Search papers, labs, and topics across Lattice.
This paper addresses the issue of temporal credit dilution in learned dynamics models, where models inaccurately attribute credit to smooth correlates instead of critical physical events due to trajectory-level supervision. The authors introduce Credit-in-Event to quantify event credit allocation and propose CREST, a novel label-free readout method that effectively re-anchors pooled representations by contrasting event features with background noise. Experiments demonstrate that CREST significantly reduces out-of-distribution errors across various systems while restoring accurate event credit assignment, outperforming traditional methods that rely on stability or locality.
Temporal credit dilution can lead models to misattribute importance, but CREST effectively re-anchors event credit, enhancing robustness in global readouts.
Learned dynamics models often answer global physical questions, such as fault severity or impact stiffness, by pooling a per-step feature sequence into one readout vector. This sequence-to-global interface creates an under-studied temporal credit problem: with only trajectory-level supervision, a model can predict accurately in training conditions while reading from abundant smooth correlates rather than the brief physical events that determine the target. We call this failure temporal credit dilution. It is not exposed by the training loss and is not removed by standard physics-informed residuals, because the error lies in where the global readout assigns functional credit. We introduce Credit-in-Event, an interface-level probe for measuring how much pooled credit lands on event steps, and prove in closed form that a pooled linear reader routes credit to a spurious background channel as the event fraction shrinks. We then propose CREST, a training-free and label-free readout that estimates a transient event core from learned features and re-anchors the pooled representation through event-versus-rest contrast. Across simulated gear and impact systems, recurrent and attention encoders, and public bearing vibration data, CREST reduces out-of-distribution error while restoring event credit. Ablations show that stable-step selection and receptive-field shrinking fail, confirming that the gain comes from event-core credit re-anchoring rather than a generic locality or stability prior.
