arXiv cs.AIFriday · May 29, 2026FREE

ParaTool: Shifting Tool Representations from Context to Parameters

ParaTool, a new framework detailed in an arXiv paper published on May 29, 2026, addresses key limitations in large language model (LLM) tool-calling. Current in-context learning (ICL) methods often incorporate extensive tool documentation and usage examples directly into the LLM's context. This practice results in substantial inference overhead and increased risks of hallucination as context lengths grow. While tuning-based methods can improve general tool-calling, they frequently fail to effectively internalize specific tool details, maintaining a persistent dependency on in-context documentation. ParaTool proposes to project each tool into a dedicated, loadable set of parameters, allowing LLMs to perform tool calling without relying on in-context documents or examples. The framework operates in three stages: parametric tool pre-training, which encapsulates the knowledge of various tools into independent parameter modules; soft tool selection; and dynamic integration of these parameterized tools. This approach aims to provide a more efficient and robust mechanism for LLMs to interact with external executable interfaces, thereby supporting environment-coupled problem solving by decoupling tool knowledge from the immediate context and improving overall performance.