Title: TInR: Exploring Tool-Internalized Reasoning in Large Language Models

Abstract: The integration of external tools into the reasoning processes of Large Language Models (LLMs) has established Tool-Integrated Reasoning (TIR) as a highly promising avenue. Nevertheless, conventional TIR approaches generally depend on accessing external tool documentation during the reasoning phase. This dependency introduces significant challenges, including difficulties in mastering tools, limitations on tool size, and inefficient inference speeds. To overcome these obstacles, this study investigates Tool-Internalized Reasoning (TInR), a paradigm designed to embed tool knowledge directly within LLMs to facilitate reasoning. Realizing this objective entails addressing two critical challenges: tool internalization and the coordination between tool usage and reasoning. In response, we introduce TInR-U, a framework that unifies reasoning and tool application through internalized tool knowledge. The training of TInR-U follows a three-stage pipeline: first, tool internalization is achieved via a bidirectional knowledge alignment strategy; second, a supervised fine-tuning warm-up is conducted using high-quality reasoning annotations; and third, reinforcement learning is applied with rewards specifically tailored to TInR. Our comprehensive evaluation, covering both in-domain and out-of-domain scenarios, demonstrates that TInR-U delivers superior performance in both contexts, underscoring its effectiveness and efficiency.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC