Tracing Logit Trajectories Across Layer Depth: Dataset-Level Explainability for Language Models

摘要

Sentence-level explanations can miss the bigger picture of how a black-box model behaves across data, which matters most for complex criteria like safety that cannot be defined by a single rule. We trace **Logit-Trajectory**, which tracks adjacent-layer logit updates as vectors and aggregates them into a reproducible dataset-level trajectory pattern, enabling depth-wise explainability through signals such as coherence and angular rotation. Across 6 languages and 5 NLP tasks, we show these trajectory summaries reveal consistent depth-wise patterns that divergence- and similarity-based baselines often wash out due to scalarization. As a case study where dataset-level intermediate decision structure matters, we evaluate safety classification, reporting both trajectory-level visual separability and classification performance.