Αποθηκεύτηκε σε:
| Κύριοι συγγραφείς: | , , , |
|---|---|
| Μορφή: | Preprint |
| Έκδοση: |
2025
|
| Θέματα: | |
| Διαθέσιμο Online: | https://arxiv.org/abs/2502.03708 |
| Ετικέτες: |
Προσθήκη ετικέτας
Δεν υπάρχουν, Καταχωρήστε ετικέτα πρώτοι!
|
| _version_ | 1866908383199625216 |
|---|---|
| author | Beaglehole, Daniel Radhakrishnan, Adityanarayanan Boix-Adserà, Enric Belkin, Mikhail |
| author_facet | Beaglehole, Daniel Radhakrishnan, Adityanarayanan Boix-Adserà, Enric Belkin, Mikhail |
| contents | Modern AI models contain much of human knowledge, yet understanding of their internal representation of this knowledge remains elusive. Characterizing the structure and properties of this representation will lead to improvements in model capabilities and development of effective safeguards. Building on recent advances in feature learning, we develop an effective, scalable approach for extracting linear representations of general concepts in large-scale AI models (language models, vision-language models, and reasoning models). We show how these representations enable model steering, through which we expose vulnerabilities, mitigate misaligned behaviors, and improve model capabilities. Additionally, we demonstrate that concept representations are remarkably transferable across human languages and combinable to enable multi-concept steering. Through quantitative analysis across hundreds of concepts, we find that newer, larger models are more steerable and steering can improve model capabilities beyond standard prompting. We show how concept representations are effective for monitoring misaligned content (hallucinations, toxic content). We demonstrate that predictive models built using concept representations are more accurate for monitoring misaligned content than using models that judge outputs directly. Together, our results illustrate the power of using internal representations to map the knowledge in AI models, advance AI safety, and improve model capabilities. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2502_03708 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Toward universal steering and monitoring of AI models Beaglehole, Daniel Radhakrishnan, Adityanarayanan Boix-Adserà, Enric Belkin, Mikhail Computation and Language Artificial Intelligence Machine Learning Modern AI models contain much of human knowledge, yet understanding of their internal representation of this knowledge remains elusive. Characterizing the structure and properties of this representation will lead to improvements in model capabilities and development of effective safeguards. Building on recent advances in feature learning, we develop an effective, scalable approach for extracting linear representations of general concepts in large-scale AI models (language models, vision-language models, and reasoning models). We show how these representations enable model steering, through which we expose vulnerabilities, mitigate misaligned behaviors, and improve model capabilities. Additionally, we demonstrate that concept representations are remarkably transferable across human languages and combinable to enable multi-concept steering. Through quantitative analysis across hundreds of concepts, we find that newer, larger models are more steerable and steering can improve model capabilities beyond standard prompting. We show how concept representations are effective for monitoring misaligned content (hallucinations, toxic content). We demonstrate that predictive models built using concept representations are more accurate for monitoring misaligned content than using models that judge outputs directly. Together, our results illustrate the power of using internal representations to map the knowledge in AI models, advance AI safety, and improve model capabilities. |
| title | Toward universal steering and monitoring of AI models |
| topic | Computation and Language Artificial Intelligence Machine Learning |
| url | https://arxiv.org/abs/2502.03708 |