Research · Oct 26, 2026, 00:00 UTC

Hugging Face paper proposes FSPO, a single-policy RL method for language model training

FSPO stores sparse historical next-token distributions as fixed temporal targets, avoiding a persistent teacher model. It reports about 4 points average gain over the strongest baseline under a 5-hour budget and peak memory dropping from about 70 GB to 56 GB on Olmo3-7B-Instruct.

• ModelOlmo3-7B-Instruct
• Topicresearch