DecisionRank Uses On-Policy Distillation for Efficient Decision Models
October 29, 2026
DecisionRank bridges the gap between generative LLM ranking and score-based decision models. It uses item-space on-policy distillation to train student models using conditional feedback from a generative teacher.
HOW THIS AFFECTS YOU
●
builderThis technique can reduce autoregressive decoding latency by using efficient decision models for ranking.
●
researcherThis provides a method to transfer complex ranking knowledge to simpler, faster architectures.