The Geometry of Harmfulness in LLMs through Subconcept Probing
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_ | 1866908470252404736 |
|---|---|
| author | Shah, McNair Angeline, Saleena Kumar, Adhitya Rajendra Chheda, Naitik Zhu, Kevin Sharma, Vasu O'Brien, Sean Cai, Will |
| author_facet | Shah, McNair Angeline, Saleena Kumar, Adhitya Rajendra Chheda, Naitik Zhu, Kevin Sharma, Vasu O'Brien, Sean Cai, Will |
| contents | Recent advances in large language models (LLMs) have intensified the need to understand and reliably curb their harmful behaviours. We introduce a multidimensional framework for probing and steering harmful content in model internals. For each of 55 distinct harmfulness subconcepts (e.g., racial hate, employment scams, weapons), we learn a linear probe, yielding 55 interpretable directions in activation space. Collectively, these directions span a harmfulness subspace that we show is strikingly low-rank. We then test ablation of the entire subspace from model internals, as well as steering and ablation in the subspace's dominant direction. We find that dominant direction steering allows for near elimination of harmfulness with a low decrease in utility. Our findings advance the emerging view that concept subspaces provide a scalable lens on LLM behaviour and offer practical tools for the community to audit and harden future generations of language models. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2507_21141 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | The Geometry of Harmfulness in LLMs through Subconcept Probing Shah, McNair Angeline, Saleena Kumar, Adhitya Rajendra Chheda, Naitik Zhu, Kevin Sharma, Vasu O'Brien, Sean Cai, Will Artificial Intelligence Recent advances in large language models (LLMs) have intensified the need to understand and reliably curb their harmful behaviours. We introduce a multidimensional framework for probing and steering harmful content in model internals. For each of 55 distinct harmfulness subconcepts (e.g., racial hate, employment scams, weapons), we learn a linear probe, yielding 55 interpretable directions in activation space. Collectively, these directions span a harmfulness subspace that we show is strikingly low-rank. We then test ablation of the entire subspace from model internals, as well as steering and ablation in the subspace's dominant direction. We find that dominant direction steering allows for near elimination of harmfulness with a low decrease in utility. Our findings advance the emerging view that concept subspaces provide a scalable lens on LLM behaviour and offer practical tools for the community to audit and harden future generations of language models. |
| title | The Geometry of Harmfulness in LLMs through Subconcept Probing |
| topic | Artificial Intelligence |
| url | https://arxiv.org/abs/2507.21141 |