Researchers have developed Gradient-Routed Auxiliary Modules (GRAM), a novel pre-training method designed to address the dual-use dilemma in AI development. GRAM allows for the selective disabling of specific capabilities within a single AI model, approximating the effect of training separate models with filtered data at a fraction of the cost. This approach enables fine-grained access control, allowing sensitive knowledge to be restricted to trusted deployments while preserving general performance. Experiments show GRAM effectively isolates capabilities across various domains, including virology and cybersecurity, and maintains this isolation even after fine-tuning. AI
IMPACT Enables more granular control over AI capabilities, potentially mitigating risks associated with dual-use technologies.
RANK_REASON The cluster describes a new pre-training method for AI models published in an academic paper.
- Addie Foote
- AE Studio
- Alex Cloud
- Anthropic
- Cem Anıl Kenar
- Diogo Schwerz de Lucena
- Erick Martinez-Herrera
- Ethan Roland
- Gradient Routed Auxiliary Modules
- gram
- Judd Rosenblatt
- Keenan Pepper
- Mike Vaiana
- Murat Çubuktepe
- Stijn Servaes
- arXiv
- Chinchilla
AI-generated summary · Google Gemini · from 3 sources. How we write summaries →
