NeuralStackStanford University researchers released Recursive Preference Optimization on July 20 2026 demonstrating a 40 percent gain in multi-step reasoning accuracy across model sizes from 7B to 70B parameters. The method reduces required training compute by 35 percent compared with standard reinforcement learning from human feedback pipelines.
Experiments used the GSM8K and MATH benchmarks with the 70B model reaching 94.2 percent accuracy. The paper will be presented at NeurIPS 2026 in December.
The research was funded by a 25 million dollar grant from the National Science Foundation awarded in 2024. Stanford plans to release training code and model checkpoints on July 28.
Prior work at Stanford focused on efficient fine-tuning techniques that influenced several commercial alignment pipelines in 2025.
Why this matters
The efficiency gains could lower barriers for academic and startup labs to train competitive reasoning models. Reduced compute requirements also lessen environmental impact of large-scale training runs.
Major labs are already testing the method internally and may incorporate it into next-generation releases by early 2027.
Stanford will host a workshop on Recursive Preference Optimization in September 2026.