DeepSeek AI has published a research paper outlining SPCT, a new technique designed to improve how general reward models scale during inference. The work targets a practical bottleneck: applying reward models effectively when models are already deployed, not just when they are being trained.
The paper positions SPCT as a novel approach to inference-time scaling for GRMs. DeepSeek has not released an R2 model, but coverage of the paper interprets it as a signal that such a model is in development. The details in the paper are technical, and the broader product implications remain speculative.