Skip to content
InfoResearchPreprintLLM-specific

Alignment with Reflect-Reasoning: Learning How To Follow Principles

Published
Record updated
View JSON

Summary

Reflect-Reasoning is a training method that teaches LLMs how to reason about alignment to principles, rather than only which principles to follow. Across five models, it outperforms system prompting and explicit prompting for in-context reflective reasoning, including on out-of-distribution medical and financial principles. The authors also report fewer rare but serious principle violations, no erosion of safety engineering, and robustness against jailbreaking with adversarial constitutions.