Alibaba's Qwen team makes AI models think deeper with new algorithm

Alibaba's Qwen team introduced FIPO, an RL algorithm that weights tokens by their downstream influence.

FIPO measures cumulative probability shifts to assign token-level rewards without a separate value model.

On Qwen2.5-32B-Base trained with DAPO data, FIPO extended chain-of-thought length from ~4,000 to over 10,000 tokens.

FIPO raised AIME-2024 accuracy from 50% to 56% (peaking at 58%) and improved AIME-2025 scores from 38% to 43%.

Training stability required discounting and filtering to prevent instabilities and crashes.