Direct evaluation of TAIA under distributional shift

Evaluate the TAIA inference strategy directly under distributional shift in joint next-activity and remaining-time predictive process monitoring, rather than only contrasting its ablation across process-entropy settings.

Background

D-TAIA incorporates the TAIA inference strategy by removing feed-forward-network LoRA updates at inference while retaining attention updates, with the aim of improving robustness when test-time prefixes differ from the fine-tuning distribution. The ablation study compares the full framework with a version that omits TAIA, but evaluates this comparison across high- and low-entropy event logs rather than under an experimentally induced distributional shift.

Because the reported ablation does not directly test the condition that TAIA is designed to address, the paper leaves unresolved whether TAIA actually improves D-TAIA’s robustness to distributional shift in predictive process monitoring.