HART: Hybrid Attention-RWKV Time series Foundation Model
HART is a hybrid time-series foundation model combining Attention and RWKV blocks to improve long-horizon forecasting. It addresses RWKV's degradation during non-autoregressive multi-step tasks by alternating blocks, allowing attention to retrieve history and RWKV to model temporal dynamics. Experiments show HART outperforms pure RWKV and Transformer on medium/long horizons while maintaining good short-term performance.