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

Luan Zhang, Dandan Song, Zhijing Wu, Zhengyu Chen, Chen Zhang, Yuhang Tian, Huipeng Ma, Chenhao Li, Changzhi Zhou, Xudong Li, Shuhao Zhang · 2026

Read PruneTIR : Inference-Time Tool Call Pruning for Effective yet Efficient Tool-Integrated Reasoning by Luan Zhang, Dandan Song, Zhijing Wu, Zhengyu Chen, Chen Zhang, Yuhang Tian, Huipeng Ma, Chenhao Li, Changzhi Zhou, Xudong Li, Shuhao Zhang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-09931.txt.

Description

Zhang and ten colleagues improve tool-using language models at inference time, without any extra training. Tool-integrated reasoning lets a model call code interpreters and other tools mid-thought, yet failed tool calls drag answers down. The authors observe that errors usually clear within a few turns or not at all, with long retry spirals rarely helping. Their framework, PruneTIR, acts on that pattern.

Three components share the work. Success-Triggered Pruning cuts a trajectory once it succeeds, Stuck-Triggered Pruning and Resampling restarts stalled tool calls, and Retry-Triggered Tool Suspension stops tool use after repeated failure. Together they limit the damage from bad calls and keep models from circling failed fixes. Experiments show clear gains in Pass@1, the share of first answers that are right, plus better efficiency.

Catalog record

CREATORLuan Zhang, Dandan Song, Zhijing Wu, Zhengyu Chen, Chen Zhang, Yuhang Tian, Huipeng Ma, Chenhao Li, Changzhi Zhou, Xudong Li, Shuhao Zhang
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE52 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.09931

Browse the full library · About the Tower of Babel