Toolweaver Achieves Scalable Tool Use with Large Language Models and Unified Selection

Researchers are tackling a critical limitation in large language model (LLM) tool use: the difficulty of scaling to a large number of tools without sacrificing semantic understanding. Bowen Fang, Wen Ye, and Yunyue Su, from the New Laboratory of Pattern Recognition (NLPR), Institute of Automation, Chinese Academy of Sciences (CASIA), alongside Jinghao Zhang, Qiang Liu, and Yesheng Liu et al., present ToolWeaver, a novel framework which encodes tools into hierarchical sequences rather than assigning each a unique token. This innovative approach circumvents the vocabulary explosion and semantic isolation common in existing methods, allowing LLMs to learn collaborative relationships between tools more effectively from dense co-occurrence patterns. Demonstrating significant performance gains across nearly 47,000 tools, ToolWeaver establishes a more scalable and semantically-aware foundation for building advanced, tool-augmented agents.

This innovative approach circumvents the vocabulary explosion and semantic isolation common in existing methods, allowing LLMs to learn collaborative relationships between tools more effectively from dense co-occurrence patterns.

Hierarchical Encoding for Scalable Tool Use

Current tool-use pipelines often struggle with a dual semantic challenge: encoders frequently fail to capture complex meanings, and LLMs lack inherent knowledge of the tools themselves. Generative methods, which task the LLM with directly learning and generating tool identifiers, offer a promising alternative, but commonly map each tool to a unique token, creating scalability and generalisation crises as vocabulary size expands. This approach also hinders the learning of relationships between tools, relying on sparse co-occurrences of isolated tool IDs. The resulting codes are then integrated into the LLM through a generative alignment stage, where the model is fine-tuned to produce these hierarchical code sequences.

For example, tools like ‘Realtime Weather’ and ‘Air Quality’ can share a parent code grouping them under a shared context like “outdoor conditions”, enabling the model to infer their relationship from the frequent co-occurrence of the shared code. This approach overcomes the limitations of previous methods by enabling the LLM to natively generate hierarchical code sequences for complex tool invocation. The team’s findings pave the way for more robust and intelligent agents capable of complex reasoning and real-world problem-solving.

Hierarchical Encoding for Scalable Tool Use

The research team identified that prevalent retrieval-based methods often employ encoders incapable of capturing nuanced semantics, while Large language models lack inherent tool knowledge from pre-training. This approach hinders generalisation and creates a semantic bottleneck, impeding the learning of relationships between tools. Experiments employed nearly 47,000 tools to rigorously evaluate ToolWeaver’s performance. The team’s innovative tokenisation process weaves together intrinsic tool semantics with extrinsic co-usage patterns, allowing the LLM to better understand and utilise tools in complex scenarios.

This work establishes a new paradigm for tool learning, moving beyond simple token mapping to a more nuanced and scalable representation. The research highlights the importance of encoding collaborative relationships between tools, enabling the LLM to reason more effectively and provide more comprehensive responses, as demonstrated by the example of integrating weather and air quality checks for a single query. The resulting framework offers a substantial advancement in the field of AI agents, paving the way for more powerful and versatile applications.

Logarithmic scaling enables vast tool integration and data

The research tackles the dual challenge of capturing complex semantics and a lack of intrinsic tool knowledge within Large Language Models (LLMs). Experiments reveal that ToolWeaver achieves logarithmic vocabulary expansion, a significant improvement over the linear growth experienced with traditional one-token-per-tool methods. This scalability is crucial, as integrating a benchmark of nearly 47,000 tools into a model like Llama-3-8B, with its existing vocabulary of 128,256 tokens, would otherwise require adding a substantial number of out-of-vocabulary tokens. For example, tools like Realtime Weather and Air Quality, when queried together, can share a parent code indicating “outdoor conditions”, allowing the model to infer their relationship and provide more comprehensive answers.

This contrasts with previous methods where the model might only check the weather, failing to consider air quality. Measurements confirm that the framework overcomes the semantic bottleneck inherent in assigning unique tokens to each tool, allowing for a richer understanding of tool relationships. Scientists recorded that the collaborative-aware vector quantization process generates these hierarchical structures in an unsupervised manner. This process encourages functionally related tools to share codes, fostering a dense collaborative signal within the model. The work establishes a new paradigm for tool representation, moving beyond flat identifiers to compositional codes that enhance both scalability and reasoning capabilities.

Hierarchical Coding Improves Tool Use Efficiency by reducing

Current tool-use systems often struggle because they assign unique, isolated tokens to each tool, leading to an unmanageable vocabulary size and hindering the model’s ability to understand relationships between tools. Importantly, the framework preserves the model’s core linguistic abilities, as evidenced by consistent summarisation performance and minimal impact on general understanding benchmarks. The authors acknowledge that the ToolBench dataset used in their evaluations was not specifically audited for biases or privacy risks, highlighting the need for careful consideration in real-world deployments. Future research directions include utilising reinforcement learning to enable the autonomous discovery of collaborative patterns between tools.