ReHoPER Improves LLM Reasoning via Receding-Horizon Planning
October 2, 2026
ReHoPER is a zero-shot, inference-only method that improves compositional reasoning by iteratively generating and answering intermediate questions. The approach plans a horizon of candidate questions and replans based on updated history without requiring task-specific prompt design.
HOW THIS AFFECTS YOU
●
builderYou can improve your application's reasoning capabilities by implementing a receding-horizon planning loop.
●
researcherThis offers a task-agnostic way to boost reasoning performance via iterative planning at inference time.