← Back to digest
HF Trending Papers

DecoEvo: co-evolving solver and rubric-generator

What's this about?

DecoEvo co-evolves two skills — a problem-solver and the rubric-generator that grades it — while decoupling their scores so one can't game the other into collapse.

Why it matters

Self-grading systems tend to collapse when solver and judge co-adapt; decoupling the scores (57 upvotes) is a clean idea for keeping automated curricula honest.

← CoRT: counterfactual replay for rubric-guided policy optimizationCLBench-V: multimodal in-context learning benchmark →