token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
Why it matters
Test-time RL without answer voting or an external judge: the model evolves its own discriminating rubrics and scores rollouts from verdict-tokenThe chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.Full definition → likelihoods, reporting ~20-point gains on HealthBench and ResearchQA with OOD transfer.
articleInducing Reward-Free Judging Rubrics that Reduce Over-Crediting in Agent EvaluationDarragh Quinn, David Dylan, Roisin Healy, Fionn Carroll, Maeve Donnelly, Cormac Sheehan
articleDirect Preference Optimization: Your Language Model is Secretly a Reward ModelRafael Rafailov et al.