Self-Rewarding LLM
LLM acts as both generator AND reward model on its own outputs — used for iterative DPO without human raters. Meta paper January 2024.
Three iterations on Llama 2 70B beat GPT-4 0613 on AlpacaEval. Inspired by debate + Constitutional AI. Risk: bias accumulation, since the judge has the same blind spots as the generator. Combined with synthetic data + SPIN ⇒ "model improves itself without human feedback" pipelines.