Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Vitenko, Ihor, Ibrahim, Noha, Amer-Yahia, Sihem
Format: Preprint
Veröffentlicht: 2026
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2604.20174
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866917442131853312
author Vitenko, Ihor
Ibrahim, Noha
Amer-Yahia, Sihem
author_facet Vitenko, Ihor
Ibrahim, Noha
Amer-Yahia, Sihem
contents Reinforcement learning (RL) policies are typically trained for fixed objectives, making reuse difficult when task requirements change. We study inference-time policy reuse: given a library of pre-trained policies and a new composite objective, can a high-quality policy be constructed entirely offline, without additional environment interaction? We introduce lever (Leveraging Efficient Vector Embeddings for Reusable policies), an end-to-end framework that retrieves relevant policies, evaluates them using behavioral embeddings, and composes new policies via offline Q-value composition. We focus on the support-limited regime, where no value propagation is possible, and show that the effectiveness of reuse depends critically on the coverage of available transitions. To balance performance and computational cost, lever proposes composition strategies that control the exploration of candidate policies. Experiments in deterministic GridWorld environments show that inference-time composition can match, and in some cases exceed, training-from-scratch performance while providing substantial speedups. At the same time, performance degrades when long-horizon dependencies require value propagation, highlighting a fundamental limitation of offline reuse.
format Preprint
id arxiv_https___arxiv_org_abs_2604_20174
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Lever: Inference-Time Policy Reuse under Support Constraints
Vitenko, Ihor
Ibrahim, Noha
Amer-Yahia, Sihem
Machine Learning
Reinforcement learning (RL) policies are typically trained for fixed objectives, making reuse difficult when task requirements change. We study inference-time policy reuse: given a library of pre-trained policies and a new composite objective, can a high-quality policy be constructed entirely offline, without additional environment interaction? We introduce lever (Leveraging Efficient Vector Embeddings for Reusable policies), an end-to-end framework that retrieves relevant policies, evaluates them using behavioral embeddings, and composes new policies via offline Q-value composition. We focus on the support-limited regime, where no value propagation is possible, and show that the effectiveness of reuse depends critically on the coverage of available transitions. To balance performance and computational cost, lever proposes composition strategies that control the exploration of candidate policies. Experiments in deterministic GridWorld environments show that inference-time composition can match, and in some cases exceed, training-from-scratch performance while providing substantial speedups. At the same time, performance degrades when long-horizon dependencies require value propagation, highlighting a fundamental limitation of offline reuse.
title Lever: Inference-Time Policy Reuse under Support Constraints
topic Machine Learning
url https://arxiv.org/abs/2604.20174