SimpleTIR

A plug-and-play algorithm that stabilizes multi-turn Tool-Integrated Reasoning training via RL (Zhenghai Xue, Longtao Zheng, Qian Liu et al., 2025). It identifies and removes from policy updates problematic trajectories with “void turns” — turns that produce neither a code block nor an answer — which, after tool feedback, trigger distribution drift and gradient explosions. It raised AIME24 from 22.1 to 50.5 on Qwen2.5-7B.