Muath Alyobi

تخصص الذكاء الاصطناعي · دفعة 2026

ReadEnough: Training-Free Early Stopping for Long-Context Reading

ReadEnough is a training-free early-stopping layer for sequential long-context reading. The input is divided into fixed-order chunks, and a frozen language model maintains compact running notes that are updated after each chunk. For multiple-choice tasks, a short probe extracts the model’s exact probability distribution over the answer options. ReadEnough accumulates evidence of distributional stability while requiring sufficient answer confidence, stopping when both conditions are satisfied. It then returns the most probable answer and the confidence measured at the stopping point. For open-ended tasks, the probe instead produces a short draft answer, and changes between successive drafts provide the stability signal. ReadEnough requires no labeled evidence spans, activation-based classifiers, task-specific training, or reinforcement-learned exit policy. The stopping confidence is available as a byproduct and may be used by downstream selective-prediction or abstention mechanisms.

اقرأ التقرير النهائي (PDF)