SING: Synthetic Intention Graph for Scalable Active Tool Discovery in LLM Agents

2026-06-15

AI Summary

Existing tool retrieval methods in large language model agents struggle to adapt to dynamically evolving task intents, resulting in low efficiency and poor generalization. This work proposes the first active tool discovery framework that integrates intent awareness with graph structure, constructing an intent-tool graph to jointly model user intent, tool capabilities, and collaborative relationships through graph neural networks and a dynamic retrieval mechanism. By incorporating task decomposition, observational feedback, and subgoal generation, the approach enables context-aware tool selection. It overcomes the limitations of static tool repositories and supports efficient open-world expansion. Evaluated on three real-world benchmarks, the method achieves up to a 59.8% improvement in Global Recall@5, a 28.9% increase in downstream task success rate, and reduces full-tool exposure by 99.8%.

Abstract

Large language model (LLM) agents increasingly rely on agent harnesses that manage context, tools, and multi-turn execution, making tools a central interface for acting in realistic digital environments. As harness-connected tool ecosystems expand to hundreds or thousands of APIs, services, and task-specific skills, exhaustive tool schema injection becomes costly and imposes a closed-world assumption that limits agents to a predefined static inventory. Retrieval-augmented tool selection offers a natural alternative, but existing one-shot retrieval methods often fail to align isolated tool descriptions with the agent's true task intention, especially in long-horizon tasks where required capabilities emerge through decomposition, observations, and newly induced subgoals. We propose SING, an intention-aware active tool discovery framework that builds an intention-tool graph linking user intentions, tool capabilities, and tool collaboration patterns, and dynamically retrieves tools according to evolving task states. Using a unified corpus of 7,471 tools, we evaluate SING on three real-world tool-use benchmarks. SING improves Global Recall@5 by up to 59.8% and downstream success rate by up to 28.9% over baselines, while reducing full-corpus tool-schema exposure by 99.8%, demonstrating that intention-aware graph structure enables more accurate and context-efficient tool discovery in large-scale agentic ecosystems.

Problem

• tool discovery
• large language model agents
• intention alignment
• scalable tool retrieval
• long-horizon tasks

Innovation

• intention-aware retrieval
• tool discovery
• LLM agents
• graph-based reasoning
• scalable tool ecosystems