Characterize downstream scaling beyond the tested pretraining regime

Determine how downstream performance of wrist-IMU representations scales with participant count, the balance of recorded and simulated data, sensor placement, and larger pretraining corpora or longer training budgets.

Background

IMU-Bench varies pretraining data volume and motion coverage under a fixed architecture, data pool, and 50,000-step training budget. The authors explicitly state that these experiments do not establish how performance changes with participant count, the relative amount of recorded versus simulated data, or sensor placement.

The paper also reports no consistent downstream gain from enlarging the corpus at the tested budget, but does not determine whether larger corpora become beneficial when training for longer. This leaves the scaling behavior of the InrtlWrist representation and related wrist-IMU models unresolved outside the specific experimental regime evaluated.

References

Our scaling study does not isolate motion coverage from data volume and source composition. It also does not establish how downstream performance scales with participant count, the balance of recorded and simulated data, or sensor placement. The benefit of larger corpora at longer training budgets remains unresolved.

— IMU-Bench: Benchmarking Wrist IMU Representations Across Kinematic, Semantic and Clinical Tasks  (2610.19481 - Li et al., 14 Oct 2026) in Section Conclusion, paragraph “Limitations”; corroborated in Appendix, Section “Pretraining Volume and Motion Coverage”