Test-Time Compute
A technical paradigm that dynamically increases computational resources during the model inference phase to improve output quality. It contrasts with training-time scaling, which increases parameters or data volume, and serves as a core characteristic of the new generation of reasoning models.
Timeline (last 90 days)
Anthropic positions dynamic workflows as the first step toward directing test-time compute at difficult problems and making scientific discovery broadly achievable by the public.
A Google DeepMind paper indicates that having a smaller model spend more time in deep thinking and generating hundreds of thousands of tokens can outperform a model 14 times its size on problems where the smaller model has some baseline probability of success.
All Facts (2)
Anthropic positions dynamic workflows as the first step toward directing test-time compute at difficult problems and making scientific discovery broadly achievable by the public.
50%UnverifiedA Google DeepMind paper indicates that having a smaller model spend more time in deep thinking and generating hundreds of thousands of tokens can outperform a model 14 times its size on problems where the smaller model has some baseline probability of success.
50%