PruneTIR: Inference-Time Tool Call Pruning for Effective yet Efficient Tool-Integrated Reasoning
Researchers have introduced PruneTIR, a novel framework designed to enhance the reasoning capabilities of Large Language Models (LLMs) during inference without requiring additional training. While recent studies focus on equipping LLMs with tool-use abilities, PruneTIR addresses the underexplored area of optimizing already tool-capable models at inference time. The framework is built on the observation that erroneous tool calls negatively correlate with answer correctness and often lead to unresolved errors if not addressed quickly. PruneTIR employs three key components: Success-Triggered Pruning, Stuck-Triggered Pruning and Resampling, and Retry-Triggered Tool Suspension. These mechanisms allow the system to prune ineffective trajectories, resample tool calls, and suspend tool usage when necessary, thereby preventing LLMs from getting stuck in repeated failed attempts. Extensive experimental results demonstrate that PruneTIR significantly improves Pass@1 accuracy and operational efficiency while reducing the working context length. This development represents a significant step forward in making tool-integrated reasoning more effective and efficient for complex problem-solving tasks involving external tools like code interpreters.
Wire timeline
PruneTIR: Inference-Time Tool Call Pruning for Effective yet Efficient Tool-Integrated Reasoning
Researchers have introduced PruneTIR, a novel framework designed to enhance the reasoning capabilities of Large Language Models (LLMs) during inference without requiring additional training. While recent studies focus on equipping LLMs with tool-use abilities, PruneTIR addresses the underexplored area of optimizing already tool-capable models at inference time. The framework is built on the observation that erroneous tool calls negatively correlate with answer correctness and often lead to unresolved errors if not addressed quickly. PruneTIR employs three key components: Success-Triggered Pruning, Stuck-Triggered Pruning and Resampling, and Retry-Triggered Tool Suspension. These mechanisms allow the system to prune ineffective trajectories, resample tool calls, and suspend tool usage when necessary, thereby preventing LLMs from getting stuck in repeated failed attempts. Extensive experimental results demonstrate that PruneTIR significantly improves Pass@1 accuracy and operational efficiency while reducing the working context length. This development represents a significant step forward in making tool-integrated reasoning more effective and efficient for complex problem-solving tasks involving external tools like code interpreters.
cs.AI updates on arXiv.org