LLM Agent Architecture

Tool Learning with Frozen Language Models is rapidly emerging as a scalable strategy to empower LLMs with real-world functionality. This article introduces Chain-of-Tools (CoTools), a novel approach that enables frozen language models to reason using external tools—without modifying their weights. CoTools leverages the model’s hidden states to determine when and which tools to invoke, generalizing to massive pools of unseen tools through contrastive learning and semantic retrieval. It outperforms traditional fine-tuning and in-context learning approaches across numerical and knowledge-based tasks. The article also explores interpretability insights, showing how only a subset of hidden state dimensions drives tool reasoning. CoTools maintains the original model’s reasoning ability while expanding its practical scope, making it ideal for building robust, extensible LLM agents. Whether you’re designing enterprise AI systems or exploring advanced LLM capabilities, this is a definitive resource on scalable, efficient, and interpretable Tool Learning with Frozen Language Models.