Archit Sharma

Direct Preference Optimization paper reframes RLHF as a classification loss

The method skipped the separate reward model and reinforcement-learning loop, and was later adopted for post-training open models including Zephyr and Tulu.

Ideas & essays