Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2603.08241 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866918379944673280 |
|---|---|
| author | Loncour, Romain Bogaert, Jérémie Standaert, François-Xavier |
| author_facet | Loncour, Romain Bogaert, Jérémie Standaert, François-Xavier |
| contents | Transformer models are now a cornerstone in natural language processing. Yet, explaining their decisions remains a challenge. It was shown recently that the same model trained on the same data with a different randomness can lead to very different explanations. In this paper, we investigate how the (syntactic) context, the classes to be learned and the tasks influence this explanations' sensitivity to randomness. We show that they all have statistically significant impact: smallest for the (syntactic) context, medium for the classes and largest for the tasks. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2603_08241 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Sensivity of LLMs' Explanations to the Training Randomness:Context, Class & Task Dependencies Loncour, Romain Bogaert, Jérémie Standaert, François-Xavier Computation and Language Transformer models are now a cornerstone in natural language processing. Yet, explaining their decisions remains a challenge. It was shown recently that the same model trained on the same data with a different randomness can lead to very different explanations. In this paper, we investigate how the (syntactic) context, the classes to be learned and the tasks influence this explanations' sensitivity to randomness. We show that they all have statistically significant impact: smallest for the (syntactic) context, medium for the classes and largest for the tasks. |
| title | Sensivity of LLMs' Explanations to the Training Randomness:Context, Class & Task Dependencies |
| topic | Computation and Language |
| url | https://arxiv.org/abs/2603.08241 |