Technical Summary: PRUNETIR

Problem Statement

Tool-Integrated Reasoning (TIR) enables Large Language Models (LLMs) to interact with external tools, such as code interpreters, to overcome limitations in computational accuracy and knowledge cutoffs. While recent research has focused on equipping LLMs with tool-use capabilities via prompting, supervised fine-tuning (SFT), or reinforcement learning (RL), the optimization of reasoning at inference time for already tool-capable models remains underexplored.

The authors identify two critical phenomena that degrade TIR performance during inference:

1. Negative Correlation with Errors: The number and proportion of erroneous tool calls are negatively correlated with the correctness of the final answer. Samples with incorrect answers exhibit significantly higher mean and median counts of erroneous calls compared to correct ones.
2. The "Stuck" Phenomenon: When an LLM fails to resolve an erroneous tool call within a few subsequent turns, it tends to become "stuck," continuing to iterate without progress even when granted many additional turns. This leads to wasted computational resources and failure to generate an answer within the turn budget.

These errors accumulate in the context window, causing the model to degrade its instruction-following and reasoning abilities, often collapsing into incorrect conclusions without further reflection.

Methodology: PRUNETIR

To address these issues, the authors propose PRUNETIR, a training-free, inference-time framework designed to prune erroneous trajectories and prevent models from getting stuck. It operates through three core components:

1. Success-Triggered Pruning (STP)

When an LLM successfully resolves an error (i.e., generates a correct tool call that executes without error), STP prunes the entire error-resolution trace.

• Mechanism: It removes all intermediate erroneous tool calls and their corresponding feedback, retaining only the final successful tool call and its feedback.
• Intent Shift Detection: To ensure reasoning coherence, the framework employs Algorithm 1 to detect "coding-intent shifts" during the error resolution process. If the model switches strategies to solve the original problem rather than fixing the specific error, the reasoning content from the new strategy is concatenated to the retained history before pruning.
• Goal: This mitigates the accumulation of error-induced noise that degrades reasoning capabilities while preserving the guidance value of the resolution process.

2. Stuck-Triggered Pruning and Resampling (STPR)

If an erroneous tool call fails to reach a successful resolution within a predefined Turn Limit, the model is considered "stuck."

• Mechanism: STPR prunes the entire error-resolution trace and resamples a new tool call conditioned on the interaction history preceding the initial erroneous call.
• Goal: This forces the model to explore alternative paths rather than continuing to exploit failing resolution trajectories, thereby reducing the risk of the model remaining stuck.

3. Retry–Triggered Tool Suspension (RTTS)

If the STPR component is triggered consecutively a predefined number of times (Retry Limit), the model is instructed to temporarily suspend tool usage.

• Mechanism: The model reverts to manual reasoning (using a specific prompt) instead of attempting further tool calls.
• Goal: This serves as a conservative fallback for sustained tool-use failures, ensuring the reasoning process continues without getting trapped in an infinite loop of failed tool interactions.

Key Contributions

1. Empirical Identification of TIR Failure Modes: The paper establishes that erroneous tool calls are negatively correlated with answer correctness and that models frequently get stuck when unable to resolve errors quickly.
2. PRUNETIR Framework: The proposal of a novel, training-free framework that enhances TIR at inference time through pruning, resampling, and suspension mechanisms.
3. Comprehensive Evaluation: Extensive experiments demonstrating that PRUNETIR improves Pass@1 and tool-use efficiency while reducing working context length across multiple LLMs and benchmarks.

Experimental Results

The authors evaluated PRUNETIR on three mathematical benchmarks: AIME24, AIME25, and BeyondAIME, using Qwen3-8B, Qwen3-14B, and ReTool-32B.

• Performance Gains: Applied to Qwen3-8B on AIME24, PRUNETIR increased Pass@1 from 62.1% to 72.7% (+10.6 percentage points).
• Efficiency: The average number of tool calls (TCN) for Qwen3-8B on AIME24 dropped from 7.7 to 4.2, representing a 45.5% improvement in tool-use efficiency.
• Context Reduction: The working context token count (WTN) decreased by 17.4% (from 11.5K to 9.5K tokens), alleviating long-context burdens.
• Generalization: Consistent improvements were observed across Qwen3-14B and ReTool-32B. The method also showed effectiveness on the GPQA-diamond dataset (biology, physics, chemistry), indicating generalization beyond mathematics.
• Ablation: Removing any of the three components (STP, STPR, RTTS) resulted in lower performance, confirming the necessity of the full framework. Specifically, STP alone improved Pass@1, while STPR and RTTS further enhanced efficiency and stability.

Significance and Claims

The paper claims that PRUNETIR offers a simple yet effective approach to enhancing tool-integrated reasoning without requiring additional training or model parameter modification. By actively managing the inference trajectory, it prevents the degradation of reasoning capabilities caused by error accumulation and avoids the inefficiency of "stuck" models.

The authors position PRUNETIR as a lightweight, plug-and-play solution that can be integrated into any tool-capable LLM. They further suggest that the high-quality, pruned trajectories generated by PRUNETIR could serve as superior data for future self-training or distillation efforts, potentially enabling the self-evolution of tool-use capabilities. The work emphasizes that inference-time optimization is a critical, yet underutilized, avenue for improving the problem-solving performance of modern LLMs.