Hasty Briefsbeta

Bilingual

What Is RLCD? The Secret Behind Jev

5 hours ago
  • Jev is a preference model generalized to a software interface, not a language model.
  • RLCD extends PPRM (Pairwise Preference Reward Model) to multiway decisions using Plackett–Luce.
  • Calibration via Brier score and temperature scaling makes probabilities empirically reliable.
  • Jev provides typed primitives (Noul, Choice, Score) for binary, multiway, and ordinal decisions.
  • Parallel inference is achieved by sequence packing and tree attention, avoiding autoregressive generation.
  • The decision head computes utilities from contextual representations, enabling multi-candidate scoring.
  • RLCD focuses on calibrated distributions over actions, distinct from RLHF/RLVR which emphasize reward sources.
  • The system decomposes into RLCD for learning and packing/masking for efficient serving.