SAEs reveal steerable features in antibody language models

Researchers have explored the use of Sparse Autoencoders (SAEs) to understand and control antibody language models. They found that TopK SAEs can identify biologically relevant features but do not guarantee causal control over generation. Ordered SAEs, however, provide reliable identification of steerable features through a hierarchical structure, though they result in more complex activation patterns. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Introduces new methods for interpreting and steering protein language models, potentially aiding drug discovery and design.

RANK_REASON Academic paper detailing a new methodology for interpreting and controlling protein language models.

Read on arXiv cs.LG →

• arXiv
• Rebonto Haque
• TopK SAEs
• Ordered SAEs

• paper
• other