Reasoning Stabilization Point: A Training-Time Signal for Stable Evidence and Shortcut Reliance
Fuente:
arXiv
Gespeichert in:
| 1. Verfasser: | |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2026
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866914260321304576 |
|---|---|
| author | Dhayalkar, Sahil Rajesh |
| author_facet | Dhayalkar, Sahil Rajesh |
| contents | Fine-tuning pretrained language models can improve task performance while subtly altering the evidence a model relies on. We propose a training-time interpretability view that tracks token-level attributions across finetuning epochs. We define explanation driftas the epoch-to-epoch change in normalized token attributions on a fixed probe set, and introduce the Reasoning Stabilization Point(RSP), the earliest epoch after which drift remains consistently low. RSP is computed from within-run drift dynamics and requires no tuning on out-of-distribution data. Across multiple lightweight transformer classifiers and benchmark classification tasks, drift typically collapses into a low, stable regime early in training, while validation accuracy continues to change only marginally. In a controlled shortcut setting with label-correlated trigger tokens, attribution dynamics expose increasing reliance on the shortcut even when validation accuracy remains competitive. Overall, explanation drift provides a simple, low-cost diagnostic for monitoring how decision evidence evolves during fine-tuning and for selecting checkpoints in a stable-evidence regime. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2601_11625 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Reasoning Stabilization Point: A Training-Time Signal for Stable Evidence and Shortcut Reliance Dhayalkar, Sahil Rajesh Artificial Intelligence Machine Learning Fine-tuning pretrained language models can improve task performance while subtly altering the evidence a model relies on. We propose a training-time interpretability view that tracks token-level attributions across finetuning epochs. We define explanation driftas the epoch-to-epoch change in normalized token attributions on a fixed probe set, and introduce the Reasoning Stabilization Point(RSP), the earliest epoch after which drift remains consistently low. RSP is computed from within-run drift dynamics and requires no tuning on out-of-distribution data. Across multiple lightweight transformer classifiers and benchmark classification tasks, drift typically collapses into a low, stable regime early in training, while validation accuracy continues to change only marginally. In a controlled shortcut setting with label-correlated trigger tokens, attribution dynamics expose increasing reliance on the shortcut even when validation accuracy remains competitive. Overall, explanation drift provides a simple, low-cost diagnostic for monitoring how decision evidence evolves during fine-tuning and for selecting checkpoints in a stable-evidence regime. |
| title | Reasoning Stabilization Point: A Training-Time Signal for Stable Evidence and Shortcut Reliance |
| topic | Artificial Intelligence Machine Learning |
| url | https://arxiv.org/abs/2601.11625 |