betaFeedback

WorksComputer Science

Restoring General and Improving Domain-Specific Safety in Fine-Tuned LLMs Without Retraining

Abstract

Safety-aligned language models are known to suffer degradation in refusal behavior following domain-specific fine-tuning, even when the fine-tuning data is entirely benign. The predominant mitigation strategy, preference optimization on the fine-tuned model, is computationally expensive and requires domain-specific preference data that is often unavailable in practice. We propose SPECTRA, a training-free, weight-level method that restores refusal behavior by transferring the base model's refusal direction into the fine-tuned model's weights via an activation-aware, low-rank update, requiring no gradient-based optimization or domain-specific safety data. The low-rank design restricts the intervention to the subspace most causally active on harmful inputs, leaving benign behavior largely undisturbed by construction. Empirically, SPECTRA substantially improves refusal behavior across a comprehensive safety evaluation, while preserving general coherence, domain-specific knowledge, and producing negligible change in false refusal rates.

Introduction

Modern transformer-based large language models (LLMs) ([19] (Touvron et al. 2023a; Team et al. 2024; Team et al. 2023)) are typically safety-aligned before deployment to ensure safe and trustworthy behavior ([16, 6] (Rafailov et al. 2023)). Recent work has shown that preference-based alignment can lead to meaningful improvements across multiple dimensions of model safety ([2] (Zhang et al. 2025)). Despite this progress, deploying LLMs in specialized settings often requires domain-specific fine-tuning ([8, 3]), which can inadvertently degrade the very safety properties instilled during alignment, even when the fine-tuning data is entirely benign ([17] (Qi et al. 2024b)).

This tension between domain adaptation and safety preservation presents a fundamental challenge for the practical deployment of aligned models.

The vulnerability of aligned models to fine-tuning is well documented. Even a small amount of adversarial data in the fine-tuning set can cause a dramatic drop in safety, while fine-tuning on benign data alone has been shown to erode refusal behavior ([17] (Qi et al. 2024b)). Existing approaches to mitigating this degradation include additional preference optimization ((Han et al., 2024)), curating domain-specific safety data for further tuning ((Jan et al., 2024)), modifying the fine-tuning objective ((Qi et al., 2024a)), embedding safety backdoors via prefix prompts ((Wang et al., 2024a)), and preventing alignment drift through optimization-state separation ((Huang et al., 2024a)). Although effective, these methods typically require additional gradient-based computation, modified training procedures, or domain-specific calibration data, which limits their practicality in resource-constrained or data-scarce settings.

A complementary line of research in mechanistic interpretability has revealed that safety-relevant behaviors in language models are encoded as linear directions in the residual stream. Arditi et al. [5] identified a refusal direction and showed that ablating it from the model’s weights constitutes an effective white-box jailbreak, demonstrating that refusal behavior is concretely encoded in weight-space structure. Subsequent work has discovered related steering vectors via sparse autoencoders ([14] (O’Brien et al. 2024)) and applied them at inference time to elicit or suppress specific behaviors ((O’Brien et al., 2025; Lee et al., 2025; Panickssery et al., 2023)). However, inference-time activation steering has fundamental limitations: it tends to increase false refusal rates ((O’Brien et al., 2025)), requires careful per-model hyperparameter tuning, and introduces runtime overhead that complicates deployment. These two lines of research together motivate our approach. The problem of fine-tuning-induced safety degradation calls for a lightweight, training-free remedy, while mechanistic interpretability offers tools for identifying and manipulating the weight-space structures that mediate refusal behavior. Rather than retraining the model or steering activations at inference time, we propose to directly inject the base model’s refusal direction into the weights of the fine-tuned model. To avoid indiscriminate perturbations to model parameters, which could degrade unrelated capabilities, the injection is restricted to an activation-aware, low-rank subspace of the weight matrix that is most causally active on harmful inputs. This ensures the intervention strengthens refusal behavior on harmful content while remaining largely dormant on benign inputs, without requiring any runtime modification. We refer to the resulting method as SPECTRA (Spectral Refusal Alignment). The key components and findings of our work are as follows:

  1. We propose SPECTRA, a training-free, closed-form weight-space intervention for restoring safety after domain-specific fine-tuning. The method injects a refusal direction from the safety-aligned base model into the weights of the fine-tuned model, requiring no additional gradient updates, domain-specific safety annotations, or inference-time steering.

  1. We develop an activation-aware low-rank update that restricts the intervention to the subspace of the weight matrix most active on harmful inputs. This design strengthens refusal behavior where it is needed while minimizing changes to model behavior on benign inputs.

  1. We empirically show that SPECTRA substantially improves safety across multiple fine-tuned models and evaluation benchmarks while preserving linguistic quality and domain-specific capabilities. Notably, these gains are not driven by indiscriminate over-refusal, as false refusal rates remain statistically negligible.

  1. We provide evidence that refusal directions transfer across fine-tuned variants of a safety-aligned base model, suggesting that fine-tuning preserves a shared geometric structure for refusal and that safety can be restored through targeted weight-space manipulation rather than full retraining.

Problem Formulation

Background

Transformer notation and residual-stream activations. Consider a decoder-only Transformer ((Radford et al., 2019; Vaswani et al., 2017)) that maps token a sequence of token indices t=(t1,…,tn)∈Vn\mathbf{t} = (t_1,\ldots,t_n) \in\mathcal{V}^n to token-level output distributions y∈Rn×∣V∣\mathbf{y} \in\mathbb{R}^{n \times|\mathcal{V}|}. Let xi(l)∈Rdmodel\mathbf{x}_i^{(l)} \in\mathbb{R}^{d_{\mathrm{model}}} denote the residual-stream activation for token position ii at the input of layer ll. The embedding layer initializes xi(1)=Embed⁡(ti)\mathbf{x}_i^{(1)} = \operatorname{Embed}(t_i). For each layer l∈[L]l \in[L], the model applies causal self-attention followed by an MLP, each with a residual connection:

hi(l)=xi(l)+Attn⁡(l)(x1:i(l)),(1)\mathbf{h}_i^{(l)} = \mathbf{x}_i^{(l)} + \operatorname{Attn}^{(l)}\left(\mathbf{x}_{1:i}^{(l)}\right), \tag*{(1)}
xi(l+1)=hi(l)+MLP⁡(l)(hi(l)).(2)\mathbf{x}_i^{(l+1)} = \mathbf{h}_i^{(l)} + \operatorname{MLP}^{(l)}\left(\mathbf{h}_i^{(l)}\right). \tag*{(2)}

The final residual-stream activations are projected back to vocabulary space via logits⁡i=Unembed⁡(xi(L+1))\operatorname{logits}_i = \operatorname{Unembed}\left(\mathbf{x}_i^{(L+1)}\right), producing yi=softmax⁡(logits⁡i)\mathbf{y}_i = \operatorname{softmax}(\operatorname{logits}_i) ([5]).

From activations to a refusal direction. A growing body of interpretability work has shown that high-level behavioral properties of language models, including safety-relevant ones such as refusal, are often linearly encoded in the residual stream ([7, 5] (Marks and Tegmark 2023)). Motivated by this, we seek a single direction in activation space that is causally predictive of refusal behavior. Following the difference-in-means methodology ([7, 5] (Marks and Tegmark 2023)), we compute class-conditional mean activations over a labeled training set of harmful and harmless prompts. For layer ll and token position ii:

μi(l)=Et∼Dharmful(train)[xi(l)(t)],\boldsymbol{\mu}_i^{(l)} = \mathbb{E}_{\mathbf{t} \sim\mathcal{D}_{\mathrm{harmful}}^{(\mathrm{train})}}\left[\mathbf{x}_i^{(l)}(\mathbf{t})\right],
νi(l)=Et∼Dharmless(train)[xi(l)(t)].\boldsymbol{\nu}_i^{(l)} = \mathbb{E}_{\mathbf{t} \sim\mathcal{D}_{\mathrm{harmless}}^{(\mathrm{train})}}\left[\mathbf{x}_i^{(l)}(\mathbf{t})\right].

The candidate refusal direction at (l,i)(l,i) is defined as the centroid difference ri(l)=μi(l)−νi(l)\mathbf{r}_i^{(l)} = \boldsymbol{\mu}_i^{(l)} - \boldsymbol{\nu}_i^{(l)}.

Selection the best refusal direction from the candidates The construction above yields a family of candidate directions indexed by layer l∈[L]l \in[L] and token position i∈Ii \in I. To select a single, causally effective refusal direction, we evaluate each candidate through intervention experiments on a held-out validation set. A direction is considered effective if (i) ablating it from activations induced by harmful prompts in Dharmful(val)D_{\mathrm{harmful}}^{(\mathrm{val})} should suppress refusal, and (ii) injecting it into activations induced by harmless prompts in Dharmless(val)D_{\mathrm{harmless}}^{(\mathrm{val})} should elicit refusal. We select the candidate (i∗,l∗)(i^*, l^*) that maximizes this combined causal effect, and denote the resulting direction as r:=ri∗(l∗)\mathbf{r} := \mathbf{r}_{i^*}^{(l^*)}, with normalized version r^\hat{\mathbf{r}}.

Research Objective

Safety-aligned language models exhibit strong refusal behavior when prompted with harmful requests. However, this behavior is known to degrade after domain-specific fine-tuning ([17] (Qi et al. 2024b)): when a safety-aligned base model M1M_1 is fine-tuned on domain-specific data to obtain M2M_2, the resulting model often exhibits a substantially lower refusal rate on harmful prompts. This motivates the central question of our work:

Can we restore or strengthen refusal behavior in a fine-tuned model without any retraining?

Arditi et al. [5] showed that ablating the refusal direction directly from a model’s output projection weights, rather than from activations at inference time, produces a persistent white-box jailbreak that reliably bypasses safety guardrails. This suggests that refusal behavior is not merely an emergent property of activations, but is concretely encoded in specific weight-space structures. We build on this insight in the opposite direction: rather than removing weight-space structure to weaken safety, we propose to inject it to restore safety. Specifically, given a fine-tuned model M2M_2 with degraded refusal behavior, our goal is to transfer the refusal direction of the base model M1M_1 directly into the weight matrices of M2M_2, achieving a training-free, weight-level safety repair that requires no inference-time activation manipulation.

Formally, let r1\mathbf{r}_1 and r2\mathbf{r}_2 be the refusal directions of M1M_1 and M2M_2, respectively. We seek a targeted, gradient-free modification to a subset of M2M_2’s parameters such that refusal behavior improves on harmful prompts from Dharmful(val)D_{\mathrm{harmful}}^{(\mathrm{val})}, while performance on benign prompts from Dharmless(val)D_{\mathrm{harmless}}^{(\mathrm{val})} is preserved.

Method: Spectral Refusal Alignment

A naïve implementation of this idea would inject r2\mathbf{r}_2 into the full weight matrices of M2M_2. However, such a full-matrix intervention is problematic. Recent work has identified the existence of superweights in large language models: a small number of individual parameters with disproportionate influence over model behavior (Qin et al., 2026). Indiscriminate perturbation of weight matrices risks disrupting these superweights, causing catastrophic degradation in capabilities unrelated to safety. This motivates our approach: rather than modifying the full weight matrix, we restrict our intervention to the low-dimensional subspace most causally active on harmful inputs, thereby preserving benign behavior by design. We refer to the resulting method as Spectral Refusal Alignment (SPECTRA).

Specifically, let W∈Rdout×dinW \in\mathbb{R}^{d_{\mathrm{out}} \times d_{\mathrm{in}}} denote a weight matrix of interest in M2M_2 (either an attention output projection or an MLP projection) and let Xin∈Rdin×nX_{\mathrm{in}} \in\mathbb{R}^{d_{\mathrm{in}} \times n} denote the matrix of pre-projection activations collected by forwarding a set of harmful prompts (e.g., ALIGN dataset (Wei et al., 2024) in our experiments) through M2M_2, where each column corresponds to a response token activation. Our method proceeds in two steps.

Step 1: Activation-Aware Low-Rank Decomposition. To identify the subspace of WW most active on harmful inputs, we follow (Wei et al., 2024; Hsu et al., 2021; Yuan et al., 2023) and seek a rank-ρ\rho approximation W~\widetilde{W} that minimizes the distortion to the output specifically on harmful-prompt activations:

W~=arg⁡min⁡rank⁡W~≤ρ ∥WXin−W~Xin∥F2.\widetilde{W} = \underset{\operatorname{rank}\widetilde{W}\leq\rho}{\arg\min}\ \left\|WX_{\mathrm{in}}-\widetilde{W}X_{\mathrm{in}}\right\|_{F}^{2}.

This is solved by performing SVD on the output matrix WXin∈Rdout×nWX_{\mathrm{in}} \in\mathbb{R}^{d_{\mathrm{out}}\times n}:

UΣVT≈WXin,U\Sigma V^{\mathsf{T}} \approx WX_{\mathrm{in}},

where U∈Rdout×ρU \in\mathbb{R}^{d_{\mathrm{out}}\times\rho} is the orthogonal matrix corresponding to the top ρ\rho left singular vectors. The minimizer in closed form is:

W~=ΠW, where Π=UUT\widetilde{W} = \Pi W,\ \text{where }\Pi= UU^{\mathsf{T}}

is the orthogonal projector onto the dominant output subspace induced by harmful activations. Intuitively, W~\widetilde{W} captures the component of WW that is most "active" when the model processes harmful content, which is precisely the subspace we wish to steer.

Step 2: Spectral Refusal Injection. Having identified the safety-relevant subspace via W~\widetilde{W}, we inject the refusal direction r1\mathbf{r}_1 of base model M1M_1 by modifying WW as follows:

W′=W+αr1r1TW~(3)W' = W + \alpha\mathbf{r}_1\mathbf{r}_1^{\mathsf{T}}\widetilde{W} \tag*{(3)}

where α>0\alpha> 0 is a scalar steering hyperparameter. The update αr1r1TW~\alpha\mathbf{r}_1\mathbf{r}_1^{\mathsf{T}}\widetilde{W} adds a rank-one perturbation in the direction of r1r_1 to the weight matrix, but critically, this perturbation is gated by W~\widetilde{W}: it only modifies the output along r1r_1 for inputs that activate the harmful subspace Π\Pi, rather than uniformly modifying for all inputs.

Equivalence to Conditional Activation Steering. The following theorem shows that our weight-level update is equivalent to a data-adaptive form of activation steering, whose strength automatically scales with the extent to which the input representation lies in the harmful-activation subspace.

Theorem 1 (SPECTRA as conditional activation steering). Let x∈Rdinx \in\mathbb{R}^{d_{\mathrm{in}}} be an input activation, and let h=Wx∈Rdouth = Wx \in\mathbb{R}^{d_{\mathrm{out}}} denote the corresponding pre-update output. Under the weight update (3), the modified output satisfies:

h′=h+β⋅r1,h' = h + \beta\cdot r_1,

where β=α⋅r1TΠh\beta= \alpha\cdot r_1^{\mathsf{T}}\Pi h is a scalar that depends on the alignment of hh with the harmful subspace Π\Pi.

Proof. Substituting the definition of W′W' and W~=ΠW\widetilde{W} = \Pi W:

h′=W′⋅x=Wx+αr1r1TΠWx=h+(α⋅r1TΠh)r1=h+β⋅r1,(4)\begin{aligned} h' &= W' \cdot x = Wx + \alpha r_1r_1^{\mathsf{T}}\Pi Wx \\ &= h + (\alpha\cdot r_1^{\mathsf{T}}\Pi h)r_1 = h + \beta\cdot r_1, \tag*{(4)} \end{aligned}

where β=α⋅r1TΠh\beta= \alpha\cdot r_1^{\mathsf{T}}\Pi h is a scalar. □

Theorem 1 reveals the key property that makes SPECTRA both effective and capability-preserving. When xx is induced by a harmful prompt, h=Wxh = Wx expected to lie substantially within the span of UU, since UU is constructed from harmful-prompt activations. This yields a large β\beta and therefore strong steering toward the refusal direction. Conversely, when xx is from a benign prompt, hh is less likely to align with uu, causing β≈0\beta\approx0 and leaving the output effectively unchanged. In this sense, SPECTRA is geometrically self-gating: the strength of the refusal intervention scales automatically with how harmful the input representation is, without requiring any explicit runtime classifier or threshold.

Experimental Setup

Models and domain fine-tuned variants. We study safety-aligned, widely used LLMs: LLAMA2-CHAT-7B (Touvron et al., 2023b) and GEMMA2-9B-IT (Team et al., 2024), together with publicly available domain fine-tuned variants. For LLAMA2-CHAT-7B, we include medical (Rohanian et al., 2024), law ([8]), and finance ([8]) models. For GEMMA2-9B-IT, we include medical (OpenMeditron, 2024) and finance ([1]) variants.

Coherence evaluation. We evaluate model coherence using zero-shot accuracy on both general-purpose and domain-specific benchmarks.

  • General coherence: We use five tasks from EleutherAI’s LM Harness ([18]): HellaSwag (Zellers et al., 2019), BoolQ ([11]), RTE (Wang et al., 2019), ARC Challenge ([12]), Winogrande (Sakaguchi et al., 2021).

  • Domain coherence: We use standard benchmarks in each domain: MedMCQA (Pal et al., 2022) and PubMedQA (Jin et al., 2019) for medical, FinanceBench (Islam et al., 2023) for finance, and LegalBench (Guha et al., 2023) for *law`; for LegalBench we report results on the Abercrombie and Hearsay tasks.

Safety evaluation and threat models. We evaluate safety using attack success rate (ASR)1 under harmful instructions, reporting both general and domain-specific safety.

  • General safety: We prompt each model with AdvBench-Eval (the first 100 prompts from AdvBench) and compute ASR as the fraction of prompts for which the model’s response does not exhibit refusal patterns, following Zou et al. (2023b). Following Wei et al. (2024), we evaluate two threat models: standard decoding (ASRvanilla_{\mathrm{vanilla}}) and adversarial decoding (ASRAdv−Decoding_{\mathrm{Adv-Decoding}}) (Huang et al., 2024b), where the attacker manipulates the decoding process. For ASRAdv−Decoding_{\mathrm{Adv-Decoding}}, we additionally report results with and without the chat template wrapper, denoted ASRAdv−DecodingT^{\mathrm{T}}_{\mathrm{Adv-Decoding}} and ASRAdv−DecodingX^{\mathrm{X}}_{\mathrm{Adv-Decoding}}, respectively.2

  • Domain safety: For the medical domain, we evaluate ASR on MedSafeEval (Han et al., 2024). For law and finance, where standardized domain safety benchmarks are limited, we construct LawSafeEval (201 prompts) and FinSafeEval (142 prompts): we first generate 2,000 candidate safety-related prompts per domain by prompting Gemini 2.5 Pro ([13]) with

domain-specific safe prompts as references, and then manually filter and label candidates to select the final prompt sets.

Jailbreak robustness (adaptive safety evaluation). In addition to ASR on fixed prompt sets, we evaluate robustness to adaptive jailbreak attacks, including GCG (Zou et al., 2023b), GPTFuzz (Yu et al., 2023), and TAP (Mehrotra et al., 2024).

Refusal labeling and ASR computation. We use the HarmBench Llama2-13B Classifier (Mazeika et al., 2024) to label outputs as refusal (refusal-score = 1) or successful attack (refusal-score = 0), following prior work (Lermen et al., 2023; Liu et al., 2024; Robey et al., 2023; Shah et al., 2023; Xu et al., 2023; Zou et al., 2023b). For TAP (Mehrotra et al., 2024) and GPTFuzz (Yu et al., 2023), we use GPT-3.5-Turbo (OpenAI, 2023) to assign refusal scores.

Datasets for estimating refusal direction and safety-critical ranks. In addition to evaluation benchmarks, our method uses auxiliary datasets to estimate (i) a refusal direction in activation space and (ii) safety-critical low-rank structure in fine-tuned weight matrices.

  • Refusal direction estimation. We follow the procedure outlined by Arditi et al. [5], see Appendix § 18.

  • Safety-critical ranks isolation. We use the ALIGN (Wei et al., 2024) dataset to isolate safety-critical ranks in fine-tuned weight matrices.

On the efficacy of refusal direction for inducing safety

We first ask which refusal direction to inject, and into which layer. Let r1r_1 be the refusal direction extracted at layer l1l_1 of the base model M1M_1, and r2r_2 the one extracted at layer l2l_2 of the fine-tuned model M2M_2; fine-tuning can shift the layer and token position at which the direction is found (Appendix § 17.§ 1). The two natural choices are to steer layer l1l_1 with r1r_1, or layer l2l_2 with r2r_2. We additionally test steering layer l2l_2 with r1r_1, since the layer that carries refusal may move during fine-tuning while the base model’s direction can still apply. We also include r2−r1r_2-r_1 as a control that combines both directions. To isolate this choice, we inject the direction into the full attention-output and MLP projection weights of the target layer, without any low-rank

approximation:

W′=W+α rrTW,(5)W' = W + \alpha\, r r^{\mathsf{T}} W, \tag*{(5)}

a special case of SPECTRA (Eq. (3)) with W~=W\widetilde{W}=W. We evaluate every direction–layer pair on Llama-2-7b-chat fine-tuned on medical data, using our standard safety metrics and GCG (Zou et al., 2023b) (Table 1).

DirectionLayerASRAdv−Dec._{\mathrm{Adv-Dec.}}ASRVanilla_{\mathrm{Vanilla}}ASRAdv−Dec.×^{\times}_{\mathrm{Adv-Dec.}}GCG
No steering (base FT)1.40.018.431
r1r_1l2l_2 (default)2.8010.84
r1r_1l1l_13.21.011.620
r2r_2l2l_22.00.016.425
r2r_2l1l_11.81.018.627
r2−r1r_2-r_1l2l_28.03.013.64
r2−r1r_2-r_1l1l_14.61.014.418

Table 1. Full-matrix injection (Eq. (5)) of each refusal direction into the attention and MLP projections of a given layer, on medical Llama-2-7b-chat. α=0.015\alpha=0.015 (attention) and 0.01 (MLP). Lower is better; bold marks the best value per column among steered variants.

Injecting r1r_1 at l2l_2 gives the lowest ASRAdv−Dec.×^{\times}_{\mathrm{Adv-Dec.}} and, together with r2−r1r_2-r_1 at l2l_2, the lowest GCG ASR (4 vs. 31 unsteered), while staying near zero on other settings. Among the two l2l_2 variants with low GCG ASR, only r1r_1 avoids the rise on ASRVanilla_{\mathrm{Vanilla}} and ASRAdv−Dec.I^{I}_{\mathrm{Adv-Dec.}} seen with r2−r1r_2-r_1. We therefore use r1r_1 at layer l2l_2 in all subsequent experiments.

Why does the base model’s direction work better? At the same layer l2l_2, injecting r1r_1 outperforms injecting r2r_2. To see why, we compare M1M_1 with M2M_2 (Llama-2-7b-chat and its medical fine-tune) at layer l2=13l_2=13 on 200 harmful HARMFUL_BEHAVIORS prompts (Zou et al., 2023b) and 200 harmless Alpaca prompts (Taori et al., 2023), using three measures. Separation is the discriminability d′d' (difference of class means over the pooled standard deviation) between harmful and harmless activations projected onto a direction at the last prompt token. We compare r1r_1, r2r_2, and a held-out difference-in-means (DiM) direction computed on a random half of the prompts and scored on the other half. Write is the mean of rTWxr^{\mathsf{T}}Wx over harmful prompts, summed over the attention and MLP output projections WW of layer l2l_2 with inputs xx, evaluated at each direction’s extraction position (token −1-1 for r1r_1, −5-5 for r2r_2). We also compute rT(W1−W2)xr^{\mathsf{T}}(W_1-W_2)x on M2M_2’s inputs to separate the effect of the layer’s weights from that of its inputs. Gate selectivity is the ratio of ∣rTWx∣\lvert r^{\mathsf{T}}W x\rvert on harmful versus harmless prompts, where SPECTRA’s update adds α r(r⊤ΠWx)\alpha\,r\left(r^\top\Pi W x\right) to the output and Π\Pi projects onto the harmful subspace; a ratio of 1 means no selectivity.

Fine-tuning lowers separation, but not specifically for r1r_1: all three directions separate the classes less well in M2M_2 than in M1M_1 (Figure 1). Each direction also loses roughly a quarter of its write in M2M_2 at its own extraction position, and almost none of this loss is due to the weights at l2l_2 (Figure 2). The shortfall therefore originates in earlier layers: injection at l2l_2 restores a signal that upstream layers no longer deliver, rather than reverting a weight edit.

Figure 1. Separation (d′d') of harmful from harmless prompts at layer l2l_2 and the last prompt token, projected onto r1r_1, r2r_2, and DiM, for M1M_1 and M2M_2.

Figure 2. (a) Write along each direction at its extraction position, over harmful prompts. (b) Share of the M1M_1–M2M_2 drop in (a) due to the weight difference W1−W2W_1-W_2 at l2l_2; the remainder comes from differing inputs.

Since r1r_1 and r2r_2 are eroded to a similar degree, erosion does not explain why r1r_1 is the better direction. The gate does (Figure 3). At the last prompt token, from which the first response token is predicted, the gate is strongly selective for r1r_1 but only weakly so for r2r_2, and at token −5-5 it is close to unselective for both; correspondingly, the update along r1r_1 adds substantially more write at the last token than the update along r2r_2. The advantage of r1r_1 thus comes from its alignment with the harmful-specific write pathway at the decisive position, so injecting it acts selectively on harmful inputs, whereas injecting r2r_2 largely does not.

Figure 3. Gate selectivity of the SPECTRA update at layer l2l_2: ratio of the mean ∣r⊤ΠWx∣\lvert r^\top\Pi W x\rvert on harmful to harmless prompts, for attention and MLP.

Experimental Results

Safety Evaluations

General Safety. We evaluate the general safety of models steered via SPECTRA against their base fine-tuned counterparts. Table 2 reports ASR scores (in %)3^{3}. It demonstrates that SPECTRA substantially improves refusal behavior across all evaluated models, while remaining lightweight and requiring no retraining. The improvement is most pronounced for both variants of Gemma-2-9b-it, where SPECTRA yields the largest reductions in ASR. We note that each model and its fine-tuned variants require distinct values of α\alpha and low-rank projection rank ρ\rho; mild hyperparameter tuning was performed to obtain the reported results. We refer the reader to Appendix § 13 and Appendix § 14 for the full parameter choices, and to Appendix § 17.§ 1 for the layers on which steering was applied.

ModelDomainMethodGCGAdvDecodingVanillaGPTFuzzTAPIn-Domain
Llama2MedicalBASE31.018.400.08.822.028.78
SPECTRA6.09.60.00.00.011.56
LawBASE6629.60.015.058.017.0
SPECTRA43.023.20.013.00.06.5
FinanceBASE62.029.00.07.828.08.45
SPECTRA47.029.00.07.80.07.05
Gemma2MedicalBASE36.09.41.068.636.020.0
SPECTRA0.05.60.07.00.05.44
FinanceBASE66.049.221.064.78.071.83
SPECTRA1.016.00.00.010.07.04

Table 2. ASR (%) of base fine-tuned models vs. SPECTRA across models and domains. Lower is better; best results per domain block are bolded. SPECTRA consistently reduces ASR across attack types, with the most pronounced improvements for Gemma2.

Robustness Against Jailbreaks. Table 2 shows that SPECTRA substantially increases each model variant’s robustness against jailbreak attacks. Importantly, this improvement is method-agnostic: SPECTRA safeguards the model against both white-box attacks that require access to model weights, such as GCG (Zou et al., 2023b), and black-box attacks that do not, such as GPTFuzz (Yu et al., 2023) and TAP (Mehrotra et al., 2024).

In-Domain Safety. SPECTRA also yields strong improvements in domain-specific safety, notably without requiring any domain-specific data in its procedure. To the best of our knowledge, SPECTRA is the first training-free method to improve the in-domain safety of fine-tuned language models without any backpropagation, making it uniquely practical for deployment settings where retraining is infeasible.

Coherence Evaluations

General Coherence. We evaluate general coherence via zero-shot performance on HellaSwag, WinoGrande, ARC Challenge, and BoolQ. As shown in Table 4, the impact of SPECTRA on general coherence is statistically negligible, confirming that safety restoration comes at little cost to general capability. In some cases, SPECTRA marginally improves benchmark performance; while this effect is minor, it suggests an interesting direction for future work exploring whether weight-level steering can simultaneously improve both safety and capability. The most notable exception is Gemma-2–9b-it fine-tuned on finance data, which exhibits the largest decrease in general coherence across all tested models; however, this decline is mild and is accompanied by the largest gain in safety performance of any model in our evaluation, representing a favorable overall trade-off.

ModelDomainRTEARCBoolQWinograndeHellaSwag
Llama2Medical68.5 / 67.5(+1.0)45.0 / 45.0(+0.0)78.0 / 78.5(−0.5)69.0 / 68.5(+0.5)60.0 / 58.5(+1.5)
Law69.5 / 69.0(+0.5)44.5 / 44.0(+0.5)78.0 / 80.0(−2.0)65.5 / 68.0(−2.5)57.0 / 57.5(−0.5)
Finance70.0 / 70.5(−0.5)41.5 / 40.5(+1.0)81.0 / 79.5(+1.5)73.5 / 70.0(+3.5)52.0 / 52.0(+0.0)
Gemma2Medical72.5 / 72.0(+0.5)55.5 / 57.5(−2.0)88.0 / 88.0(+0.0)72.5 / 71.0(+1.5)52.5 / 53.0(−0.5)
Finance77.0 / 80.5(−3.5)53.0 / 52.5(+0.5)86.0 / 91.0(−5.0)70.5 / 72.5(−2.0)53.5 / 53.5(+0.0)

Table 4. General coherence benchmark scores across models and domains. For each entry, black denotes the SPECTRA result and gray denotes the base fine-tuned model result, with green indicating improvement and red indicating degradation relative to the base model.

In-Domain Coherence. To further assess the effect of SPECTRA on task utility, we evaluate domain-specific knowledge retention for all models modified by SPECTRA, with results reported in Table 3. We find that these models retain their domain-specific knowledge, with no statistically significant decrease in domain performance. Together with the general coherence results above, this confirms that SPECTRA introduces no significant degradation in either general or domain-specific capability, achieving a favorable safety–utility trade-off without any retraining.

ModelDomainMedicalLawFinance
PubMedQAMedMCQAAbercrombieHearsayFinanceBench
Llama2Medical77.5 / 77.5(+0.0)41.7 / 41.7(+0.0)–––
Law––29.4 / 28.4(+1.0)63.0 / 65.2(−2.0)–
Finance––––70.0 / 66.7(+3.3)
Gemma2Medical95.5 / 95.5(+0.0)52.5 / 52.5(+0.0)–––
Finance––––66.7 / 68.0(−1.3)

Table 3. In-domain performance of base fine-tuned models and SPECTRA across models and domains. Medical domains are evaluated on PubMedQA and MedMCQA, Law on Abercrombie and Hearsay from LegalBench, and Finance on FinanceBench.

False Refusal and Comparison to other methods

A known limitation of refusal steering is its tendency to trigger refusal on unrelated, benign prompts (O’Brien et al., 2025; Lee et al., 2025). To assess whether this limitation applies to SPECTRA, we measure false refusal rates before and after applying SPECTRA across all model variants.

Concretely, we prompt each model with 1000 benign instructions from the Alpaca dataset (Taori et al., 2023) and compute the percentage of responses that constitute false refusals. Model responses are evaluated using Gemini 2.5 Pro [13] as an automated judge, with results reported in Table 5. We find that SPECTRA produces negligible change in false refusal rates across nearly all model variants. The one exception is Gemma-2-9b-it fine-tuned on finance data, which exhibits a small increase of 1.3%, notably the same model that achieves the largest safety gains across all evaluated models, suggesting a mild safety–utility trade-off in this case. Crucially, unlike activation-based steering methods (O’Brien et al., 2025), SPECTRA does not meaningfully increase false refusal rates.

ModelDomainActAddROSICASTSteer2EditSafeMERGESPECTRA
Gemma2Finance91.27.355.818.816.41.3
Medical88.95.765.917.411.20.3
Llama2Finance93.49.561.822.513.70.1
Law97.96.962.421.619.1−0.8
Medical98.58.163.820.414.40.2
Average94.07.561.920.115.00.2

Table 5. Change in false refusal rate (Post − Pre, in %) after steering. Lower is better; best per row in bold.

Furthermore, we compare our methodology to prior methods that utilize steering vectors/ model editing to induce safety in post-trained models. We specifically focus on prior literature that aims to induce refusal directions such as Activation Addition (ActAdd) [4], ROSI (Shairah et al., 2025), CAST (Lee et al., 2025), Steer2Edit (Sun et al., 2026) and a model editing methodology,SafeMERGE, that is aimed to specifically aid fine-tune models safety [15] and present the comparison in Table 5.

Impact of SPECTRA on Multi-Turn Dialogue

Prior sections establish that SPECTRA has minimal impact on both general and in-domain task performance. To probe more subtle, fine-grained effects, we benchmark SPECTRA on MultiChallenge (Sirdeshmukh et al., 2025), a multi-turn conversational evaluation designed to assess capabilities such as inference memory, self-coherence, instruction retention, and reliable editing. As shown in Table 6, the effect of SPECTRA on multi-turn capabilities varies across models, domains, and evaluation dimensions, with no consistent directional trend. We note that the base models themselves achieve relatively low scores on this benchmark, which limits the strength of conclusions that can be drawn from these results alone. Evaluating SPECTRA on larger, more capable models would be necessary to generalize these findings, and we identify this as a promising direction for future work.

DomainModelMethodInference MemorySelf CoherenceInstruction RetentionReliable Editing
FinanceGemma2Base2.159.0929.410.00
SPECTRA4.4210.0013.045.00
Llama2Base4.422.0414.4914.63
SPECTRA5.3110.005.809.76
MedicalGemma2Base8.856.0010.144.88
SPECTRA4.4214.0023.534.88
Llama2Base3.5410.0017.399.76
SPECTRA5.316.1220.590.00
LawLlama2Base7.9618.0020.297.32
SPECTRA5.316.0014.497.32

Table 6. Model performance on MultiChallenge (Sirdeshmukh et al., 2025) across domains.

Limitations

The adversarial landscape surrounding language model safety is rapidly evolving, and it is likely that sufficiently adaptive attacks could be formulated to circumvent the guardrails SPECTRA provides. Generalizing the method to maintain robustness against emerging attack strategies remains an imimportant direction for future work. Additionally, while our evaluations demonstrate that SPECTRA has negligible impact on general coherence and domain-specific knowledge, modern language models possess a wide range of capabilities that were not assessed in this work. It is possible that subtle, task-specific capabilities could be affected by weight-level steering in ways not captured by our benchmarks, and a more comprehensive capability evaluation would be necessary to fully characterize the method’s impact and support broader generalization.

Acknowledgment

This material is based upon work supported by the U.S. National Science Foundation under Grant Nos. IIS-2202699, IIS-2416895, IIS-2301599, CMMI-2301601, DMS-2529302, and OIA-2616279.

Discussion

In this work, we introduced SPECTRA, a training-free method for restoring safety behaviors in fine-tuned language models that suffer from alignment degradation, even when fine-tuned on entirely benign data. Our evaluations demonstrate that SPECTRA substantially improves refusal behavior and robustness against a diverse set of jailbreak attacks across multiple models and domains, with no significant degradation in general coherence, domain-specific knowledge, or false refusal rates. Notably, SPECTRA also improves domain-specific safety without requiring any domain-specific safety data. The effectiveness of SPECTRA stems from a key insight: while full-rank weight perturbation along the refusal direction is effective, it is indiscriminate and disrupts unrelated model capabilities. By restricting the intervention to an activation-aware low-rank subspace in which safety-relevant components are concentrated (Wei et al., 2024), SPECTRA achieves targeted refusal steering with minimal collateral effects. Practically, this requires modifying at most two projection matrices per layer, making the method both simple and computationally lightweight.

Future Work. Several promising directions emerge from this work. Beyond its application to fine-tuning safety restoration, SPECTRA’s weight-level steering framework could generalize to a broader class of behavioral steering tasks, offering a training-free mechanism for eliciting or suppressing model behaviors beyond refusal. The simplicity and low computational cost of SPECTRA also make it well-suited for deployment in production systems, where it could be evaluated against a wider range of adversarial attacks and assessed across a broader set of model capabilities and application domains.

Ethical Considerations We believe mechanistic interpretability techniques can alleviate many AI safety concerns and assist in creating safe and reliable AI systems. However, dual-use remains a concern, as research in mechanistic interpretability can aid malicious intentions for exploiting/creating unsafe AI, potentially leading our methodology to be used for malicious intent. We believe future work in this direction would aid mitigate such downsides.

Related Work

Improving Safety of Fine-Tuned Models: Works have aimed at improving the safety of fine-tuned language models via further preference optimization (Han et al., 2024), curating specific safety data to improve overall safety evaluation by further tuning (Jan et al., 2024), introducing a novel finetuning objective to mitigate fine-tuning-based safety attacks (Qi et al., 2024a), creating safety prefix prompts as backdoors during the fine-tuning procedure (Wang et al., 2024a) or via separating states during the fine-tuning procedure and optimization state drift to prevent alignment degradation (Huang et al., 2024a). All of the aforementioned methods rely on further gradient calculations/tuning or altering the fine-tuning objective or require further domain-specific calibration data, which can pose constraints on the efficacy and cost of the fine-tuning procedure. Our method distinguishes itself by directly injecting a steering vector into a low-rank safety-related space of the fine-tuned variant, making it simple, efficient, and having low spillover onto other capabilities. Erstwhile, ([10] (Jain et al. 2023)) focus on mechanistically interpreting the effects of fine-tuning on know language model circuits.

Refusal Steering Vectors: Prior work has identified the refusal steering vector [4] and utilized its significance to create a white box jailbreak method that incurs low cost and enjoys a high Attack Success Rate. Other works have discovered such steering vectors via the use of sparse autoencoders (O’Brien et al., 2024). Prior research has also utilized steering vectors to mitigate false refusals in language models (Wang et al., 2024b). While some works have applied steering vectors to steer refusal and improve model safety (O’Brien et al., 2025), they have shown to have significant downsides on the general coherence of the language model and/or require inference time interventions (Lee et al., 2025) to steer model activations. Chhabra and Khalili [9] also utilized steering vectors to improve the safety of compressed models.

Compute Statement

For running experiments, we utilized 6 NVIDIA A6000 GPUs with total computations for the experiments covered to be around 300 GPU hours.

values of the reported experiments

The value of α\alpha is model-dependent and was selected via mild hyperparameter tuning for each model variant. Table 7 reports the values used in all experiments.

Model familyDomainAttentionα_{\alpha}MLPα_{\alpha}
LLAMA-2 CHATMedical0.070.05
LLAMA-2 CHATLaw0.10.03
LLAMA-2 CHATFinance0.10.05
GEMMA2-9B-ITMedical0.010.00
GEMMA2-9B-ITFinance0.010.00

Table 7. Values of the steering hyperparameter α\alpha used for each model variant and domain in all experiments.

Hyperparameter sensitivity. we perform a sweep over α\alpha and report GCG attack success rates (Zou et al., 2023b) on the medically fine-tuned LLAMA-2-CHAT (Rohanian et al., 2024) as a function of α\alpha. We consider two cases. In the first, we fix the MLP steering coefficient at α=0.05\alpha= 0.05 and sweep the attention steering coefficient; results are shown in Figure 4. In the second, we fix the attention steering coefficient at α=0.07\alpha= 0.07 and sweep the MLP steering coefficient; results are shown in Figure 5.

ASR of GCG attacks vs α for attention heads

Figure 4. ASR of GCG attacks vs α\alpha for attention heads, with fixed α\alpha for MLPs.

ASR of GCG attacks vs α for MLP

Figure 5. ASR of GCG attacks vs α\alpha for MLP, with fixed α\alpha for attention.

Rank of the Low Matrix for Weight Steering

We report the rank ρ\rho of the low-rank projection matrices used in SPECTRA for each model variant and domain in Table 8.

Model familyDomainAttentionMLP
LLAMA-2 CHATMedical2096/40962096/4096
LLAMA-2 CHATLaw3096/40963096/4096
LLAMA-2 CHATFinance2096/40962096/4096
GEMMA2-9B-ITMedical1048/20482048/3048
GEMMA2-9B-ITFinance1048/20482048/3048

Table 8. Low-rank projection ranks in SPECTRA, reported as selected rank ρ\rho/full rank.

Refusal Direction Selecting Algorithm

The construction in § 2.1 yields a family of candidate directions indexed by layer l∈[L]l \in[L] and token position i∈Ii \in I. To select a single, causally effective refusal direction, we evaluate each candidate through intervention experiments on a held-out validation set. A direction is considered effective if (i) ablating it from activations induced by harmful prompts in Dharmful(val)D_{\mathrm{harmful}}^{(\mathrm{val})} should suppress refusal, and (ii) injecting it into activations induced by harmless prompts in Dharmless(val)D_{\mathrm{harmless}}^{(\mathrm{val})} should elicit refusal. We select the candidate (i∗,l∗)(i^*, l^*) that maximizes this combined causal effect, and denote the resulting direction as r:=ri∗(l∗)\mathbf{r} := \mathbf{r}_{i^*}^{(l^*)}, with normalized version r^\hat{\mathbf{r}}. We borrow the refusal direction selection algorithm from Arditi et al. [5]. Given a collection of difference-in-means vectors, denoted as {ri(l)∣i∈I,l∈[L]}\{\mathbf{r}_i^{(l)} \mid i \in I, l \in[L]\}, we evaluate the following key metrics:

  • bypass_score: Measures the average refusal rate on the validation set of harmful prompts (Dharmful(val))(D_{\mathrm{harmful}}^{(\mathrm{val})}) when applying directional ablation to ri(l)\mathbf{r}_i^{(l)}.

  • induce_score: Assesses the average refusal rate on the validation set of harmless prompts (Dharmless(val)\mathcal{D}_{\mathrm{harmless}}^{(\mathrm{val})}) when the activation addition of ri(l)\mathbf{r}_{i}^{(l)} is applied.

  • kl_score: Computes the average Kullback-Leibler (KL) divergence between the model’s probability distributions at the final token position when evaluated on Dharmless(val)\mathcal{D}_{\mathrm{harmless}}^{(\mathrm{val})} with and without directional ablation of ri(l)\mathbf{r}_{i}^{(l)}.

To identify the optimal direction ri∗(l∗)\mathbf{r}_{i^*}^{(l^*)}, we select the vector with the lowest bypass_score, while ensuring the following constraints are met:

  • induce_score >0> 0

    • Ensures that the selected direction is capable of inducing a refusal response.

  • kl_score <0.1< 0.1

    • Prevents the selection of directions that excessively alter model behavior on benign prompts.

  • l<0.8Ll < 0.8L

    • Restricts the selection to earlier layers, avoiding interference with unembedding representations.

GCG Attack Details

We borrow and modify the methodology of Wei et al. (2024) to generate adversarial suffixes, which is: Run the GCG attack (Zou et al., 2023b) for 500 iterations, with adversarial string initiated as “!!!!!!!!!!!!!!!!!!!!” and a batch size of 256, top-kk as 128, with optimization over the fine-tuned models (including the SPECTRA Variants), with the system prompts removed, for three independent trials. We then identify the top three suffixes with the highest attack success rates on AdvBench, and use them in our evaluation.

Details of Zero-Shot Evaluations

  1. ARC-Challenge: (a) Downstream Task: Science Question Answering.

    (b) Overview: This metric gauges model performance on the ARC-Challenge portion of the AI2 Reasoning Challenge dataset. It comprises grade-school science questions that necessitate complex reasoning and an in-depth understanding of scientific principles4.

  1. HellaSWAG: (a) Downstream Task: Commonsense Reasoning.

    (b) Overview: HellaSWAG is designed to test commonsense reasoning capabilities. It presents a context followed by several multiple-choice endings, with the objective of selecting the most plausible continuation. The dataset challenges models to interpret and reason about everyday situations5.

  1. WinoGrande: (a) Downstream Task: Commonsense Reasoning.

    (b) Overview: WinoGrande is a large-scale dataset for assessing commonsense reasoning. Presented as a fill-in-the-blank task with binary choices, the aim is to select the appropriate option, demanding robust commonsense understanding while mitigating dataset-specific biases6.

  1. BoolQ:

(a) Downstream Task: Yes/No Question Answering.

(b) Overview: BoolQ is a dataset focused on yes/no questions, featuring 15,942 naturally occurring examples. Each instance comprises a question, a passage, and the corresponding answer, with optional contextual information such as the page title. The setup is akin to text-pair classification tasks found in natural language inference research7.

  1. RTE (Recognizing Textual Entailment):

(a) Downstream Task: Textual Entailment.

(b) Overview: The RTE task involves deciding whether a hypothesis can be logically inferred from a given premise. The dataset consists of sentence pairs, where the goal is to classify each pair as either "entailment" (if the hypothesis logically follows from the premise) or "not entailment" (if it does not)8.

Refusal Directional Change After Fine-Tuning

Table Table 9 reports the layer and token position of the refusal direction in both the base and fine-tuned models, showing that domain fine-tuning shifts the source position of the refusal direction across all evaluated variants.

ModelDomainl2/l1l_2/l_1tp2/tp1tp_2/tp_1
Llama2Medical13/14−5/−1-5/-1
Llama2Law12/14−2/−1-2/-1
Llama2Finance12/14−5/−1-5/-1
Gemma2Medical21/31−1/−1-1/-1
Gemma2Finance31/31−5/−1-5/-1

Table 9. Layer and token position of the refusal direction in the base model (l1l_1, tp1tp_1) and fine-tuned model (l2l_2, tp2tp_2) for each variant.

Datasets for Refusal Direction Estimation

Following Arditi et al. [5], we form (Dharmful,Dharmless)(\mathcal{D}_{\mathrm{harmful}}, \mathcal{D}_{\mathrm{harmless}}) for estimating refusal directions. We construct Dharmful\mathcal{D}_{\mathrm{harmful}} from harmful instructions in ADVBENCH (Zou et al. (2023a)), MALICIOUSINSTRUCT (Huang et al. (2024b)), TDC2023 (Mazeika et al. (2024)), and HARMBENCH (Mazeika et al. (2024)), and construct Dharmless\mathcal{D}_{\mathrm{harmless}} from ALPACA (Taori et al. (2023)). Each set contains 160 samples, split into 128 training and 32 validation examples, which we use to compute candidates and select the final normalized refusal direction r^\hat{r}.

References

  1. [1]Cengiz Asmazoğlu Abdullah Bezir, Furkan Burhan Türkaу. 2025. Wiroai/wiroai-finance-gemma-9b. Hugging Face repository.
  2. [2]Reda Alami, Ali Khalifa Almansoori, Ahmed Alzubaidi, Mohamed El Amine Seddik, Mugariya Farooq, and Hakim Hacid. 2024. Alignment with preference optimization is all you need for llm safety. Preprint, arXiv:2409.07772.arxiv.org/abs/2409.07772
  3. [3]DM Anisuzzaman, Jeffrey G Malins, Paul A Friedman, and Zachi I Attia. 2025. Fine-tuning large language models for specialized use cases. Mayo Clinic Proceedings: Digital Health, 3(1):100184.DOI
  4. [4]Andy Arditi and Oscar Obeso. 2023. Refusal mechanisms: initial experiments with Llama-2-7b-chat. Alignment Forum.
  5. [5]Andy Arditi, Oscar Obeso, Aaquib Syed, Daniel Paleka, Nina Panickssery, Wes Gurnee, and Neel Nanda. 2024. Refusal in language models is mediated by a single direction. Preprint, arXiv:2406.11717.arxiv.org/abs/2406.11717
  6. [6]Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, and 1 others. 2022. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862.arxiv.org/abs/2204.05862
  7. [7]Nora Belrose. 2023. Diff-in-means concept editing is worst-case optimal: Explaining a result by Sam Marks and Max Tegmark. https://blog.eleuther.ai/diff-in-means/. Accessed on: May 20, 2024.
  8. [8]Daixuan Cheng, Shaohan Huang, and Furu Wei. 2024. Adapting large language models to domains via reading comprehension. Preprint, arXiv:2309.09530.arxiv.org/abs/2309.09530
  9. [9]Vishnu Kabir Chhabra and Mohammad Mahdi Khalili. 2025. Towards understanding and improving refusal in compressed models via mechanistic interpretability. Preprint, arXiv:2504.04215.arxiv.org/abs/2504.04215
  10. [10]Vishnu Kabir Chhabra, Ding Zhu, and Mohammad Mahdi Khalili. 2024. Neuroplasticity and corruption in model mechanisms: A case study of indirect object identification. In ICML 2024 Workshop on Mechanistic Interpretability.
  11. [11]Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. 2019. BoolQ: Exploring the Surprising Difficulty of Natural Yes/No Questions. In NAACL.arxiv.org/abs/1905.10044
  12. [12]Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. 2018. Think you have solved question answering? Try ARC, the AI2 reasoning challenge. arXiv preprint arXiv:1803.05457.arxiv.org/abs/1803.05457
  13. [13]Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pşţpat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, and 1 others. 2025. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261.arxiv.org/abs/2507.06261
  14. [14]Hoagy Cunningham, Aidan Ewart, Logan Riggs, Robert Huben, and Lee Sharkey. 2023. Sparse autoencoders find highly interpretable features in language models. arXiv preprint arXiv:2309.08600.arxiv.org/abs/2309.08600
  15. [15]Aladin Djuhara, Swanand Ravindra Kadhe, Farhan Ahmed, Syed Zawad, and Holger Boche. 2026. Safe-merge: Preserving safety alignment in fine-tuned large language models via selective layer-wise model merging. In Findings of the Association for Computational Linguistics: ACL 2026, page 35316–35335. Association for Computational Linguistics.
  16. [16]Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. 2024. Kto: Model alignment as prospect theoretic optimization. Preprint, arXiv:2402.01306.arxiv.org/abs/2402.01306
  17. [17]Kathleen C Fraser, Hillary Dawkins, Isar Njieadgholi, and Svetlana Kiritchenko. 2025. Fine-tuning lowers safety and disrupts evaluation consistency. arXiv preprint arXiv:2506.17209.arxiv.org/abs/2506.17209
  18. [18]Leo Gao, Jonathan Tow, Baber Abbasi, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Alain Le Noac’h, Haonan Li, Kyle McDonnell, Niklas Muennighoff, Chris Ociepa, Jason Phang, Laria Reynolds, Hailey Schoelkopf, Aviya Skowron, Lintang Sutawika, and 5 others. 2023. A framework for few-shot language model evaluation.
  19. [19]Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. The Llama 3 herd of models. arXiv preprint arXiv:2407.21783.arxiv.org/abs/2407.21783

Paper details

vlnc:W_xfxymnrp
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents