DecoEvo: Co-Evolving LLM Solver and Rubric Generator in Text Space

A new method called DecoEvo (Decoupled Co-Evolution) has been developed by researchers to enhance large language models (LLMs) in text space without altering the model weights. Unlike traditional text-space methods that maintain a static evaluation, DecoEvo simultaneously evolves a solver skill and a rubric-generator skill with separate objectives. The solver benefits from feedback at the criterion level, while the rubric-generator undergoes revisions through complementary processes. This innovation tackles the limitations of fixed rubrics that fail to capture unmeasured aspects and prevents the solver from exploiting unreliable rubric changes. Notably, gold rubrics are not utilized during the optimization process. The research paper can be found on arXiv.