A new arXiv paper introduces FinPILOT, a method for reinforcement-learning (RL) trading agents that enables policy adjustments at inference time. The authors argue that typical RL agents for portfolio management are trained and deployed as static policies, with no way to incorporate fresh price forecasts once trading begins. FinPILOT directly addresses this gap by planning before each trade, using current forecasts to guide the agent's actions.

The paper's abstract is brief, but the core idea is clear: instead of treating the trained policy as fixed, FinPILOT optimizes the policy at inference time based on new information. This could make RL trading agents more responsive to changing market conditions, potentially improving their performance in dynamic environments.

Because only the abstract is available, details on the algorithm's architecture, experimental results, and comparisons to baselines are not yet described. The work is posted on arXiv under the title "Plan Before You Trade: Inference-Time Optimization for RL Trading Agents" (arXiv:2605.12653).