Saved in:
Bibliographic Details
Main Authors: Xu, Qianyi, Habib, Gousia, Wu, Feng, Du, Yanrui, Chen, Zhihui, Mishra, Swapnil, Perera, Dilruk, Feng, Mengling
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2602.03305
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866917247485739008
author Xu, Qianyi
Habib, Gousia
Wu, Feng
Du, Yanrui
Chen, Zhihui
Mishra, Swapnil
Perera, Dilruk
Feng, Mengling
author_facet Xu, Qianyi
Habib, Gousia
Wu, Feng
Du, Yanrui
Chen, Zhihui
Mishra, Swapnil
Perera, Dilruk
Feng, Mengling
contents Reinforcement Learning (RL) offers a powerful framework for optimizing dynamic treatment regimes (DTRs). However, clinical RL is fundamentally bottlenecked by reward engineering: the challenge of defining signals that safely and effectively guide policy learning in complex, sparse offline environments. Existing approaches often rely on manual heuristics that fail to generalize across diverse pathologies. To address this, we propose an automated pipeline leveraging Large Language Models (LLMs) for offline reward design and verification. We formulate the reward function using potential functions consisted of three core components: survival, confidence, and competence. We further introduce quantitative metrics to rigorously evaluate and select the optimal reward structure prior to deployment. By integrating LLM-driven domain knowledge, our framework automates the design of reward functions for specific diseases while significantly enhancing the performance of the resulting policies.
format Preprint
id arxiv_https___arxiv_org_abs_2602_03305
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle medR: Reward Engineering for Clinical Offline Reinforcement Learning via Tri-Drive Potential Functions
Xu, Qianyi
Habib, Gousia
Wu, Feng
Du, Yanrui
Chen, Zhihui
Mishra, Swapnil
Perera, Dilruk
Feng, Mengling
Machine Learning
Reinforcement Learning (RL) offers a powerful framework for optimizing dynamic treatment regimes (DTRs). However, clinical RL is fundamentally bottlenecked by reward engineering: the challenge of defining signals that safely and effectively guide policy learning in complex, sparse offline environments. Existing approaches often rely on manual heuristics that fail to generalize across diverse pathologies. To address this, we propose an automated pipeline leveraging Large Language Models (LLMs) for offline reward design and verification. We formulate the reward function using potential functions consisted of three core components: survival, confidence, and competence. We further introduce quantitative metrics to rigorously evaluate and select the optimal reward structure prior to deployment. By integrating LLM-driven domain knowledge, our framework automates the design of reward functions for specific diseases while significantly enhancing the performance of the resulting policies.
title medR: Reward Engineering for Clinical Offline Reinforcement Learning via Tri-Drive Potential Functions
topic Machine Learning
url https://arxiv.org/abs/2602.03305