SING framework boosts agent tool discovery by up to 59.8% recall
SING, a new intention-aware active tool discovery framework, improves LLM agent tool retrieval by building a dynamic graph linking user intentions, tool capabilities, and collaboration patterns across a corpus of 7,471 tools.
Score breakdown
SING reduces full-corpus tool-schema exposure by 99.8% while simultaneously improving retrieval recall and task success, directly addressing the context-cost and closed-world limitations that arise as agentic tool ecosystems scale to thousands of APIs.
- 01SING stands for Synthetic Intention Graph, an intention-aware active tool discovery framework for LLM agents.
- 02It builds an intention-tool graph linking user intentions, tool capabilities, and tool collaboration patterns.
- 03Tools are retrieved dynamically according to evolving task states, not in a single one-shot retrieval.
LLM agents operating in realistic digital environments increasingly depend on harnesses that manage context, tools, and multi-turn execution. As these harness-connected ecosystems scale to hundreds or thousands of APIs, services, and task-specific skills, exhaustive tool schema injection grows costly and imposes a closed-world assumption that locks agents into a predefined static inventory. Existing retrieval-augmented approaches attempt to address this, but one-shot retrieval methods frequently fail to align isolated tool descriptions with the agent's true task intention — particularly in long-horizon tasks where required capabilities only emerge through decomposition, observations, and newly induced subgoals.
Rather than retrieving tools once at task start, SING dynamically updates retrieval according to evolving task states.
SING (Synthetic Intention Graph) proposes an intention-aware active tool discovery framework that builds a graph structure linking user intentions, tool capabilities, and tool collaboration patterns. Rather than retrieving tools once at task start, SING dynamically updates retrieval according to evolving task states. Authored by Qiao Xiao, Haochen Shi, and Yisen Gao, the paper evaluates SING on three real-world tool-use benchmarks using a unified corpus of 7,471 tools. 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 at scale.
Key facts
- 01SING stands for Synthetic Intention Graph, an intention-aware active tool discovery framework for LLM agents.
- 02It builds an intention-tool graph linking user intentions, tool capabilities, and tool collaboration patterns.
- 03Tools are retrieved dynamically according to evolving task states, not in a single one-shot retrieval.
- 04Evaluated on a unified corpus of 7,471 tools across three real-world tool-use benchmarks.
- 05SING improves Global Recall@5 by up to 59.8% over baselines.
- 06Downstream task success rate improves by up to 28.9% over baselines.
- 07Full-corpus tool-schema exposure is reduced by 99.8%, cutting context costs significantly.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 16, 2026 · 23:11 UTC. How this works →