ICML 2026｜机制不是开关，而是一条轨迹：TRACE 如何捕捉系统的连续变化

TL;DR - TRACE models evolving causal mechanisms as continuous mixtures of atomic mechanisms rather than discrete switches, enabling recovery of how a system changes over time. The ICML 2026 paper provides identifiability guarantees and strong results on synthetic, vehicle-turning, and gait-transition data.

• Uses a mixture-of-experts architecture to learn atomic mechanisms from labeled pure-domain data, then estimates continuous mixture trajectories via least-squares projection.
• Establishes latent-variable identifiability up to permutation and component-wise transformations, plus finite-sample error bounds for trajectory recovery.
• Achieves trajectory correlations of 0.94 ± 0.05 on synthetic data, 0.960 on UAVDT vehicle turns, and 0.856 ± 0.043 on walking-to-running sequences.
• Assumes known mechanism count and labeled pure-mechanism training data; recovery becomes difficult when mechanism bases are nearly collinear.