JAUNT: Joint Alignment of User Intent and Network State for QoE-centric LLM Tool Routing Conference: JAUNT; JAUNT; 2025 CCS: Information systems Language models CCS: Networks Network management

Enhan Li, Hongyang Du · 2025

Read JAUNT: Joint Alignment of User Intent and Network State for QoE-centric LLM Tool Routing Conference: JAUNT; JAUNT; 2025 CCS: Information systems Language models CCS: Networks Network management by Enhan Li, Hongyang Du (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-18550.txt.

Introduction

Enhan Li and Hongyang Du look at how language models choose which external tool to call for a user's request. Matching the words of a request to a tool's function is not enough, they argue. Requests are often vague, and the user's real experience also depends on network delay and whether servers are actually available, facts that pure meaning matching cannot see.

Their framework, JAUNT, aligns two views at once. It interprets what the user intends, and it uses model agents to turn raw network numbers into profiles written in the same semantic space, so routing can weigh both together. The authors also built a benchmark that pairs varied request styles with mixed network conditions. In tests, JAUNT improved quality of experience over several baselines, supporting their claim that scalable tool services must route by intent and network state together.

Catalog record

CREATOREnhan Li, Hongyang Du
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE41 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-18550

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

JAUNT: Joint Alignment of User Intent and Network State for QoE-centric LLM Tool Routing Conference: JAUNT; JAUNT; 2025 CCS: Information systems Language models CCS: Networks Network management Enhan Li, Hongyang Du Abstract . Large Language Models (LLMs) increasingly rely on emerging protocols such as the Model Context Protocol (MCP) to invoke external tools and services. However, current tool routing mechanisms remain fragile because they only consider functional matching between users’ queries and tools. In practice, user intent expressed through queries can be vague or underspecified, and the actual Quality of Experience (QoE) also depends on external factors such as link latency and server availability that are not captured by semantics alone. To address this challenge, we propose JAUNT, a framework for J oint A lignment of U ser intent and N etwork state in QoE-centric T ool routing. JAUNT introduces a dual-view alignment strategy that interprets user intent while employing LLM agents to construct network profiles, mapping numerical performance indicators into the semantic space to guide routing. We further design a benchmark that integrates diverse user request patterns with heterogeneous network states, enabling systematic evaluation of QoE outcomes. Experimental results show that JAUNT significantly improves QoE compared with several baselines, demonstrating the importance of aligning both intent and network state for scalable LLM service orchestration. 1. Introduction Large Language Models (LLMs) have evolved from text-based conversational systems ( Dubey et al. 2024 ; Achiam et al. 2023 ; Yang et al. 2025 ) into general-purpose intelligence engines interacting with external tools and services ( Zhuang et al. 2023 ; Nam et al. 2024 ; Chang et al. 2024 ; Ferrag et al. 2025 ; Li 2025 ) . This transition reflects an inevitable trajectory toward integration, where LLMs no longer operate in isolation but act as coordinators across heterogeneous digital resources. Emerging protocols such as the Model Context Protocol (MCP) ( to Model Context Protocol MCP ; Hou et al. 2025 ; Singh et al. 2025 ) have enabled LLMs to connect seamlessly with applications such as booking, design, learning, and data analysis within a unified conversational interface. Such integration transforms LLMs into a new operating paradigm that manages user intents, orchestrates tools, and delivers end-to-end task execution through natural dialogue ( Yang et al. 2023 ) . As this ecosystem expands, the ability of LLMs to understand user goals and invoke the right external services becomes a fundamental capability shaping the next generation of AI-driven platforms, analogous to an intelligent operating system for human–AI interaction ( Shah et al. 2025 ; Bodonhelyi et al. 2024 ) . While external tool integration greatly expands the functional capacity of LLMs, it also reveals the significant role of the network as both an enabler and a constraint of intelligence. Modern network infrastructure enables LLMs running on a host device to access diverse remote services through distributed computation, allowing a single model to compose tools across cloud platforms, knowledge bases, and specialized APIs in real time and thus form an open and scalable ecosystem unattainable by standalone models. However, this advantage comes with inherent dependence on network conditions. Each tool invocation must traverse communication links and remote servers whose latency, bandwidth, and congestion vary with physical distance and traffic load. These variations influence the task response speed and the stability of task execution. For example, routing a user request to a distant or heavily loaded server may introduce substantial delay or timeout, degrading user-perceived Quality of Experience (QoE) ( Jain 2004 ) despite the correct semantic match. However, current orchestration protocols, such as MCP, still treat the network as an ideal and static medium, overlooking its dynamic impact on service quality. Figure 1. Motivation of JAUNT: Intent is obscured by semantic ambiguity, emotional expression, and network trade-offs. Incorporating network awareness addresses part of the orchestration challenge but does not fully determine the QoE. The fundamental purpose of LLM-based systems is to serve diverse users, yet current routing methods often assume that user queries are clear and complete. In practice, this assumption rarely holds ( Shah et al. 2025 ) . Public-facing LLM systems receive inputs that vary widely in specificity, tone, and contextual precision, leading to semantic ambiguity followed by inaccurate tool selection ( Wang et al. 2025b ) . More importantly, users’ emotional expression in their queries inherently encodes expectations about network responsiveness and computational effort. Different user personalities imply distinct tolerances for latency, accuracy, and interaction cost ( Du et al. 2023 ) . For instance, a frustrated or time-sensitive user implicitly values low delay and immediate feedback, while an accuracy-focused user may prefer detailed responses even at longer execution times. As shown in Fig. 1 , these latent expectations link user intent with network states, making them inseparable when maximizing users’ QoE. However, current orchestration frameworks treat semantic understanding and network execution as independent processes, overlooking how emotional tone and behavioral patterns shape sensitivity to service performance. Bridging this gap requires a unified perspective that aligns user intent with dynamic network states to achieve effective and adaptive tool routing. Addressing the above research gap requires tackling two challenges. The first is the lack of datasets that jointly reflect user intent diversity and network variability. Existing benchmarks mainly evaluate tool routing under well-defined queries and stable connections ( Guo et al. 2025 ; Gao et al. 2025 ; Wang et al. 2025a ; Fei et al. 2025 ) , and while some works assess tool routing performance across diverse network conditions ( Li et al. 2025 ) , they fail to capture how real users express vague or emotionally biased requests that implicitly affect routing outcomes. The second challenge lies in LLM tool routing algorithmic integration, where semantic analysis of the user query must be aligned with network states to optimize end-to-end QoE. To address these challenges, we first construct the Tool-Routing Intent and Persona (TRIP) benchmark for modeling user diversity. TRIP includes user profiles capturing behavioral traits and queries generated from these profiles across common application scenarios such as travel planning and information retrieval. Each query is augmented with linguistic fuzziness and emotional tone to reflect realistic user intent and expression variations. Building upon this, we leverage the intrinsic understanding capability of LLMs to jointly interpret users’ long-term profiles, short-term emotional queries, and real-time network states to construct JAUNT, a framework for achieving J oint A lignment of U ser intent and N etwork state in QoE-centric T ool routing. Our contributions are summarized as • We construct the TRIP benchmark to fill the current gap in evaluating LLM-based tool routing under diverse user intents. TRIP provides structured user profiles and intent-driven query sets that capture realistic variations in linguistic fuzziness, tone, and behavioral preferences, enabling systematic analysis of subjective QoE in tool orchestration. • We propose the JAUNT framework, which achieves dual-view alignment between semantic intent and network performance. An LLM agent infers both long-term intent from user profiles and short-term intent from current queries, then combines this with real-time network information to rank candidate tools and select the one expected to maximize user QoE. • We conduct extensive experiments demonstrating that JAUNT balances task success rate, latency, and reliability to achieve high QoE, validating the importance of jointly

Read the complete text