PruneTIR: Inference-Time Tool Call Pruning for Effective yet Efficient Tool-Integrated Reasoning

Researchers introduce PruneTIR, an inference-time optimization framework that improves tool-integrated reasoning in large language models by pruning failed trajectories, resampling tool calls, and suspending tool usage when errors persist. The approach enhances LLM performance without requiring additional training, demonstrating significant improvements in accuracy and efficiency.

PruneTIR addresses a critical gap in tool-integrated reasoning optimization. While extensive research has focused on enabling LLMs to use external tools like code interpreters, the framework tackles a less-explored problem: improving reasoning quality during inference once models already possess tool capabilities. This distinction matters because inference-time optimizations provide immediate performance gains without computational costs of retraining.

The research identifies a key failure pattern in tool-capable LLMs: erroneous tool calls accumulate during reasoning chains, creating compounding errors that models struggle to recover from even with additional attempts. By observing that most recoverable errors resolve within a few turns while persistent errors rarely resolve regardless of additional attempts, PruneTIR implements targeted interventions. The three-component system works synergistically—pruning unsuccessful trajectories prevents wasted computation, resampling generates alternative tool calls to escape local failure states, and suspension recognizes when tool use becomes counterproductive.

For the AI development community, PruneTIR demonstrates that tool-integrated reasoning optimization can be achieved through intelligent trajectory management rather than architectural changes or fine-tuning. The efficiency gains—reduced context length and improved Pass@1 metrics—are particularly valuable as LLMs scale to handle increasingly complex multi-step reasoning tasks. This approach parallels broader trends in AI optimization, where inference-time techniques like speculative decoding and dynamic batching extract additional performance from existing models.

Looking ahead, this research signals that tool-integrated reasoning remains an active optimization frontier. The framework's success suggests future work may explore adaptive pruning strategies, learned suspension policies, and integration with other inference-time optimization techniques to maximize both accuracy and computational efficiency.

• →PruneTIR improves tool-integrated reasoning at inference time without requiring model retraining or fine-tuning.
• →The framework identifies that erroneous tool calls either resolve within a few turns or persist indefinitely, enabling targeted intervention strategies.
• →Three mechanisms—success-triggered pruning, stuck-triggered resampling, and retry-triggered suspension—collectively mitigate cascading tool-use errors.
• →Experimental results show significant improvements in Pass@1 accuracy while reducing computational overhead and context length requirements.
• →Inference-time optimization of tool use represents an underexplored but high-impact frontier for improving LLM reasoning capabilities.