Refusal Steering: Fine-grained Control over LLM Refusal Behaviour for Sensitive Topics

TL;DR

Refusal Steering enables fine-grained, inference-time control over LLM refusal behavior on politically sensitive topics without retraining. It replaces pattern-based refusals with an LLM-as-a-judge to produce refusal confidence scores and introduces ridge-regularized steering vectors (RMD, WRMD) to isolate the refusal–compliance direction, enabling targeted modification across model layers. Experiments on Qwen3-Next-80B-A3B-Thinking show substantial reductions in political refusals (down to ~23.8%) while preserving safety on JailbreakBench and maintaining general capabilities, with the approach generalizing to 4B and 80B architectures. Steering signals are shown to concentrate in deeper layers and spread across many dimensions, supporting robust, multi-dimensional control and offering a practical path to transparent, inference-time moderation.

Abstract

We introduce Refusal Steering, an inference-time method to exercise fine-grained control over Large Language Models refusal behaviour on politically sensitive topics without retraining. We replace fragile pattern-based refusal detection with an LLM-as-a-judge that assigns refusal confidence scores and we propose a ridge-regularized variant to compute steering vectors that better isolate the refusal--compliance direction. On Qwen3-Next-80B-A3B-Thinking, our method removes the refusal behaviour of the model around politically sensitive topics while maintaining safety on JailbreakBench and near-baseline performance on general benchmarks. The approach generalizes across 4B and 80B models and can also induce targeted refusals when desired. We analize the steering vectors and show that refusal signals concentrate in deeper layers of the transformer and are distributed across many dimensions. Together, these results demonstrate that activation steering can remove political refusal behaviour while retaining safety alignment for harmful content, offering a practical path to controllable, transparent moderation at inference time.

Figures (7)

• Figure 1: Example of our steering method applied to Qwen3-Next-80B-A3B-Thinking model to either remove political refusal behaviour or introduce it.
• Figure 2: Refusal answer by Qwen3-Next-80B-A3B-Thinking. The model doesn't explicitly refuse the question, it produces a state-aligned answer.
• Figure 3: 2D PCA projection of last-token activations at layer 42 for positive (refusal), negative (non-refusal), and neutral examples.
• Figure 4: Rejection rates for world politics and general prompts applying different positive steering coefficients to induce refusal behavior. The baseline (green) shows the model's natural rejection rate.
• Figure 5: Steering vector analysis for Qwen3-Next-80B-A3B-Thinking on the Extended dataset. (a) Layer-wise correlation between activation projections and refusal confidence. (b) Magnitude distribution of steering vector at layer 42 (log scale). (c) Correlation heatmap across all layers. (d) Full steering vector heatmap for layer 42.