Compute as Teacher (CaT) was jointly proposed by Oxford University, Meta Super Intelligence Laboratory and other institutions in September 2025. The relevant research results were published in the paper "Compute as Teacher: Turning Inference Compute Into Reference-Free Supervision".

Where does the learning signal come from when there are no true labels after training? CaT transforms the model's self-exploration into no-reference supervision by synthesizing a single reference from a set of parallel runs. The model's exploration at inference time is then transformed into no-reference supervision and optimized towards this reference. As a test-time procedure, CaT improves the performance of Gemma 3 4B, Qwen 3 4B, and Llama 3.1 8B models (up to +27% on MATH-500 and +12% on HealthBench).