CBF-LLM: Safe Control for LLM Alignment

TL;DR

This paper introduces CBF-LLM, a control barrier function-based safety filter for aligning large language models like Llama 3, reducing unsafe outputs and improving user alignment with minimal interventions.

Contribution

It presents a novel control-based framework using CBFs to ensure safe and aligned text generation in LLMs, with implementation and experimental validation.

Findings

Effective safety filtering reduces unsafe outputs.

Fewer interventions needed for alignment.

Framework applicable to large models like Llama 3.

Abstract

This paper proposes a control-based framework for aligning large language models (LLMs) by leveraging a control barrier function (CBF) to ensure user-desirable text generation. The presented framework applies the safety filter, designed based on the CBF, to the output generation of the baseline LLM, i.e., the sequence of the token, with the aim of intervening in the generated text. The overall text-generation system is implemented with Llama 3 and a RoBERTa model, and the source code is available at https://github.com/Mya-Mya/CBF-LLM. The experiment demonstrates its control ability and effectiveness in reducing the number of interventions needed for user-specified alignment tasks.