Lightweight constitutional adapters improve jailbreak defenses without adversarial training

Adapters distilled from synthetic constitution-following text transferred across model checkpoints and remained effective in long-context and multi-turn evaluations.

Big Tech
Adam S. Lowet · Mark Kurzeja

Anthropic Fellows Program · Independent

Research Digest··2 min read
Lowet and Kurzeja convert a written model constitution into low-rank adapters and activation steering vectors using synthetic training data.

The authors generated synthetic corpora from two reference documents: a value-bearing constitution and a control handbook without comparable normative content.

Why this paper

From Anthropic Fellows Program and Independent

In one line

Constitutional adapters, formed by subtracting control-trained from constitution-trained lightweight adapters, improve jailbreak defense and measured alignment without training on harmful requests.

What we could check

  • ·No code link found
  • ·No weights link found
  • ·No dataset link found
  • ·No compute details found
  • ·No stated limitations found
  • ·No benchmark numbers found

Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.

§
newspaper

Research Digest

Articles published under the Zotpaper byline are synthesized from multiple source publications by our AI editor and reviewed by our editorial process. Each story combines reporting from credible outlets to give readers a balanced, comprehensive view.