SafeConstellations: Mitigating Over-Refusals in LLMs Through Task-Aware Representation Steering
Fuente:
arXiv
Enregistré dans:
| Auteurs principaux: | , , , |
|---|---|
| Format: | Preprint |
| Publié: |
2025
|
| Sujets: | |
| Accès en ligne: | |
| Tags: |
Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
|
| _version_ | 1866911606744547328 |
|---|---|
| author | Maskey, Utsav Yadav, Sumit Dras, Mark Naseem, Usman |
| author_facet | Maskey, Utsav Yadav, Sumit Dras, Mark Naseem, Usman |
| contents | LLMs increasingly exhibit over-refusal behavior, where safety mechanisms cause models to reject benign instructions that seemingly resemble harmful content. This phenomenon diminishes utility in production applications that repeatedly rely on common prompt templates or applications that frequently rely on LLMs for specific tasks (e.g. sentiment analysis, language translation). Through extensive evaluation, we demonstrate that LLMs persist in refusing inputs containing harmful content, even when they are reframed with tasks that have benign intent. Our mechanistic analysis reveals that LLMs follow distinct "constellation" patterns in embedding space as representations traverse layers, with each NLP task maintaining consistent trajectories that shift predictably between refusal and non-refusal cases. We introduce SafeConstellations, an inference-time trajectory-shifting approach that tracks task-specific trajectory patterns and guides representations toward non-refusal pathways. By selectively guiding model behavior only on tasks prone to over-refusal, our method reduces over-refusals with minimal impact on utility -- offering a principled and conditional approach to mitigating over-refusals. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2508_11290 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | SafeConstellations: Mitigating Over-Refusals in LLMs Through Task-Aware Representation Steering Maskey, Utsav Yadav, Sumit Dras, Mark Naseem, Usman Computation and Language LLMs increasingly exhibit over-refusal behavior, where safety mechanisms cause models to reject benign instructions that seemingly resemble harmful content. This phenomenon diminishes utility in production applications that repeatedly rely on common prompt templates or applications that frequently rely on LLMs for specific tasks (e.g. sentiment analysis, language translation). Through extensive evaluation, we demonstrate that LLMs persist in refusing inputs containing harmful content, even when they are reframed with tasks that have benign intent. Our mechanistic analysis reveals that LLMs follow distinct "constellation" patterns in embedding space as representations traverse layers, with each NLP task maintaining consistent trajectories that shift predictably between refusal and non-refusal cases. We introduce SafeConstellations, an inference-time trajectory-shifting approach that tracks task-specific trajectory patterns and guides representations toward non-refusal pathways. By selectively guiding model behavior only on tasks prone to over-refusal, our method reduces over-refusals with minimal impact on utility -- offering a principled and conditional approach to mitigating over-refusals. |
| title | SafeConstellations: Mitigating Over-Refusals in LLMs Through Task-Aware Representation Steering |
| topic | Computation and Language |
| url | https://arxiv.org/abs/2508.11290 |