# Risk-Gated Residual Control

The work introduces BlindBias, a framework for decoding-time jailbreaking through text-only continuation interfaces that permit repeated sampling and assistant-prefix continuation: combining sample-based distribution reconstruction, risk-gated residual control, and speculative multi-token execution, it achieves the highest mean score in 20 of 24 comparisons across four target endpoints and three benchmarks, and shows that selective intervention trades attack effectiveness for less frequent controller invocation.

The work introduces BlindBias, a framework for decoding-time jailbreaking through text-only continuation interfaces that permit repeated sampling and assistant-prefix continuation: combining sample-based distribution reconstruction, risk-gated residual control, and speculative multi-token execution, it achieves the highest mean score in 20 of 24 comparisons across four target endpoints and three benchmarks, and shows that selective intervention trades attack effectiveness for less frequent controller invocation.