Noiser: Bounded Input Perturbations for Attributing Large Language Models
Fuente:
arXiv
Salvato in:
| Autori principali: | , , , , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2025
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866909596024569856 |
|---|---|
| author | Madani, Mohammad Reza Ghasemi Gema, Aryo Pradipta Sarti, Gabriele Zhao, Yu Minervini, Pasquale Passerini, Andrea |
| author_facet | Madani, Mohammad Reza Ghasemi Gema, Aryo Pradipta Sarti, Gabriele Zhao, Yu Minervini, Pasquale Passerini, Andrea |
| contents | Feature attribution (FA) methods are common post-hoc approaches that explain how Large Language Models (LLMs) make predictions. Accordingly, generating faithful attributions that reflect the actual inner behavior of the model is crucial. In this paper, we introduce Noiser, a perturbation-based FA method that imposes bounded noise on each input embedding and measures the robustness of the model against partially noised input to obtain the input attributions. Additionally, we propose an answerability metric that employs an instructed judge model to assess the extent to which highly scored tokens suffice to recover the predicted output. Through a comprehensive evaluation across six LLMs and three tasks, we demonstrate that Noiser consistently outperforms existing gradient-based, attention-based, and perturbation-based FA methods in terms of both faithfulness and answerability, making it a robust and effective approach for explaining language model predictions. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2504_02911 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Noiser: Bounded Input Perturbations for Attributing Large Language Models Madani, Mohammad Reza Ghasemi Gema, Aryo Pradipta Sarti, Gabriele Zhao, Yu Minervini, Pasquale Passerini, Andrea Computation and Language Artificial Intelligence Feature attribution (FA) methods are common post-hoc approaches that explain how Large Language Models (LLMs) make predictions. Accordingly, generating faithful attributions that reflect the actual inner behavior of the model is crucial. In this paper, we introduce Noiser, a perturbation-based FA method that imposes bounded noise on each input embedding and measures the robustness of the model against partially noised input to obtain the input attributions. Additionally, we propose an answerability metric that employs an instructed judge model to assess the extent to which highly scored tokens suffice to recover the predicted output. Through a comprehensive evaluation across six LLMs and three tasks, we demonstrate that Noiser consistently outperforms existing gradient-based, attention-based, and perturbation-based FA methods in terms of both faithfulness and answerability, making it a robust and effective approach for explaining language model predictions. |
| title | Noiser: Bounded Input Perturbations for Attributing Large Language Models |
| topic | Computation and Language Artificial Intelligence |
| url | https://arxiv.org/abs/2504.02911 |