Backtracking Improves Generation Safety

TL;DR

This paper introduces a backtracking method using a [RESET] token that enables language models to undo unsafe generations, significantly improving safety without sacrificing helpfulness, and resisting adversarial attacks.

Contribution

The paper proposes a novel backtracking technique with a [RESET] token for language models to recover from unsafe outputs, enhancing safety and robustness.

Findings

Backtracking Llama-3-8B is four times safer than baseline.

Models trained with backtracking maintain helpfulness.

Backtracking provides protection against adversarial attacks.

Abstract

Text generation has a fundamental limitation almost by definition: there is no taking back tokens that have been generated, even when they are clearly problematic. In the context of language model safety, when a partial unsafe generation is produced, language models by their nature tend to happily keep on generating similarly unsafe additional text. This is in fact how safety alignment of frontier models gets circumvented in the wild, despite great efforts in improving their safety. Deviating from the paradigm of approaching safety alignment as prevention (decreasing the probability of harmful responses), we propose backtracking, a technique that allows language models to "undo" and recover from their own unsafe generation through the introduction of a special [RESET] token. Our method can be incorporated into either SFT or DPO training to optimize helpfulness and harmlessness. We show that models trained to backtrack are consistently safer than baseline models: backtracking Llama-3-8B is four times more safe than the baseline model (6.1\% $\to$ 1.5\%) in our evaluations without regression in helpfulness. Our method additionally provides protection against four adversarial attacks including an adaptive attack, despite not being trained to do so.