, BS positions, and UE positions) and wireless channel data (CSIs).
   • Usage in this study: For training and testing AI2MMUM, the authors used 10,000 samples from two previously unseen WAIR-D areas: #00032 and #00247. The use of unseen areas is crucial for validating the model's generalization capabilities to new environments.
   • Example Data Sample: While not explicitly shown in the paper, a WAIR-D sample for physical environment data would include an image or coordinate representation of a city area map, specific (x,y) coordinates for BS and UE locations. A wireless channel data sample would be a complex-valued CSI matrix $\mathbf{H}$ (as defined in Section II), derived from simulations or measurements in that specific environment.

2. DeepMIMO Dataset [10]:

   • Source: A generic deep learning dataset for millimeter wave (mmWave) and Massive MIMO applications.
   • Characteristics: The authors employ the "Outdoor 1 (O1)" scenario. This scenario features 18 BSs and UEs positioned in a cross-shaped area surrounded by buildings, simulating a dense urban environment typical for mmWave deployments.
   • Usage in this study: AI2MMUM was trained and tested using 10,000 samples from the DeepMIMO O1 scenario, specifically from BS#12.
   • Example Data Sample: A DeepMIMO sample would include CSI data specific to mmWave frequencies, often characterized by sparse channels due to blockages, and physical environment data like the layout of buildings and positions of BSs and UEs within the O1 scenario.

Why these datasets were chosen:

• WAIR-D: Provides real-world environmental complexity and a large volume of multi-modal data for pre-training and testing, allowing evaluation of generalization to unseen real-world areas.
• DeepMIMO: Offers a standardized and well-controlled simulation environment, particularly suitable for mmWave and Massive MIMO scenarios, which are highly relevant for 6G research.
• The combination allows for testing AI2MMUM's performance and generalization across both real-world and simulated 6G-relevant environments and modalities.

5.2. Evaluation Metrics

For each of the five downstream tasks, specific evaluation metrics are used:

1. Direct Positioning:

   • Conceptual Definition: Measures the accuracy of predicting the User Equipment (UE)'s position. Given that positioning can have varying degrees of error, it's common to look at the error distribution. CDF90 represents the error value below which 90% of the predicted positions fall. A smaller CDF90 indicates higher positioning accuracy.
   • Mathematical Formula: The positioning error for a single sample $i$ is typically the Euclidean distance between the predicted position $(\hat{x}_i, \hat{y}_i)$ and the true position $(x_i, y_i)$: $ e_i = \sqrt{(\hat{x}_i - x_i)^2 + (\hat{y}_i - y_i)^2} $ CDF90 is then found from the cumulative distribution function (CDF) of these errors. Let F(e) be the CDF of the errors, which gives the proportion of errors less than or equal to $e$. CDF90 is the value $e_{90}$ such that $F(e_{90}) = 0.90$. $ \mathrm{CDF90} = \min {e \mid P(\text{error} \le e) \ge 0.90 } $
   • Symbol Explanation:
     • $e_i$: Positioning error for sample $i$.
     • $(\hat{x}_i, \hat{y}_i)$: Predicted (x,y) coordinates of the UE for sample $i$.
     • $(x_i, y_i)$: True (x,y) coordinates of the UE for sample $i$.
     • $P(\text{error} \le e)$: The probability that the positioning error is less than or equal to $e$.

2. LOS/NLOS Identification:

   • Conceptual Definition: This is a classification task where the model predicts whether the communication path is Line-of-Sight (LOS) or Non-Line-of-Sight (NLOS). Classification accuracy measures the percentage of correctly predicted LOS/NLOS statuses out of all predictions.
   • Mathematical Formula: $ \text{Accuracy} = \frac{\text{Number of Correct Predictions}}{\text{Total Number of Predictions}} $
   • Symbol Explanation:
     • Number of Correct Predictions: The count of instances where the model's predicted LOS/NLOS status matches the true status.
     • Total Number of Predictions: The total number of samples evaluated.

3. MIMO Precoding:

   • Conceptual Definition: This task involves predicting the optimal precoding matrix to improve signal quality in MIMO systems. Squared Generalized Cosine Similarity (SGCS) is used to evaluate how closely the predicted precoding matrix aligns with the optimal (ground truth) precoding matrix. A higher SGCS indicates better precoding performance.
   • Mathematical Formula: For two complex vectors $\mathbf{u}$ and $\mathbf{v}$, the cosine similarity is $\frac{|\mathbf{u}^H \mathbf{v}|}{\|\mathbf{u}\| \|\mathbf{v}\|}$. For matrices, it often generalizes to a Frobenius norm-based similarity. Assuming the optimal precoding matrix is $\mathbf{W}_{\text{opt}}$ and the predicted precoding matrix is $\hat{\mathbf{W}}$, the SGCS can be defined as: $ \mathrm{SGCS}(\mathbf{W}{\text{opt}}, \hat{\mathbf{W}}) = \frac{\left| \hat{\mathbf{W}}^H \mathbf{W}{\text{opt}} \right|_F^2}{\left| \hat{\mathbf{W}} \right|F^2 \left| \mathbf{W}{\text{opt}} \right|_F^2} $
   • Symbol Explanation:
     • $\mathbf{W}_{\text{opt}}$: The optimal (ground truth) precoding matrix.
     • $\hat{\mathbf{W}}$: The predicted precoding matrix.
     • $(\cdot)^H$: Conjugate transpose (Hermitian transpose).
     • $\|\cdot\|_F$: Frobenius norm of a matrix, defined as $\|\mathbf{A}\|_F = \sqrt{\sum_{i=1}^m \sum_{j=1}^n |a_{ij}|^2}$. It essentially measures the "size" or magnitude of the matrix elements.

4. Beam Selection:

   • Conceptual Definition: In beam selection, the goal is to choose the best beam index from a predefined set (e.g., a DFT codebook). Top-1 accuracy measures the percentage of times the model correctly identifies the single best beam.
   • Mathematical Formula: Same as Classification Accuracy. $ \text{Top-1 Accuracy} = \frac{\text{Number of Correct Top-1 Beam Predictions}}{\text{Total Number of Beam Predictions}} $
   • Symbol Explanation:
     • Number of Correct Top-1 Beam Predictions: The count of instances where the model's highest-ranked predicted beam index matches the true optimal beam index.
     • Total Number of Beam Predictions: The total number of samples evaluated.

5. Path Loss Prediction:

   • Conceptual Definition: Path loss prediction is a regression task where the model estimates the signal power loss. Root Mean Squared Error (RMSE) is a common metric for regression, quantifying the average magnitude of the errors. It is the square root of the average of the squared differences between prediction and actual observation. Lower RMSE indicates higher accuracy.
   • Mathematical Formula: $ \mathrm{RMSE} = \sqrt{\frac{1}{N} \sum_{i=1}^{N} (\hat{y}_i - y_i)^2} $
   • Symbol Explanation:
     • $N$: The total number of samples.
     • $\hat{y}_i$: The predicted path loss value for sample $i$.
     • $y_i$: The true (actual) path loss value for sample $i$.

5.3. Baselines

The paper conducts ablation studies to evaluate the necessity and contribution of each module's innovative design. Six benchmarks are used for comparison, essentially representing different configurations or simplified versions of the proposed AI2MMUM:

1. Fixed Prompt (FP):

   • Description: This method uses only fixed task key descriptions (e.g., "position", "LOS status") as text input, without the learnable prefix prompts.
   • Purpose: To emphasize the contribution of the learnable prefix prompts in enhancing model expressiveness and guiding task-related feature extraction.

2. Same Prompt (SP):

   • Description: This method uses a single, identical instruction (comprising both fixed and learnable prompts, e.g., "user information") to perform all tasks.
   • Purpose: To highlight the importance of distinct, task-specific instructions for effectively guiding task-related feature extraction and avoiding task-agnostic features that can degrade performance, especially for complex outputs.

3. Train EPNN/CFENN (TE/TC):

   • Description: This method involves training the EPNN or CFENN (the multi-modal radio encoders) from scratch using the local dataset for the specific task, instead of using the robust pre-trained versions from large-scale multi-modal alignment.
   • Purpose: To underscore the robust representation capabilities gained by EPNN and CFENN through large-scale multi-modal alignment and to demonstrate the limitations of relying solely on limited local data for feature extraction.

4. Without LoRA (WL):

   • Description: In this method, the LLM backbone processes wireless data solely using its original pre-trained language knowledge, meaning LoRA matrices are not used for fine-tuning to incorporate domain-specific knowledge. The original LLM weights are preserved but not adapted.
   • Purpose: To demonstrate the role of LoRA in efficiently learning and integrating wireless domain-specific knowledge, thereby enhancing task performance beyond the LLM's inherent language understanding.

5. Random LLM (RL):

   • Description: This method uses a randomly initialized (instead of pre-trained) and frozen LLM backbone, which is then trained with LoRA for wireless tasks.
   • Purpose: To assess whether the underlying language knowledge within the pre-trained LLM backbone actually benefits communication task execution, or if the LoRA adaptation alone is sufficient. It tests the compatibility between language and wireless domain knowledge.

6. Without LLM (WM):

   • Description: This benchmark represents a traditional wireless AI method. It excludes the Task Instruction Module and the LLM backbone. Instead, the adapter layer (connecting the radio encoders) is directly connected to the task-specific heads for end-to-end supervised training.

   • Purpose: To highlight the advantages of leveraging the LLM backbone's generalization capability and the discriminative power of task instructions over traditional, specialized wireless AI models. This serves as a strong baseline to show the overall benefit of the AI2MMUM architecture.

     These benchmarks are strategically chosen to dissect the contributions of AI2MMUM's innovative components, providing a clear understanding of why each element is necessary for achieving superior performance.

6. Results & Analysis

6.1. Core Results Analysis

The experimental results demonstrate that the proposed AI2MMUM method consistently outperforms all six benchmarks across diverse modalities, datasets, and tasks. This robust performance validates the effectiveness of AI2MMUM's integrated design and its innovative modules.

The following are the results from Table I of the original paper:

(WC and PE denote wireless channel data and physical environment data, respectively.)

Let's analyze the performance trends shown in Figures 4 and 5, which present the performance of AI2MMUM and the six benchmarks across various tasks and datasets.

该图像是图表，展示了我们提出的方法与六个基准在信道基础的直接定位、LOS/NLOS 识别和 MIMO 预编码任务中的表现。左侧为 WAIR-D 区域 #00032，右侧为 DeepMIMO O1 BS#12。

Fig. 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.

该图像是图表，展示了我们的方法与六个基准在 WAIRD 区域 ext{#00247} 的环境基础波束选择和路径损耗预测任务上的性能对比。在波束选择任务中，我们的方法显示出88.00%的最高准确率；而在路径损耗预测任务中，表现出的 RMSE 约为 5.6 dB。

Fig. 5. The performance of our proposed method and six benchmarks across the environment-based beam selection and path loss prediction tasks in WAIRD area $\# 0 0 2 4 7$ .

Overall Superiority of AI2MMUM:

• In direct positioning (CDF90 error, lower is better), AI2MMUM achieves the lowest error across both WAIR-D and DeepMIMO. For instance, on WAIR-D #00032, it achieves roughly 1.5m CDF90 error, significantly better than the WM method's ~3.5m.

• For LOS/NLOS identification (classification accuracy, higher is better), AI2MMUM consistently shows the highest accuracy, nearing 90% on both datasets.

• In MIMO precoding (SGCS, higher is better), AI2MMUM outperforms, achieving an SGCS close to 0.95 on WAIR-D #00032.

• For beam selection (top-1 accuracy, higher is better), AI2MMUM reaches around 88% accuracy on WAIR-D #00247.

• Finally, in path loss prediction (RMSE, lower is better), AI2MMUM demonstrates the lowest RMSE of approximately 5.6 dB on WAIR-D #00247.

  These results strongly validate that AI2MMUM effectively integrates multi-modal features and task instructions through its LLM backbone, leading to superior performance.

Analysis of Benchmarks (Ablation Studies):

1. Fixed Prompt (FP) Method:

   • Performance: The FP method generally performs worse than AI2MMUM across all tasks. For example, its positioning CDF90 error is higher, and LOS/NLOS accuracy is lower.
   • Implication: This highlights the importance of the learnable prefix prompts. Explicitly training these prompts allows the model to capture more subtle and implicit task instructions, leading to better feature extraction and task fulfillment. Simply relying on fixed keywords is insufficient for optimal performance.

2. Same Prompt (SP) Method:

   • Performance: SP shows significant degradation, particularly for high-dimensional tasks like MIMO precoding and beam selection. For MIMO precoding, its SGCS is noticeably lower than AI2MMUM. For beam selection, its top-1 accuracy is substantially worse. However, for simpler, low-dimensional outputs like position or LOS status, the performance gap might be smaller but still present.
   • Implication: This confirms that distinct task instructions are crucial. When all tasks share the same instruction, the LLM backbone extracts task-agnostic features. While these might partially work for simple regression or binary classification, they become inadequate for complex, high-dimensional outputs, where biases from other tasks can severely degrade accuracy.

3. Train EPNN/CFENN (TE/TC) Method:

   • Performance: TE/TC consistently performs worse than AI2MMUM. For instance, its CDF90 positioning error is higher than AI2MMUM.
   • Implication: This underscores the limitations of training multi-modal radio encoders from scratch (TE/TC) using limited local data. The robust representation capabilities of the pre-trained EPNN and CFENN, achieved through large-scale multi-modal alignment on vast datasets, are essential for providing comprehensive insights into wireless characteristics and enabling strong generalization and task adaptability.

4. Without LoRA (WL) Method:

   • Performance: The WL method generally performs acceptably, but still worse than AI2MMUM. For example, AI2MMUM achieves better SGCS for MIMO precoding than WL.
   • Implication: This suggests that the original pre-trained telecom LLM backbone already possesses some capability for communication tasks due to its pre-training on a telecommunication corpus. However, LoRA plays a vital role in fine-tuning specific modules, allowing the LLM to more effectively absorb new wireless domain-specific knowledge. LoRA specifically adapts the LLM to the unique characteristics of wireless data and tasks, leading to further performance gains.

5. Random LLM (RL) Method:

   • Performance: RL shows significant performance degradation in most tasks compared to AI2MMUM. Its positioning error is much higher, and classification accuracies are lower.
   • Implication: This is a strong indicator that the inherent language knowledge within the pre-trained LLM backbone is indeed compatible with and beneficial for wireless domain knowledge. Random initialization of the LLM backbone severely hampers performance, even with LoRA adaptation, implying that the LLM's pre-trained understanding of patterns and relationships (even from language) is crucial. LoRA can partially compensate for the deficiencies of random initialization, but it cannot fully replace the rich language knowledge base.

6. Without LLM (WM) Method:

   • Performance: The WM method consistently yields the worst performance across all tasks and datasets. Its CDF90 positioning error is significantly higher (e.g., ~3.5m on WAIR-D #00032 vs. ~1.5m for AI2MMUM), and accuracies are notably lower.

   • Implication: This is the most critical benchmark, representing traditional wireless AI approaches that lack an LLM backbone and explicit task instructions. Its poor performance definitively demonstrates that AI2MMUM's core innovation—leveraging the LLM backbone's generalization capability and the discriminative power of task instructions—is essential. Traditional methods struggle to adapt to multiple tasks with high precision because they lack the contextual understanding and flexibility provided by the LLM.

     In conclusion, the comprehensive evaluations and ablation studies clearly validate that each component of AI2MMUM contributes significantly to its SOTA performance. The integration of pre-trained multi-modal encoders, task-aware instructions, an LLM backbone fine-tuned with LoRA, and lightweight task-specific heads creates a powerful and flexible universal model for 6G wireless systems.

7. Conclusion & Reflections

7.1. Conclusion Summary

This paper successfully proposes AI2MMUM, an Artificial Intelligence-Air Interface Multi-Modal Universal Model, designed to address the growing complexity and diverse requirements of 6G wireless systems. AI2MMUM leverages the generalization capability of Large Language Models (LLMs) and the discriminative power of task instructions to process 6G-oriented multi-modal data (physical environment and wireless channel information) and flexibly execute a variety of downstream tasks. The framework meticulously integrates four key components: a multi-modal radio feature extraction module (employing pre-trained EPNN and CFENN with adapter layers), a task instruction module (combining fixed keywords and learnable prefix prompts), a telecom LLM backbone enhanced with LoRA for domain adaptation, and lightweight task-specific heads for direct output. Extensive ablation experiments demonstrate AI2MMUM's superior performance, consistently outperforming traditional non-LLM methods and models lacking its innovative design across five representative tasks: direct positioning, LOS/NLOS identification, MIMO precoding, beam selection, and path loss prediction. These compelling results highlight the profound compatibility and synergistic potential between radio and language knowledge, paving the way for unified wireless multi-modal intelligence.

7.2. Limitations & Future Work

The authors highlight several implicit and explicit directions for future work, primarily focusing on scalability and further generalization:

• Scalability to New Modalities: The paper mentions that when new modalities such as radar and LiDAR are introduced, only the adapter layers need to be updated while the corresponding modality encoders remain frozen. This implies a future direction to integrate more diverse 6G modalities beyond just physical environment and wireless channel data, potentially expanding the scope of AI2MMUM.

• API Encapsulation and Function Calls: The authors suggest that task-specific heads can be encapsulated within diverse Application Programming Interfaces (APIs), allowing AI2MMUM to identify and invoke the appropriate API based on task instructions and function calls. This points towards future work in developing a more sophisticated API management layer and function-calling capabilities for AI2MMUM, enabling it to interact with external systems and tools.

• Further Domain-Specific Knowledge Integration: While LoRA is effective, continued exploration of more advanced and efficient fine-tuning techniques or new pre-training strategies for the telecom LLM could further enhance its ability to absorb and reason with wireless domain-specific knowledge.

  Potential limitations (not explicitly stated by authors but inferable):

• Computational Resources for Training: Although LoRA reduces fine-tuning parameters, the initial pre-training of the teleMA LLaMA2-7B on a telecommunication corpus (which precedes AI2MMUM's specific training) would have required substantial computational resources. The overall energy consumption and carbon footprint of such large models are ongoing concerns.

• Data Availability and Diversity: The success hinges on large-scale, diverse, and high-quality multi-modal datasets. While WAIR-D and DeepMIMO are used, collecting and curating sufficiently diverse real-world 6G data across all envisioned modalities remains a significant challenge.

• Real-time Performance for 6G: 6G applications often demand extremely low latency. While the single-pass approach for task-specific heads aims to reduce inference time, the overall latency of a large LLM backbone for critical real-time physical layer tasks in dynamic 6G environments needs further investigation.

• Interpretability and Trustworthiness: LLMs are often considered "black boxes." For critical wireless infrastructure, understanding why a model makes a particular precoding or beam selection decision is vital for reliability, fault diagnosis, and regulatory compliance. The paper does not delve into the interpretability of AI2MMUM.

7.3. Personal Insights & Critique

This paper presents a highly significant step towards realizing the AI-native 6G vision. The idea of leveraging LLMs as a backbone for a universal wireless model is intuitive and powerful, given LLMs' inherent generalization and contextual understanding capabilities. The careful integration of multi-modal radio encoders and task instruction modules is particularly clever, bridging the gap between raw wireless data and the LLM's symbolic reasoning.

Transferability and Application: The methods and conclusions of AI2MMUM have strong potential for transferability. The LoRA approach, in particular, is a game-changer for adapting large foundation models to specialized domains without prohibitive retraining costs. This principle could be applied to other vertical industries where LLMs need to interact with diverse sensor data and execute domain-specific commands (e.g., smart manufacturing, smart agriculture, robotics, or even medical imaging paired with diagnostic instructions). The core idea of "instruction-aware multi-modal AI" is broadly applicable.

Potential Issues/Areas for Improvement:

1. Complexity of Instruction Engineering: While learnable prefix prompts are introduced, the process of task instruction engineering (both fixed keywords and optimizing learnable prompts) could still be complex. As the number of tasks and modalities grows, managing and designing effective instructions might become a challenge. A more explicit framework for automated instruction generation or dynamic prompt optimization could be beneficial.

2. Robustness to Adversarial Attacks: In wireless communication, especially for critical infrastructure, AI models are vulnerable to adversarial attacks. The robustness of LLM-based wireless models to malicious inputs or corrupted CSI data is an important area not discussed.

3. Cross-Lingual Instructions: While LLMs are generally multi-lingual, the paper implies English task instructions. Extending this to support multi-lingual instructions could enhance global applicability.

4. Beyond CSI and Environment Data: While the paper mentions radar and LiDAR as future modalities, the current evaluation is primarily focused on CSI and physical environment data. A deeper dive into how AI2MMUM specifically handles the unique characteristics and challenges of other modalities (e.g., temporal dynamics of video, sparse nature of LiDAR point clouds) would be valuable. The adapter layers are designed for this, but the actual performance with these new modalities remains to be seen.

5. Energy Efficiency and Green AI: The trend towards Large AI models raises concerns about energy consumption. While LoRA helps during fine-tuning, the operational energy cost of continuously running such a large model for 6G tasks, especially at the edge, needs to be considered for truly sustainable AI-native networks.

   Overall, AI2MMUM offers a compelling vision and a robust initial architecture for AI-native 6G. The paper's strength lies in its systematic approach to integrating advanced AI techniques to solve complex wireless problems, providing a strong foundation for future research in unified multi-modal intelligence in communication systems.