Back to GitHub trending

fars-analemma/transferable-defensive-tokens

Python15 stars0 forks
View on GitHubShare on X

What it does

AlignDefTok enables the transfer of pre-optimized DefensiveTokens between closely related LLMs without the need for training, using Orthogonal Procrustes alignment. This approach enhances the robustness of language models against adversarial attacks, making it significant for AI safety and security.

Star history

Not enough history yet — 1 day(s) recorded. The daily snapshot builds this up.

Tracking

Last trending
2026-02-21

Creator kit

Hook

Discover how to enhance LLM security effortlessly with AlignDefTok's innovative token transfer method!

Content angles

  • Create a tutorial on implementing Orthogonal Procrustes alignment for LLMs.
  • Discuss the implications of using pre-optimized tokens in AI safety.
  • Analyze the performance of DefensiveTokens across different LLMs.

Who should care

AI researchers, developers working on language models, and cybersecurity professionals.