Mechanism of long-horizon degradation in xLSTM-based TiRex

Determine whether the xLSTM-based TiRex time-series foundation model shares the write-compensation mechanism associated with Placeholder Idling Decay in RWKV-7 during non-autoregressive multi-step forecasting, using per-architecture analysis.

Background

The paper’s mechanistic analysis focuses on RWKV-7 and explains its long-horizon degradation under non-autoregressive forecasting through a hypothesized imbalance between recurrent state transitions and placeholder-driven writes. The xLSTM-based TiRex model exhibits qualitatively similar amplitude degradation, but the reported experiments do not establish that the same state-update mechanism causes both behaviors.

The authors therefore identify a model-specific unresolved question: verifying whether TiRex’s degradation results from the same write-compensation mechanism requires a separate analysis of its architecture and state dynamics.

References

Figure 1 shows that the xLSTM-based TiRex also exhibits similar long-horizon amplitude degradation, but whether it shares the same write-compensation mechanism requires per-architecture analysis.

— HART: Hybrid Attention-RWKV Time series Foundation Model  (2610.12883 - Lin et al., 9 Oct 2026) in Section 6, Limitations and Future Work; related qualification in Section 3.1