Datchworth

New DATC method improves language model learning by assigning token-specific credit

Researchers have developed a new method called Direction-Aware Token Credit Assignment (DATC) to improve how language models learn from feedback. Current methods like GRPO assign a single reward advantage across all tok…