Traditionally, LLMs make decisions by autoregressive decoding, which is slow and hides probabilities.
LLMs can be repurposed as fast single-pass decision models with explicit probabilities
Diverse supervision and reasoning distillation improve new-schema accuracy while cutting latency by 182-fold on Gemma-4-31B.
Top University
Ziyi Chen · Yan Zhang · Jianhui Wei · Jianing Yin · Daoan Zhang
University of Illinois Urbana-Champaign · National University of Singapore · Zhejiang University · University of Pennsylvania · University of Rochester
Research Digest··3 min read
The authors propose Large Decision Models (LDMs), which repurpose large language models to return explicit probabilities over labels in a single prefill forward pass.
§