s1

Hi everyone, Sharing s1, a new approach to improving LLM performance at test time. This work comes from researchers at Stanford University and the University of Washington, and offers some exciting results: · Strong Reasoning: Matches the performance of larger models (like o1-preview) on reasoning tasks. · Minimal Data: Achieves this with only 1,000 examples (insane, right?). · Budget Forcing: Uses a novel "budget forcing" technique during inference. · Test-Time Scaling: Improves performance without retraining the model. · Fully Open-Source: Model, data, and code are all available.