When Rubrics Fail: Error Enumeration as Reward in Reference-Free RL Post-Training for Virtual Try-On

Researchers propose Implicit Error Counting (IEC), a new reinforcement learning approach for training AI models in domains where multiple valid outputs exist and traditional rubric-based evaluation fails. The method focuses on counting what responses get wrong rather than what they get right, with validation shown in virtual try-on applications where it outperforms existing rubric-based methods.

• →Traditional rubric-based reward systems fail in domains with multiple valid outputs and no single ideal reference answer.
• →Implicit Error Counting (IEC) enumerates errors rather than checking correctness against rubrics, providing more reliable rewards for model training.
• →The method uses severity-weighted scores and group calibration to make error counting stable for optimization.
• →IEC outperformed Rubrics as Rewards (RaR) across all metrics in virtual try-on benchmarks.
• →The approach addresses a significant gap in current post-training methods for AI systems.