Decoupled Predictive Decision Making: A Shared Conditional Predictor for Agents
DPDM improves agent decisions using deployment history without updating either model’s neural parameters.
A language model represents current and past records, while a shared, frozen LimiX-2 predictor estimates task-defined targets that a decision rule turns into choices. With Qwen3.5-27B as the encoder, the authors report mean accuracy of 84.92% across 50 benchmarks, versus 78.80% for few-shot prompting, using each method’s evaluated history budget. With identical histories and queries, DPDM reaches 82.33%. The same predictor serves five task categories, and accuracy improves as history grows.
HF Daily Papers' note
A language model represents current and past records, while a shared, frozen LimiX-2 predictor estimates task-defined targets that a decision rule turns into choices. With Qwen3.5-27B as the encoder, the authors report mean accuracy of 84.92% across 50 benchmarks, versus 78.80% for few-shot prompting, using each method’s evaluated history budget. With identical histories and queries, DPDM reaches 82.33%. The same predictor serves five task categories, and accuracy improves as history grows.
HF Daily Papers' note
score 4