GIFT: LLM-Guided State-Reward Interface for Financial Reinforcement Learning

A new framework called GIFT uses large language models to design better state and reward interfaces for reinforcement learning agents that trade financial portfolios, according to a paper submitted to arXiv in 2026 [1][2]. Financial portfolio trading can be formulated as a reinforcement learning problem where an agent sequentially rebalances assets to balance return, risk, and transaction costs [2]. In non-stationary markets, raw OHLCV states and short-horizon return rewards often provide an under-specified learning interface [2]. The GIFT framework addresses this by using large language models to inject financial knowledge into state and reward design while constraining open-ended generation, rather than having the LLM make trading decisions directly [1][2]. GIFT operates through three components. Factor-guided State Enhancement generates state features from financial-factor primitives. Risk-rule-guided Reward Shaping generates auxiliary rewards from portfolio-risk rules. Diagnostic-guided Refinement revises candidate interfaces using PPO rollout diagnostics [1][2]. After refinement, GIFT fixes the selected state-reward interface before evaluation, with no further LLM queries or interface updates at test time [2]. Comprehensive rolling-window experiments across diverse market regimes and portfolio scenarios demonstrate that GIFT improves learning-signal quality and out-of-sample risk-adjusted portfolio performance over baselines [1][2]. Code and data for the framework have been made publicly available on GitHub [2]. The paper was submitted to arXiv on 7 June 2026 under the Computer Science and Artificial Intelligence category [1]. The work builds on a growing body of research exploring the intersection of large language models and financial reinforcement learning, as evidenced by related preprints catalogued on the same preprint server [3][4][5].