A new preprint on arXiv (2610.12010) examines predictive representation learning from photoplethysmography (PPG) signals. The authors point out that even when a model uses causal attention, it can still violate causal information access. The culprit is not the attention mechanism itself but other parts of the pipeline: normalization steps, nonlocal transforms, or companion views that may depend on samples the model is supposed to treat as withheld.
The paper argues that this leakage is a subtle but serious problem for future-beat state forecasting. If a model implicitly sees future data through these indirect dependencies, its performance on real-world tasks may be overstated. The authors propose PulseBound, a method designed to enforce an explicit information boundary so that the model cannot access future samples through any route.
Because the abstract is truncated, the preprint's full methodology and experimental results are not yet visible in the source. What is clear is the central claim: causal attention is necessary but not sufficient for causal inference in PPG modeling, and explicit safeguards are needed to ensure that predictive models are truly learning from the past only.