PruneTIR: Improving Tool-Integrated Reasoning Efficiency of Large Language Models via Inference-Time Tool-Call Pruning

The PruneTIR framework significantly improves the reasoning efficiency and accuracy of tool-augmented LLMs through three inference-time optimization strategies—success-triggered pruning, stuck-triggered pruning and resampling, and retry-triggered tool pausing—without additional training. Addressing the gap of neglecting inference-time optimization in tool-integrated reasoning, it provides a highly cost-effective solution for practical applications.

Tool-Integrated Reasoning (TIR) of Large Language Models (LLMs) can break through the limitations of parametric knowledge and solve complex problems. However, current research mostly focuses on enabling LLMs to learn to use tools, while neglecting how LLMs that already have tool-using capabilities can utilize tools more efficiently and accurately during inference. Inference-time optimization requires no additional training costs and can directly improve practical application performance. The PruneTIR framework is proposed precisely to address this gap.

The research team discovered two patterns: 1. The rate of incorrect tool calls is significantly negatively correlated with the correctness of the final answer; reducing incorrect call chains can improve reasoning quality. 2. There is a "golden time window" for error recovery—beyond this window, the model tends to fall into the dilemma of repeated attempts, lacking metacognitive ability to monitor its own reasoning trajectory.

PruneTIR consists of three components: 1. Success-triggered pruning: Prune other paths after finding a feasible solution. 2. Stuck-triggered pruning and resampling: Monitor the error recovery window; if it times out, abandon the failed trajectory and resample. 3. Retry-triggered tool pausing: Temporarily disable relevant tools after multiple failed retries. The workflow is monitoring → evaluation → decision-making → execution. All operations are completed during inference without fine-tuning the model, ensuring strong generality.

In benchmark tests, PruneTIR showed significant performance: 1. The Pass@1 metric improved, making the model more likely to take the correct path. 2. The number of reasoning steps decreased, reducing latency and token consumption. 3. Context length is controlled to avoid hitting window limits, enabling handling of more complex problems.

1. Inference-time optimization is highly cost-effective, with lower costs than fine-tuning or pre-training. 2. Metacognitive ability (monitoring and adjusting one's own reasoning) is key for next-generation AI systems. 3. Failure recovery strategies need to be refined to identify the right time to persist, give up, or switch directions.

Limitations: Trigger condition thresholds are heuristically set and may vary by task; pruning strategies may be aggressive under complex reasoning chains; resampling depends on model generation diversity. Future directions: Adaptive threshold learning, fine-grained trajectory evaluation, and integration with reinforcement learning.

PruneTIR is an important advancement in the field of tool-integrated reasoning, proving that inference-time strategy design can significantly improve the efficiency and accuracy of LLM tool usage. As AI Agents become more prevalent, such optimization techniques will play a key role in enhancing user experience, reducing costs, and expanding application boundaries.