Recursa paper reports legal task accuracy gain from self-built knowledge repository

Recursa researchers propose Recursa, a recursively self-improving foundation model system, and report that it raises DeepSeek-V4.1-Flash from 30.0% to 55.0% all-pass accuracy on held-out legal tasks.

The gain is reported without a stronger teacher model or human experts, according to the paper linked in a post by Elvis Saravia (@omarsar0).

The figures come from the authors' paper as cited by the post. The post does not describe how the knowledge repository is built or tested beyond the legal task result.