Research paper
GIFT: LLM-Guided State-Reward Interface for Financial Reinforcement Learning
arXiv preprint
Abstract
An LLM-guided framework for designing state and reward interfaces for PPO-based financial reinforcement learning, using factor-guided state enhancement, risk-rule-guided reward shaping, and diagnostic-guided refinement.
Overview
GIFT uses large language models as constrained interface designers for financial reinforcement learning rather than direct trading agents. It augments raw market observations with factor-derived state features, shapes rewards with portfolio-risk rules, and refines candidate interfaces using PPO rollout diagnostics before fixing the selected interface for test-time evaluation.
Citation
@article{2026-gift,
title = {GIFT: LLM-Guided State-Reward Interface for Financial Reinforcement Learning},
author = {Yanyan Wu and Boyi Zhang and Yanlin Liu and Xinyu Fang and Jining Luan and Meiqi Zhang and Jiacheng Liu and Hao Zeng and Dexu Yu and Chang Liu and Hanwen Du and Yongxin Ni and Youhua Li},
year = {2026},
journal = {arXiv preprint},
eprint = {2606.08450},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2606.08450},
}