DaCe-DT: Data-Centric Offline Multi-Task Reinforcement Learning via Adaptive Prompts and Trajectory Correction for Heterogeneous Tasks
DaCe-DT addresses limitations in Offline Multi-Task Reinforcement Learning by tackling three data bottlenecks: fixed prompt lengths, semantically irrelevant prompts, and fragmented trajectories. The framework introduces Length-Gated Prompt Masking, Retrieval-Augmented Prompt Construction, and Value-Adaptive Return Calibration to adaptively select context and retrieve relevant behavioral examples for better performance.