Given a prompt $x\in\mathcal{D}_{\text{harmless}}$, we first use the safety teacher model $\pi_{t}$ to sample $N$ refusal trajectories, denoted as $\{r^{(n)}\}_{n=1}^{N}$. The response length of each trajectory is constrained to $H$. Compared to the rollouts of the base model $\pi_{0}$ on harmless instructions, the use of $\pi_{t}$ rollouts produces refusal trajectories, yielding a set of safety tokens with higher purity.

Subsequently, for the $n$-th trajectory, we concatenate it with the input $x$ and feed it into both $\pi_{t}$ and $\pi_{0}$ for a forward pass. This yields the conditional probability distributions at each valid step $j$, represented as $p_{t}(v\mid x,r^{(n)}_{<j})$ and $p_{0}(v\mid x,r^{(n)}_{<j})$. For each token $v\in\mathcal{V}$, we define its contrastive log probability at step $j$ of the $n$-th trajectory as:

A larger positive value of $\Delta^{(x,n)}_{j}(v)$ signifies that the token is more strongly induced by the refusal direction along this trajectory.

For each prompt $x$, trajectory index $n$, and valid step $j$, we select the top-$K^{\prime}$ tokens with the largest $\Delta$ values to form a candidate set, denoted as $\mathcal{C}_{j}^{(x,n)}$. Subsequently, we use an indicator function to aggregate the overall votes:

where $H_{x,n}\leq H$ is the valid sequence length of the $n$-th trajectory for prompt $x$ before emitting an EOT token. The final safety token subset $\mathcal{S}$ is defined as the top-$K$ tokens with the highest votes across the entire vocabulary $\mathcal{V}$:

This discretized voting mechanism, combined with multiple rollouts, effectively prevents extreme $\Delta$ values from dominating the overall ranking.

Reverse KL divergence exhibits a mode-seeking property, making it more suitable for driving the model to converge toward a single, explicit refusal mode. Given a prompt $x\in\mathcal{D}_{\text{harmful}}$, student model $\pi_{s}$ first samples $M$ complete response trajectories $\{y^{(m)}\}_{m=1}^{M}$. To simplify notation, we define the short-hand conditional probabilities for token $v\in\mathcal{V}$ at time step $t$ as $p_{s}(v)\triangleq p_{s}(v\mid x,y^{(m)}_{<t})$ and $p_{t}(v)\triangleq p_{t}(v\mid x,y^{(m)}_{<t})$. In standard OPD, the student model $\pi_{s}$ is updated by minimizing the reverse KL divergence over the vocabulary $\mathcal{V}$:

However, the standard reverse KL divergence still penalizes the entire vocabulary $\mathcal{V}$. It inevitably affects the tokens responsible for general capabilities, leading to performance degradation.

To eliminate the negative impact brought by the global KL penalty, after extracting the safety token subset $\mathcal{S}$, we restrict the calculation of the reverse KL loss to this sparse subset. Consequently, the reverse KL loss restricted to the safety tokens is modified as:

The final training loss is computed by averaging the localized penalty over all valid time steps across the $M$ rollouts for each sample in the batch:

where the expectation is taken over the batch of size $B$ and $M$ rollouts, and $T^{(b,m)}$ denotes the valid generated length of the $m$-th trajectory for the $b$-th sample in the batch. Through this design, the alignment updates are strictly localized to the safety token subset $\mathcal{S}$. This fundamentally eliminates the alignment tax while maintaining stable gradients via multiple rollouts.

To evaluate the effectiveness of our SafeSteer, we apply it to four widely used open-source LLMs: Llama-3-8B-Instruct, Llama-3.2-3B-Instruct (21), Qwen2.5-7B-Instruct (30), and Qwen3-4B-Instruct-2507 (37).

We compare our SafeSteer against several baseline approaches: MoCAN (14), W-DOOR (42), BFPO (39), NSPO (26), and DPO-Mix. Following prior work (26), the DPO-Mix variant is configured with a 50/50 mixing ratio of helpfulness and safety preferences. For training data, our method uses only 100 instructions randomly sampled from the unsafe subset of PKU-SafeRLHF(8), whereas NSPO samples 40% of the entire dataset and the remaining baselines are trained on the full dataset.

We comprehensively evaluate our SafeSteer across both safety and general capabilities. For safety benchmarks, we report the Attack Success Rate (ASR) on two categories: (i) Harmful queries, using AdvBench (46), PKU-SafeRLHF (8), HarmBench (24), JailbreakBench (5), and SORRY-Bench (36); and (ii) Red-team queries, utilizing HarmfulQA (3) and ALERT (35). To validate that SafeSteer can preserve the model’s general capabilities, we evaluate its performance across four dimensions: (i) General Knowledge on MMLU (12); (ii) Instruction Following on AlpacaEval (16); (iii) Mathematical Reasoning on GSM8K (7) and MATH500 (18); and (iv) Coding on HumanEval (6).

To extract the refusal direction, we utilize 160 harmful and 160 harmless instructions provided by previous work (2). During safety token selection, we use the same 160 harmless samples. Additionally, we set the safety token subset size to $|\mathcal{S}|=50$. In the SafeSteer training stage, we apply a learning rate of 1e-6 to Llama models and 1e-5 to Qwen models, with $M=8$ rollouts. For safety evaluation, we employ Llama-Guard-4-12B as the default judge across all benchmarks, except SORRY-Bench (36), which relies on its natively fine-tuned scoring model. To measure general capabilities, we use DeepSeek-V4-Flash (9) to report the Win Rate compared to Alpaca-7B on AlpacaEval, while using lm-eval (10) for the remaining benchmarks. During model inference, we set do_sample=False at temperature 0. When temperature is 1.0, we run the evaluation three times and report the average. More details on training and evaluation are in Appendix A.

Under both temperature settings of 0 and 1.0, SafeSteer attains the lowest ASR among all methods on the Qwen family, by a clear margin over the strongest baselines. At temperature 0, SafeSteer pushes the average ASR down to 0.91% on Qwen3-4B-Instruct, compared to 1.13% for the next-best MoCAN and 2.87% for the base model. Likewise, it lowers the ASR to 1.48% on Qwen2.5-7B-Instruct, which represents a 2.5$\times$ reduction against 3.75% for the strongest baseline BFPO. Similarly, for the Llama family, SafeSteer remains highly competitive: it not only reduces the ASR on Llama-3.2-3B-Instruct from 5.26% to 4.06%, but also achieves the second-lowest ASR on Llama-3-8B-Instruct. Notably, while some baselines slightly surpass SafeSteer in safety, most exhibit a significant degradation in general capabilities. For example, W-DOOR collapses both Llama models to near half of their base capability, dropping the average score from 63.61 to 34.04 on Llama-3.2-3B-Instruct and from 63.64 to 37.82 on Llama-3-8B-Instruct. Consistent with the observation in NSPO (26), we find that DPO-Mix can even increase ASR on both Llama-3-8B-Instruct and Qwen2.5-7B-Instruct, indicating that the alignment tax cannot be solved by naive data mixing.

Without relying on any general-purpose data necessitated by other methods, SafeSteer can still maintain the model’s general capabilities. For the Qwen family, at both temperatures 0 and 1.0, SafeSteer preserves general capabilities with negligible loss: at temperature 0, the average score shifts slightly from 71.68 to 71.39 on Qwen3-4B-Instruct and from 68.90 to 69.17 on Qwen2.5-7B-Instruct, effectively retaining both models’ original ability. For the Llama family at temperature 0, SafeSteer limits the performance degradation to 0.75 points on Llama-3.2-3B-Instruct and outperforms all baselines on Llama-3-8B-Instruct. SafeSteer also ranks second among all methods on Llama-3-8B-Instruct at temperature 1.0. This demonstrates that our SafeSteer successfully decouples the safety feature from the general capability space.

Self-distillation (41) often relies on system prompts to construct a safety teacher. However, such prompt-driven approaches often struggle to maintain stable refusal behaviors, failing to provide consistent supervisory signals during distillation. As shown in Table 3, replacing our activation-steered teacher with a prompt-based version (see Appendix C) results in a higher ASR. In contrast, by inducing refusal behaviors directly within the model’s representation space, our steered teacher can provide robust refusal signals, consistently reducing the ASR to near 0.00%.

We compare SafeSteer against two variants to verify our localized penalty design:

1. 1.

   w/ forward KL, which uses forward KL instead of reverse KL to compute the loss.

2. 2.

   w/o safety token, which penalizes the entire vocabulary $\mathcal{V}$.

As shown in Table 4, the w/o safety token variant noticeably degrades models’ performance on MATH and HumanEval. It proves that applying the penalty across the entire vocabulary negatively impacts general capabilities. Additionally, w/ forward KL underperforms reverse KL, confirming that the mode-seeking property of reverse KL better suits learning explicit refusal behaviors.

Using the base model $\pi_{0}$ to generate rollout trajectories for safety token selection causes a catastrophic degradation in safety capabilities. Specifically, training a Llama-3-8B-Instruct variant under this setting increases its JailbreakBench ASR from 1.0% to 75.0%. We observe that the resulting safety token subset $\mathcal{S}$ is dominated by formatting tokens (e.g., “\n\n” and “\r\n\r\n”) rather than genuine safety tokens. Consequently, the model merely learns to insert meaningless paragraph breaks instead of refusing. This failure stems from the shallow safety alignment of Llama-3-8B-Instruct. Because its safety guardrails are concentrated on the very first few output tokens, the unsteered $\pi_{0}$ lacks the capability to maintain refusal behaviors at deeper positions, causing the contrastive log probability computed along this trajectory to fail in capturing safety tokens. Conversely, using $\pi_{t}$ for rollouts avoids this vulnerability: the steered trajectory forces the model to maintain a coherent refusal signal throughout the sequence, ensuring that the contrastive log probability correctly localizes genuine safety tokens. While this complete collapse was most pronounced on Llama-3-8B-Instruct, it highlights the necessity of using $\pi_{t}$ to guarantee stable token extraction across different base models.

The response length $H$ controls the scope of the safety token selection (see Section 3.3). Figure 4 compares the top-voted tokens in $\mathcal{S}$ at $H=1$ with those at $H=7$. When the response length is constrained to $H=1$, the safety token selection method only captures initial refusal words (e.g., “I”, “Sorry”, “Unfortunately”). This essentially encodes how to begin refusing at a superficial level. As $H$ grows, the response extends into the model’s explanation for refusing. Consequently, the top-voted tokens in $\mathcal{S}$ shift dramatically toward core semantic safety concepts (e.g., “illegal”, “unethical”, “harmful”) that encode why to refuse. This dynamic transition demonstrates that an appropriate $H$ is essential for “deep safety alignment” (28), enabling the model to grasp safety at a semantic level rather than relying on superficial initial tokens.

When calculating the token-level localized KL divergence, we use the raw probability slices $p_{t}(v\mid x,y_{<t})$ and $p_{s}(v\mid x,y_{<t})$ for $v\in\mathcal{S}$ without re-normalizing them into a proper distribution via softmax (see Section 3.4). Empirically, we find this design to be critical. As shown in Table 5, forcing re-normalization strictly over $\mathcal{S}$ causes drastic degradation in general capabilities, by up to 4.76 points on Llama-3.2-3B-Instruct and 3.04 on Qwen2.5-7B-Instruct. This collapse is fundamentally driven by the loss of the absolute magnitude anchor. Without re-normalization, the standard Reverse KL naturally constrains the absolute probability mass of the student $p_{s}(v)$. After re-normalization, the optimizer aggressively inflates the logits of tokens in $\mathcal{S}$, crowding out probability mass from the remaining tokens and distorting the student model’s general-capability distribution.