CoCA Framework Enables Co-Evolving Policy and Rubric Training
October 30, 2026
The CoCA framework uses capability awareness to jointly adapt RL rubrics and policies through criterion-level diagnosis. This prevents reward signal saturation by ensuring rubric difficulty tracks the evolving capability frontier of the model.
HOW THIS AFFECTS YOU
●
builderThis approach can help improve the alignment of models on complex, open-ended generative tasks.
●
researcherYou can implement closed-loop RL where the evaluation criteria evolve alongside the model's performance.