DecisionRank
A reranking method based on decision models. Through on-policy distillation in item space, it transfers the ranking knowledge of a generative LLM to a low-latency decision model, enabling efficient ranking in retrieval tasks.
Timeline (last 90 days)
DecisionRank transfers the ranking knowledge of generative LLMs to decision models through item-space on-policy distillation (OPD) and achieves leading results on multiple retrieval benchmarks.
The core design of item-space OPD is: rank the document scores output by the student model into a ranked sequence, then feed this list as input to the generative teacher model, which provides conditional feedback based on the list to directly supervise the student model's scores.
DecisionRank achieves state-of-the-art performance on general and reasoning-intensive retrieval benchmarks, including TREC DL, BEIR, and BRIGHT.
The 1.7B-parameter DecisionRank surpasses its 27B generative teacher on TREC DL19–23.
The 8B-scale DecisionRank outperforms its generative listwise counterpart while running 24.5 times faster on TREC DL and 16.7 times faster on BRIGHT.
All Facts (5)
DecisionRank transfers the ranking knowledge of generative LLMs to decision models through item-space on-policy distillation (OPD) and achieves leading results on multiple retrieval benchmarks.
50%UnverifiedThe core design of item-space OPD is: rank the document scores output by the student model into a ranked sequence, then feed this list as input to the generative teacher model, which provides conditional feedback based on the list to directly supervise the student model's scores.
50%UnverifiedDecisionRank achieves state-of-the-art performance on general and reasoning-intensive retrieval benchmarks, including TREC DL, BEIR, and BRIGHT.
50%UnverifiedThe 1.7B-parameter DecisionRank surpasses its 27B generative teacher on TREC DL19–23.
50%UnverifiedThe 8B-scale DecisionRank outperforms its generative listwise counterpart while running 24.5 times faster on TREC DL and 16.7 times faster on BRIGHT.
50%