DeepSeek has introduced a new method to boost the reasoning capabilities of large language models (LLMs). Developed with Tsinghua University, the method combines generative reward modeling (GRM) and self-principled critique tuning.

The approach improves the accuracy and efficiency of LLMs in answering general queries.

According to a paper published on Apr. 4, on arXiv, DeepSeek’s GRM models showed competitive performance with existing public reward models and align AI outputs with human preferences.

DeepSeek plans to open source the GRM models, though no timeline has been shared.

The news comes amid speculation about the release of DeepSeek-R2, a possible successor to the R1 reasoning model.

Reuters reported it could launch this month, but DeepSeek has not confirmed this.

Recent DeepSeek developments