Learning Stratified Policies for Inference-Time Objectives
Splitting a fixed response budget across learned policies improved pass@k over sampling repeatedly from one policy.
Each policy uses a separate LoRA adapter on a shared language model and contributes one response to the set of k samples. The authors train the policies for pass@k and report consistent gains across multiple reasoning tasks. Their linear-setting analysis also identifies cases where a population of LoRAs achieves zero best-of-k approximation error while every single LoRA retains positive prediction error. The supplied abstract gives no numerical gains.
ArXiv · AI/CL/LG's note
Each policy uses a separate LoRA adapter on a shared language model and contributes one response to the set of k samples. The authors train the policies for pass@k and report consistent gains across multiple reasoning tasks. Their linear-setting analysis also identifies cases where a population of LoRAs achieves zero best-of-k approximation error while every single LoRA retains positive prediction error. The supplied abstract gives no numerical gains.
ArXiv · AI/CL/LG's note
score 5