LAMeTA: Intent-Aware Agentic Network Optimization via a Large AI Model-Empowered Two-Stage Approach

TL;DR

LAMeTA addresses intent-aware optimization in agentic networks by combining a compact intent-distillation stage (IoKD) with a symbiotic reinforcement learning stage (SRL). IoKD transfers intent understanding from a centralized, large LAM to lightweight edge LAMs, enabling accurate interpretation of natural-language user prompts with reduced compute. SRL then jointly optimizes GenSFC composition and E-LAM selection via a graph-based MDP, using E-LAM translations of intents to shape state representations and rewards, while leveraging in-context learning for adaptability. Experiments on an 81-agent setup show IoKD reduces intent-prediction error by up to 22.5%, and SRL outperforms baselines by up to 23.5% in maximizing intent-aware QoE, demonstrating scalable, practical gains for diverse user intents in networked GenAI services.

Abstract

Nowadays, Generative AI (GenAI) reshapes numerous domains by enabling machines to create content across modalities. As GenAI evolves into autonomous agents capable of reasoning, collaboration, and interaction, they are increasingly deployed on network infrastructures to serve humans automatically. This emerging paradigm, known as the agentic network, presents new optimization challenges due to the demand to incorporate subjective intents of human users expressed in natural language. Traditional generic Deep Reinforcement Learning (DRL) struggles to capture intent semantics and adjust policies dynamically, thus leading to suboptimality. In this paper, we present LAMeTA, a Large AI Model (LAM)-empowered Two-stage Approach for intent-aware agentic network optimization. First, we propose Intent-oriented Knowledge Distillation (IoKD), which efficiently distills intent-understanding capabilities from resource-intensive LAMs to lightweight edge LAMs (E-LAMs) to serve end users. Second, we develop Symbiotic Reinforcement Learning (SRL), integrating E-LAMs with a policy-based DRL framework. In SRL, E-LAMs translate natural language user intents into structured preference vectors that guide both state representation and reward design. The DRL, in turn, optimizes the generative service function chain composition and E-LAM selection based on real-time network conditions, thus optimizing the subjective Quality-of-Experience (QoE). Extensive experiments conducted in an agentic network with 81 agents demonstrate that IoKD reduces mean squared error in intent prediction by up to 22.5%, while SRL outperforms conventional generic DRL by up to 23.5% in maximizing intent-aware QoE.

Figures (17)

• Figure 1: The illustration of diverse applications and the corresponding preferences in the agentic network (left). The overall workflow of LAMeTA and the illustration of a $3$-step GenSFC for report writing (right). In agentic networks, the intents are incorporated into prompts explicitly or implicitly. For example, example 2 explicitly requires high capability and low latency for 3D rendering. However, the requirement for high capability (to write insightful reports) is implicitly expressed in example 3. Both explicit and implicit intents should be captured and considered.
• Figure 2: The illustration of LAMeTA for agentic network optimization. The items with different shapes and colors represent intents belonging to different applications. We can observe that applications 1 and 2 refer to business and gaming, respectively.
• Figure 3: The illustration of IoKD. The left part demonstrates the process of adaptive dataset construction, which is the prerequisite of IoKD. The right part demonstrates the IoKD process.
• Figure 4: The illustration of SRL architecture. Note that the policy optimization module is pluggable and can be implemented by diverse policy-based RL algorithms.
• Figure 5: The illustration of user preferences for two applications and the effectiveness of dataset augmentation.