Reproduce “Learning to Reason from Reference Samples Alone: Reinforced Adversarial Learning for Modern LLM Post-Training” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Albert Bou, Javier Valdes-Aleman, Jon M. Laurent, Samuel G. Rodriques, Andrew D. White, Michaela Hinks, Siddharth Narayanan
The authors introduce RAL, co-training a generator and judge on reference and generated answers; the judge’s reference probability is the generator’s reinforcement-learning reward. RAL reached 0.949 held-out exact accuracy on multiplication, versus 0.938 with verifier-reward training. On title-conditioned NIH abstracts, fooling rates were 35% for LLM judges and 55% for human evaluators, indicating difficulty distinguishing them under this protocol, not factual validity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.