cad-gpt-neo-20b-nqswap-gain

Summary

This is a measured result showing that the CAD decoding strategy dramatically boosts how well the 20-billion-parameter GPT-Neo model handles adversarial or trick questions, nearly doubling its score on one benchmark and adding over half again on another compared to its usual way of generating text. It matters because it isolates the decoding method as a major lever for robustness, meaning you can get large gains without retraining the model at all.