New Method Improves Dynamics Model Robustness by Re-Anchoring Event Credit.

Dynamics models, particularly those used for global physical questions like fault severity, often process a sequence of features and then pool them into a single readout vector. A critical, yet under-studied, problem arises here: without explicit step-level supervision, the model might accurately predict during training by relying on abundant, smooth background correlates rather than the brief, decisive physical events that truly determine the target outcome. This issue is termed "temporal credit dilution." This problem is insidious because it isn't exposed by standard training losses and isn't resolved by typical physics-informed residuals, as the error lies in *where* the global readout assigns functional credit. To address this, researchers introduce "Credit-in-Event," an interface-level probe to quantify how much pooled credit lands on event steps. They formally prove that a pooled linear reader will route credit to spurious background channels as the event fraction decreases. To combat this, they propose CREST (Credit Re-Anchoring for Event-core SpatioTemporal), a training-free and label-free readout method. CREST identifies a transient event core from learned features and then re-anchors the pooled representation by contrasting the event core with the rest of the sequence. Across various simulated systems (gear, impact) and real-world data (bearing vibration), CREST consistently reduces out-of-distribution error and restores proper event credit, demonstrating its effectiveness where other methods fail.