Instead of evaluating each token, one at a time, RLFT grades entire answers—perfect for scenarios like OpenAI’s o1, where we can’t access intermediate tokens before the “end_of_thought.”

→ View original post on X — @whats_ai