Ground-Truth Neighborhood Regularization for RL Post-Training of Time Series Foundation Models

A new paper on arXiv (2608.08010) introduces Ground-Truth Neighborhood Regularization (GTN-R) to address a phenomenon called 'suboptimal collapse' in reinforcement learning (RL) post-training of time series foundation models (TSFMs). The authors observe that during RL post-training, the output distributions of TSFMs can gradually drift away from the ground truth in certain forecast regions, limiting performance. They attribute this to difficulty in initially sampling high-quality trajectories near the ground truth. GTN-R is proposed as a regularization technique to keep the model outputs close to the ground truth, thereby improving forecasting accuracy. The paper is categorized as a cross-type announcement and is available on arXiv.