betaFeedback

WorksComputer Science

TraceRouter++: Causal Semantic Routing Realignment via Optimal Transport

Abstract

Despite their remarkable capabilities, large foundation models (LFMs) remain susceptible to adversarial manipulation. Current safety intervention methods primarily rely on the locality hypothesis, rigidly suppressing isolated neurons or features. However, harmful semantics manifest as distributed, cross-layer circuits, rendering localized ablations both brittle against complex jailbreaks and detrimental to general utility. To bridge this gap, we propose TraceRouter++, a routing-level safety intervention framework that explicitly traces safety-relevant causal routing pathways activated under harmful inputs and realigns their unsafe activation states. Our framework operates in a Discover-Trace-Realign paradigm: it first pinpoints a sensitive onset layer via attention divergence and disentangles dense activations into interpretable routing features using sparse autoencoders (SAEs). Subsequently, it maps these features to downstream causal circuits via feature influence scores (FIS). Crucially, departing from rigid suppression paradigms, TraceRouter++ introduces a constrained optimal transport (OT)-based soft routing realignment. By formulating safety intervention as a tractable coordinate-wise marginal optimal transport approximation, our method redirects unsafe activation marginals toward safe-reference marginals within causally identified routing subspaces while preserving non-selected coordinates at the intervention point. This mechanism safely redirects malicious semantic flows while leaving orthogonal computational routes intact. Extensive experiments across representative large language models (LLMs), multimodal large language models (MLLMs), and diffusion models (DMs) show that TraceRouter++ outperforms evaluated baselines, improving adversarial robustness while maintaining a favorable safety-utility trade-off. WARNING: This paper contains unsafe model responses.

Introduction

The widespread deployment of large foundation models (LFMs), spanning large language models (LLMs) [88], [78], [92], [27], multimodal large language models (MLLMs) [79], [82], [33], [91], and diffusion models (DMs) [18], [49], [11], [66], [65], is accompanied by significant adversarial risks [77], [95], [43]. Existing safety intervention methods [29] primarily rely on the locality hypothesis, attempting to mitigate harmful semantics by suppressing isolated neurons or features. However, semantic representations within LFMs are inherently

distributed [17]. Harmful concepts are rarely confined to isolated components; instead, they propagate through complex, cross-layer topological trajectories [4]. Consequently, localized interventions act as isolated roadblocks, unable to contain semantic leakage under sophisticated adversarial attacks. Furthermore, forcibly severing entangled neural components inevitably disrupts benign computations and degrades general utility. These critical vulnerabilities necessitate a structural paradigm shift from rigid node suppression toward continuous routing interventions. By smoothly realigning malicious semantic propagation, this approach mitigates harmful outputs while limiting distortion to the representation manifold.

Current internal interventions [29], [74], [93] typically impose rigid ablations on specific features or local neurons under the assumption that harmful semantics are spatially localized (see Figure 1(a)). This component-centric perspective, however, overlooks the cross-layer evolution of representations: adversarial semantics exist in superposition across layers and propagate along structured computational circuits. Recent neuroscientific evidence [39], [34] further corroborates that complex behaviors emerge from topological circuits rather than isolated nodes. Thus, ablating individual units often fails to comprehensively regulate the holistic flow of malicious information, leaving models vulnerable to adversarial semantic escape. Furthermore, indiscriminate rigid suppression can damage benign orthogonal pathways, degrading general performance. Effective safety alignment, therefore, suggests a departure from forcibly severing connections, pointing towards a necessity for continuous, optimal transport [53] (OT)-driven soft realignment over these underlying causal pathways.

Illustration of neuron-level and path-level safety intervention

Figure 1. Illustration of neuron-level and path-level safety intervention. Given an illustrative unsafe prompt, (a) neuron-level intervention suppresses isolated unsafe activations, which may disrupt benign neurons while allowing harmful information to recover through alternative propagation routes. In contrast, (b) path-level intervention identifies safety-relevant cross-layer routes that respond to harmful inputs and realigns their unsafe activation states, thereby reducing harmful outputs while limiting interference with benign computations.

Motivated by these mechanistic insights, we formalize the routing-level representation hypothesis: safety-critical semantics in LFMs are encoded and transmitted via distributed causal routing circuits rather than localized units. Crucially, abruptly disconnecting these highly entangled circuits tends to disrupt orthogonal benign computations. Accordingly, as illustrated in Figure 1(b), we posit that formulating safety intervention as a distribution alignment problem, specifically, executing an OT-driven soft realignment over these estimated routing pathways, provides a theoretically grounded and empirically robust approach for enhancing safety control while mitigating unintended utility degradation.

To address this problem, we propose TraceRouter++, a generalized routing-level safety intervention framework operating through a universal Discover–Trace–Realign paradigm. In the discovery stage, we localize a sensitive onset layer, defined as the earliest representational bottleneck where harmful-context-specific attention divergence becomes salient. Subsequently, we employ a Top-K sparse autoencoder (SAE) [15] to decompose dense layer activations into interpretable, sparse routing features, thereby isolating components that exhibit preferential activation under malicious induction relative to benign queries. In the tracing stage, we project these sparse features back into the backbone and utilize zero-out interventions to quantify the feature influence score (FIS) across downstream layers. This causal tracing fundamentally moves beyond localized node analysis, enabling the precise reconstruction of cross-layer computational circuits associated with unsafe behaviors. Finally, in the realignment stage, TraceRouter++ supersedes heuristic rigid suppression with an OT-inspired coordinate-wise soft routing policy. By framing the intervention as a tractable, coordinate-wise transport problem under a conditional independence assumption, TraceRouter++ smoothly maps unsafe activation distributions along identified malicious pathways toward safe reference distributions. This paradigm shift, from rigid neural ablation to continuous distributional redirection, effectively curtails adversarial propagation while better preserving the underlying representation manifold in our evaluations.

We highlight the following contributions of this extended work [67]:

  • We propose TraceRouter++, a routing-level intervention framework evaluated across representative LLM, MLLM, and diffusion architectures. It introduces a systematic “Discover–Trace–Realign” paradigm to explicitly identify and redirect harmful information flows within LFMs.

  • We introduce a novel OT-inspired soft routing mechanism. By formulating safety intervention as a constrained distribution alignment problem, this approach maps unsafe activation distributions toward safer reference distributions, mitigating malicious concepts while alleviating the catastrophic utility degradation often associated with brittle hard-pruning.

  • We seamlessly integrate SAEs [15] with FIS to facilitate fine-grained feature disentanglement, enabling the precise estimation of intervention-sensitive causal propagation paths at the routing level.

  • Extensive evaluations across diverse generative architectures demonstrate that our framework significantly improves adversarial robustness, consistently establishing a highly favorable safety-utility Pareto frontier compared to the evaluated baselines.

Related Work

Safety Intervention in Large Foundation Models

Safety alignment mechanisms for large foundation models (LFMs) span diverse modalities to mitigate adversarial risks. In natural language processing, training-time alignments like RLHF [52] and DPO [57] optimize generation toward human preferences, whereas inference-time defenses such as circuit breaker (CB) [94] and safety-specific neuron ablation [90] explicitly disrupt malicious representations. In the vision domain, diffusion model (DM) interventions range from inference-time guidance (e.g., SLD [62]) to closed-form parameter editing methods like ESD [19], UCE [20], and MACE [46], which mathematically update weight matrices to execute mass concept erasure. Furthermore, to secure multimodal large language models (MLLMs), approaches such as ECSO [26] neutralize visual threats via image-to-text transformation, CoCA [21] recalibrates representations constitutionally, and fine-grained token- or neuron-level methods like SafePTR [10] and SGM [74] aim to intercept malicious logic pathways. Crucially, a fundamental limitation of many of these lightweight editing and fine-grained methods is their implicit reliance on the locality hypothesis, which assumes that harmful concepts are spatially confined. Recent studies demonstrate that such localized erasures are frequently circumvented by sophisticated adversarial prompts [54] and often introduce unintended side effects [61], leaving multimodal systems particularly vulnerable to complex jailbreaks [44]. This over-reliance on semantic locality tends to result in incomplete suppression and semantic leakage. Simultaneously, indiscriminately ablating these components risks degrading benign generation fidelity by rigidly disrupting polysemantic units and orthogonal computational routes.

Causal Routing Analysis of Neural Representations

Emerging research in mechanistic interpretability posits that semantic behaviors in deep neural networks emerge from distributed computations along cross-layer routing paths rather than isolated nodes. This structural perspective is vital for comprehensive model analysis [51]. In Transformer architectures, which underpin LLMs and MLLMs, multiple interacting components dynamically assemble into functional topological circuits. By uncovering influential neuron paths [75], researchers have mapped how specific concepts are encoded in superposition and expressed through structured, often redundant, computational routing paths. This structural perspective extends seamlessly to visual generation. Prior studies on diffusion models [30, 9] demonstrate that textual concepts guide image synthesis through multi-layer attention trajectories and intermediate feature routing paths. These works confirm that modifying attention patterns or feature flows can fundamentally alter semantic expression. Parallel findings in LLMs [75] further indicate that because semantic behaviors are mediated by overlapping cross-layer pathways, harmful concepts can persist or easily re-emerge if interventions are strictly localized. Despite these diagnostic advances in interpretability across diverse architectures, most existing safety interventions have yet to fully integrate these causal semantic routing paths into their defensive mechanisms. This creates a critical gap between routing-level mechanistic understanding and practical, robust safety control. Mechanistic interpretability has enabled the localization of behaviorally relevant layers, sparse features, and cross-layer circuits through attention analysis, sparse representation learning, and activation-based causal tracing [51, 75]. In parallel, representation-steering and concept-erasure methods control model behavior by modifying activation directions, projecting out concept information, or suppressing selected components [93, 7]. Building on these advances, TraceRouter++ connects cross-layer causal route identification with coordinate-wise optimal transport, enabling unsafe activation distributions to be realigned toward safe references within the identified routing subspace.

Highly effective in bridging structural gaps, aligning neural graph representations [72, 85], and uncovering complex structural correspondences between human visual perception and LLMs [83]. By formulating feature alignment as a transportation problem, these approaches facilitate smooth distribution mappings between disparate latent spaces while preserving underlying geometric and topological structures. Despite its tremendous success in macroscopic data alignment, the potential of OT for internal causal routing re-alignment remains largely underexplored. Recently, pioneering works have applied OT to the internal architectures of LLMs: OpT-DeUs [8] fuses Transformer blocks to mitigate permutation mismatches, OTPO [40] employs unbalanced OT for token-level weighting, and multi-level optimal transport (MOT) [64] infers soft, globally consistent layer-to-layer couplings. Conversely, foundation model safety interventions predominantly rely on rigid ablations or orthogonal projections [93, 7], which can inadvertently disrupt benign feature distributions and trigger capability degradation. To address this limitation, our work introduces an OT-based soft routing realignment strategy specifically designed for safety interventions over causal semantic routing paths. By estimating coordinate-wise OT maps between unsafe and safe activations, our method realigns these paths toward safe representational manifolds, enabling targeted concept neutralization while better preserving the model’s general generative utility.

Optimal Transport in Representation Learning

Optimal transport (OT) [53] provides a principled mathematical framework for comparing and mapping probability distributions. In representation learning, OT-based methods leveraging the Wasserstein distance [5] have been extensively applied to domain adaptation [13], generative modeling [5], and cross-modal alignment [2]. For instance, OT has proven

Method

As illustrated in Figure 2, we propose TraceRouter++, a routing level safety intervention framework applicable to diverse large foundation model architectures. We first mathematically formalize safety intervention as a constrained distribution alignment problem over causal semantic routing subspaces (detailed in Section III-A). To resolve this problem, the methodology operates through a systematic pipeline consisting of three stages designed to explicitly identify, trace, and realign the specific computational routing paths associated with harmful semantics. In the discovery stage, TraceRouter++ pinpoints a sensitive onset layer by analyzing attention divergence. It then employs a Top-KK sparse autoencoder to disentangle dense layer activations into interpretable routing features (presented in Section III-B). During the tracing stage, these sparse features are projected back into the latent space to monitor their subsequent propagation. Within this process, the causal influence of each routing node is quantitatively evaluated using the feature influence score, as outlined in Section III-C. Finally, in the realignment stage, TraceRouter++ eschews rigid suppression in favor of a soft routing realignment driven by optimal transport. By estimating marginal transport maps between unsafe and safe activation distributions, the framework seamlessly maps the identified causal semantic routing paths toward safe representational manifolds (discussed in Section III-D). This distributional redirection mitigates the propagation of malicious semantics while strictly preserving orthogonal computational routes. The mathematical intuition and geometric guarantees underpinning this transport mechanism are formally established in Section III-E.

Figure 2. Overview of the TraceRouter++ framework. The pipeline executes three sequential stages: discovery, tracing, and realignment. It first isolates the sensitive onset layer and disentangles dense activations into interpretable features using a Top-KK sparse autoencoder [15]. Subsequently, it reconstructs causal semantic pathways via feature influence scoring. Finally, optimal transport-driven soft realignment maps malicious distributions toward safe reference manifolds, neutralizing harmful semantics while preserving orthogonal generative routes and general utility.

Problem Formulation

We formalize safety intervention as a distribution alignment problem over causal semantic routing subspaces. Let the latent activation vector be denoted by h∈Rdh \in\mathbb{R}^{d}. We define R⊂{1,2,…,d}\mathcal{R} \subset\{1,2,\ldots,d\} as the index set of routing features that are identified as sensitive to interventions and causally linked to harmful semantic propagation.

Let pXp_X and pYp_Y represent the probability distributions of the unsafe activations and the safe reference activations, respectively. Our objective is to discover an optimal transport map T:Rd→RdT:\mathbb{R}^{d}\to\mathbb{R}^{d} that minimizes the transport cost between these distributions while strictly preserving the components outside the selected routing subspace. We formulate the ideal constrained joint optimal transport objective as follows:

min⁡T∫Rdc(x,T(x)) dpX(x)s.t.T#pX=pY,[T(x)]k=xk, ∀k∉R,(1)\begin{aligned} \min_T \quad& \int_{\mathbb{R}^{d}} c(x,T(x))\,\mathrm{d}p_X(x) \\ \text{s.t.}\quad& T_{\#}p_X=p_Y, \\ & [T(x)]_k=x_k,\ \forall k\notin\mathcal{R}, \tag*{(1)} \end{aligned}

where c(x,T(x))c(x,T(x)) is the underlying cost function representing the geometric distance between the original and mapped activations, and T#pXT_{\#}p_X denotes the pushforward measure of pXp_X through the map TT. This formulation ensures that the transport operation is exclusively restricted to the dimensions relevant to the malicious routing pathways, thereby reducing direct interference with non-selected representations.

Sensitive Routing Source Identification

To systematically locate the origin of harmful semantic propagation, we first identify the representational bottleneck that exhibits significant attention divergence. If a specific layer ll encodes harmful semantic routing, a sensitive prompt should induce prominent attention weights directed from sensitive modifiers to target entity nouns. From Algorithm 1, let TsensT_{\mathrm{sens}} and TnT_n denote the sets of sensitive modifier tokens and target entity noun tokens, respectively. We define the sensitive attention SA(P,l)SA(P,l) at layer ll for a prompt PP as the mean attention from TsensT_{\mathrm{sens}} to TnT_n, averaged across all attention heads:

SA(P,l)=∑i=1∣Tsens∣∑j=1∣Tn∣Aˉi,j(l)∣Tsens∣∣Tn∣,(2)SA(P,l)=\frac{\sum_{i=1}^{|T_{\mathrm{sens}}|}\sum_{j=1}^{|T_n|}\bar{A}^{(l)}_{i,j}}{|T_{\mathrm{sens}}||T_n|}, \tag*{(2)}

where Aˉ(l)∈RT×T\bar{A}^{(l)}\in\mathbb{R}^{T\times T} denotes the head-averaged attention matrix at layer ll, ii indexes a sensitive modifier token, and jj indexes a target entity noun token. For Stable Diffusion 1.4 [59], Aˉ(l)\bar{A}^{(l)} is extracted from the frozen CLIP ViT-L/14 [56] text encoder, and ll indexes its 12 transformer blocks rather than the spatial layers of the UNet. Accordingly, the Stable Diffusion 1.4 curve in Figure 3 reports sensitivity across these CLIP text-encoder blocks.

Sensitivity scores across model depth for LLaMA-3-8B-Instruct, LLaVA-1.5-7B, and Stable Diffusion 1.4

Figure 3. Identification of the sensitive onset layer. We localize this critical layer by detecting the first local peak of the sensitivity score SS(l)SS(l) across the model depth. It functions as a representational bottleneck capturing the precise stage where malicious semantics initially diverge from benign representations. Consequently, this focal point serves as the topological origin for subsequent sparse feature disentanglement and optimal transport-driven soft realignment.

However, SA can also increase due to global attention redistribution induced by general prompt variations. To isolate these background effects, we measure the attention changes over unrelated tokens TunrelatedT_{\mathrm{unrelated}} between a sensitive prompt $P_{\mathrm{sens}}

Algorithm 1: Sensitive Routing Source Identification
Input: Sensitive prompts PsensP_{\mathrm{sens}}; benign prompts PbenignP_{\mathrm{benign}}
Input: SAE encoder ESAEE_{\mathrm{SAE}}; integer KK
Output: Sensitive onset layer l∗l^*; sensitive features Msens\mathcal{M}_{\mathrm{sens}}
1 for l=1l=1 to LL do
2   for k=1k=1 to NN do
3      Compute SAk(l)SA_k(l) using (2);
4      Compute CDk(l)CD_k(l) using (3);
5   Compute sensitivity score:
6      SS(l)←1N∑k=1N[SAk(l)−CDk(l)]SS(l)\leftarrow\frac{1}{N}\sum_{k=1}^{N}[SA_k(l)-CD_k(l)] using (4);
7   l∗←l^*\leftarrow index of the first local peak of SS(l)SS(l);
8   Encode activations at layer l∗l^*:
9      Zsens←ESAE(hsens(l∗))Z_{\mathrm{sens}}\leftarrow E_{\mathrm{SAE}}(h_{\mathrm{sens}}^{(l^*)});
10   Zbenign←ESAE(hbenign(l∗))Z_{\mathrm{benign}}\leftarrow E_{\mathrm{SAE}}(h_{\mathrm{benign}}^{(l^*)});
11   for each SAE feature mm do
12      Compute WFSsens(m)\mathrm{WFS}_{\mathrm{sens}}(m) and WFSbenign(m)\mathrm{WFS}_{\mathrm{benign}}(m);
13      ΔWFS(m)←WFSsens(m)−WFSbenign(m)\Delta\mathrm{WFS}(m)\leftarrow\mathrm{WFS}_{\mathrm{sens}}(m)-\mathrm{WFS}_{\mathrm{benign}}(m) using (5);
14   Msens←TopK⁡m(ΔWFS(m),K)\mathcal{M}_{\mathrm{sens}}\leftarrow\operatorname{TopK}_m(\Delta\mathrm{WFS}(m),K);
15   return l∗,Msensl^*,\mathcal{M}_{\mathrm{sens}}

Table 1.

and its benign counterpart PbenignP_{\mathrm{benign}}. The contextual disturbance CD at layer ll is defined as:

CD(l)=1∣Tunrelated∣∑t=1∣Tunrelated∣∥A^Psens(l)(t)−A^Pbenign(l)(t)∥1,(3)\mathrm{CD}(l)=\frac{1}{\lvert T_{\mathrm{unrelated}}\rvert}\sum_{t=1}^{\lvert T_{\mathrm{unrelated}}\rvert}\left\lVert\hat{A}^{(l)}_{P_{\mathrm{sens}}}(t)-\hat{A}^{(l)}_{P_{\mathrm{benign}}}(t)\right\rVert_1, \tag*{(3)}

where A^(l)\hat{A}^{(l)} is the row-normalized attention matrix at layer ll. We then define the sensitivity score SS at layer ll as the average difference between SA and CD:

SS(l)=1N∑k=1N[SAk(Psens,l)−CDk(l)],(4)\mathrm{SS}(l)=\frac{1}{N}\sum_{k=1}^{N}\left[\mathrm{SA}_k(P_{\mathrm{sens}},l)-\mathrm{CD}_k(l)\right], \tag*{(4)}

where NN denotes the number of paired sensitive and benign prompts, and kk indexes the prompt pair. By monitoring SS(l)\mathrm{SS}(l) across all layers, we pinpoint the sensitive onset layer as the earliest layer where the score exhibits its first prominent local peak above the background level. As illustrated in Figure 3, SS(l)\mathrm{SS}(l) reaches this initial peak at a specific block, which we designate as the routing source layer for subsequent feature disentanglement and causal tracing.

After locating the sensitive onset layer, we employ an SAE [15] to decompose its dense activations into sparse and interpretable routing features. We evaluate each SAE [15] neuron mm using a weighted frequency score (WFS), defined as WFS(m)=f(m)⋅μ(m)\mathrm{WFS}(m)=f(m)\cdot\mu(m), which combines its activation frequency ff and mean activation magnitude μ\mu. To isolate the specific routing features causally linked to harmful concepts, we rank the SAE [15] features based on their differential activations between the sensitive and benign prompts:

ΔWFS(m)=WFSsens(m)−WFSbenign(m),(5)\Delta\mathrm{WFS}(m)=\mathrm{WFS}_{\mathrm{sens}}(m)-\mathrm{WFS}_{\mathrm{benign}}(m), \tag*{(5)}

where WFSsens(m)\mathrm{WFS}_{\mathrm{sens}}(m) and WFSbenign(m)\mathrm{WFS}_{\mathrm{benign}}(m) denote the weighted frequency scores of neuron mm computed over the sensitive and benign sample sets, respectively. The Top-KK SAE [15] features ranked by ΔWFS\Delta\mathrm{WFS} are selected as the sensitive routing features. This differential evaluation ensures that the extracted routing features capture signals specific to malicious intents rather than generic linguistic patterns.

Causal Semantic Routing Tracing

To explicitly link the interpretable routing features with the internal dense computation of the model, we project the identified sensitive SAE [15] features Msens\mathcal{M}_{\mathrm{sens}} back into the original dense activation space at the sensitive onset layer l∗l^*. From Algorithm 2, by utilizing the decoder weight matrix Wdec\mathbf{W}_{\mathrm{dec}}, we compute the projected activation Zproj(l∗)=Wdec⋅Msens\mathbf{Z}^{(l^*)}_{\mathrm{proj}}=\mathbf{W}_{\mathrm{dec}}\cdot\mathcal{M}_{\mathrm{sens}}. Dense activation coordinates in layer l∗l^* with the largest absolute projected magnitudes are selected as the source routing nodes, denoted as Ssrc\mathcal{S}_{\mathrm{src}}. To characterize the subsequent propagation of harmful semantics, we evaluate how these source routing nodes causally influence downstream layers l>l∗l>l^* through a selective zero ablation intervention. Although rigorous causal mediation analysis typically requires computationally intensive activation resampling to strictly prevent out-of-distribution representation shifts, targeted zero ablation serves as a highly tractable and empirically robust approximation for mapping

Algorithm 2: Causal Semantic Routing Tracing

Input: Sensitive features Msens\mathcal{M}_{\mathrm{sens}}; onset layer l∗l^* Input: SAE decoder weight matrix Wdec\mathbf{W}_{\mathrm{dec}}; integer KK Output: Causal routing path P\mathcal{P}; routing mask MM

  1. Project sensitive features back to dense space and identify the source routing nodes Ssrc\mathcal{S}_{\mathrm{src}} at layer l∗l^*;

  2. Initialize path P←Ssrc\mathcal{P}\leftarrow\mathcal{S}_{\mathrm{src}};

  3. for each input sample xx do

  4.   Run one baseline forward pass and cache ax(l)a_x^{(l)} for all l>l∗l>l^*;

  5.   Apply selective zero ablation once at the source layer: h^x,r(l∗)←0, ∀r∈Ssrc\hat{h}_{x,r}^{(l^*)}\leftarrow0,\ \forall r\in\mathcal{S}_{\mathrm{src}};

  6.   Run one ablated forward pass and cache a^x(l)\hat{a}_x^{(l)} for all l>l∗l>l^*;

  7. for l=l∗+1l=l^*+1 to LL do

  8.   for each dense neuron mm in layer ll do

  9.     Compute Δablation(m)\Delta_{\mathrm{ablation}}(m) from cached a(l)a^{(l)} and a^(l)\hat{a}^{(l)};

  10.     Compute fdense(m)f_{\mathrm{dense}}(m) and μdense(m)\mu_{\mathrm{dense}}(m);

  11.     FIS(m)←fdense(m)⋅μdense(m)⋅Δablation(m)\mathrm{FIS}(m)\leftarrow f_{\mathrm{dense}}(m)\cdot\mu_{\mathrm{dense}}(m)\cdot\Delta_{\mathrm{ablation}}(m);

  12.     Sl←TopK⁡m(FIS(m),K)\mathcal{S}_l\leftarrow\operatorname{TopK}_m(\mathrm{FIS}(m),K);

  13.     P←P∪Sl\mathcal{P}\leftarrow\mathcal{P}\cup\mathcal{S}_l;

  14. Construct binary routing mask MM from path P\mathcal{P};

  15. return P,M\mathcal{P},M

causal dependencies across massive parameter spaces. Specifically, this intervention is applied directly to the backbone activation coordinates: for every sample in the offline tracing set, we set the activation values of the identified source routing nodes to exactly zero (i.e., Zi(l∗)←0Z_i^{(l^*)}\leftarrow0 for all i∈Ssrci\in\mathcal{S}_{\mathrm{src}}), while maintaining all orthogonal neurons unaltered.

We then quantify the causal impact on each downstream routing node mm using the feature influence score (FIS):

FIS(m)=fdense(m)⋅μdense(m)⋅Δablation(m).(6)\mathrm{FIS}(m)=f_{\mathrm{dense}}(m)\cdot\mu_{\mathrm{dense}}(m)\cdot\Delta_{\mathrm{ablation}}(m). \tag*{(6)}

For downstream dense neurons, we define the activation shift Δablation(m)=E[∣am−a^m∣]\Delta_{\mathrm{ablation}}(m)=\mathbb{E}[\lvert a_m-\hat{a}_m\rvert] as the expected absolute difference between the original activation (ama_m) and the post intervention activation (a^m\hat{a}_m). Furthermore, we adapt the statistical metrics for continuous signals: fdense(m)=P(am>0)f_{\mathrm{dense}}(m)=P(a_m>0) denotes the activation frequency, and μdense(m)=E[am∣am>0]\mu_{\mathrm{dense}}(m)=\mathbb{E}[a_m\mid a_m>0] represents the conditional mean magnitude.

By selecting routing nodes that exhibit high FIS values under malicious induction, we can reliably reconstruct the principal causal semantic routing pathways responsible for harmful content generation.

Causal Routing Realignment

To neutralize malicious semantic propagation without brittle structural suppression, TraceRouter++ introduces optimal transport (OT)-driven soft routing realignment. Rather than enforcing rigid activation removal, our framework maps harmful activation distributions toward safe reference manifolds strictly

Algorithm 3: Causal Routing Realignment

Input: Activation vector h∈Rd\mathbf{h} \in\mathbb{R}^{d}; routing mask M∈{0,1}d\mathbf{M} \in\{0,1\}^{d}

Input: Unsafe activation samples X={xi}i=1n\mathcal{X} = \{\mathbf{x}_{i}\}_{i=1}^{n}; safe reference activation samples Y={yj}j=1m\mathcal{Y} = \{\mathbf{y}_{j}\}_{j=1}^{m}

Output: Realigned activation h~\tilde{\mathbf{h}}

  1. Initialize marginal transport map T={Tk}k=1dT = \{T_{k}\}_{k=1}^{d};

  2. for k=1k = 1 to dd do

  3.   if Mk=1M_{k} = 1 then

  4.     Extract unsafe activation values X(k)={xi(k)}i=1n\mathcal{X}^{(k)} = \{x_{i}^{(k)}\}_{i=1}^{n};

  5.     Extract safe reference activation values Y(k)={yj(k)}j=1m\mathcal{Y}^{(k)} = \{y_{j}^{(k)}\}_{j=1}^{m};

  6.     Estimate empirical CDF FXkF_{X}^{k} from X(k)\mathcal{X}^{(k)};

  7.     Estimate empirical CDF FYkF_{Y}^{k} from Y(k)\mathcal{Y}^{(k)};

  8.     Define marginal transport map:

  9.     Tk(v)←(FYk)−1(FXk(v))T_{k}(v) \leftarrow(F_{Y}^{k})^{-1}(F_{X}^{k}(v)) using Eq. (7);

  10.   else

  11.     Set identity map Tk(v)←vT_{k}(v) \leftarrow v;

  12. Apply transport to each coordinate independently:

  13. [T(h)]k←Tk(hk), ∀k∈{1,…,d}[T(\mathbf{h})]_{k} \leftarrow T_{k}(h_{k}), \ \forall k \in\{1,\ldots,d\};

  14. Apply constrained routing realignment:

  15. h~←h+M⊙(T(h)−h)\tilde{\mathbf{h}} \leftarrow\mathbf{h} + \mathbf{M} \odot(T(\mathbf{h}) - \mathbf{h}) using Eq. (8);

  16. return h~\tilde{\mathbf{h}};

along the isolated causal pathways. For each identified routing node on a causal semantic routing path, we collect its activation distributions under unsafe contexts and safe reference contexts. Specifically, let X={xi}i=1n\mathcal{X} = \{\mathbf{x}_{i}\}_{i=1}^{n} denote the activation samples observed under unsafe inputs, and let Y={yj}j=1m\mathcal{Y} = \{\mathbf{y}_{j}\}_{j=1}^{m} denote those observed under safe reference inputs. Let F^Xk\widehat{F}_{X_{k}} and F^Yk\widehat{F}_{Y_{k}} denote the empirical cumulative distribution functions for coordinate kk. Because solving unrestricted multivariate OT over the selected activation subspace is computationally expensive, we implement coordinate-wise marginal OT as a tractable approximation. This procedure aligns the unsafe and safe-reference marginal distributions independently along each selected coordinate; it does not assume that the learned routing features are independent in practice. Exact equivalence to joint OT holds only under the factorization and separable-cost conditions stated in Proposition 1. For a scalar activation value vv in a selected coordinate kk, we define the empirical quantile map as

Tk(v)=F^Yk−1(F^Xk(v)).(7)T_{k}(v) = \widehat{F}_{Y_{k}}^{-1}\left(\widehat{F}_{X_{k}}(v)\right). \tag*{(7)}

(8) defines an empirical quantile-based mapping motivated by one-dimensional optimal transport. Algorithm 3 applies this mapping independently to each selected activation coordinate and uses the identity map outside the selected routing subspace. Under the factorization and separable-cost conditions in Proposition 1, the product of exact scalar optimal transport maps, when they exist, is an optimal joint transport map. For general activation distributions, however, the empirical coordinate-wise implementation need not recover the safe-reference joint distribution. Algorithm 3 is therefore treated as

ItemLLMsMLLMsDMs
ModelsLLaMA3-8B-Instruct [27]<br>Mistral-7B-Instruct [35]LLaVA-1.5-7B [42]<br>MiniGPT4-7B [91]<br>DeepSeek-VL2 [80]Stable Diffusion 1.4 [59]<br>FLUX.1-Dev [38]<br>AltDiffusion [86]<br>Show-o2 [84]
DatasetsAdvBench [95], HarmBench [50]<br>GCG [95], AutoDAN [43]<br>Template [81], Prefill [81]<br>XSTEST [60]<br>Global-MMLU-Lite [68]FigStep [25]<br>JailbreakV-28K [47]<br>MMBench [45]I2P-N [62], I2P-V [62]<br>P4D [12], Ring-A-Bell [73]<br>MS COCO [41]
MetricsDSR [22]<br>ORR [14]<br>Accuracy [69]DSR [22]<br>Accuracy [69]DSR [22]<br>CLIP Score [31]<br>FID [32]

Table 1 (PDF p. 6). Overview of models, datasets, and metrics across foundational architectures. Underlines indicate utility or over-refusal evaluations.

To integrate this transport mechanism with the identified causal routing structure, we define a binary routing mask M∈{0,1}dM \in\{0,1\}^{d}, so that only the selected activation dimensions are modified. The activation vector h∈Rd\mathbf{h} \in\mathbb{R}^{d} is decomposed into selected routing components and unselected components. The realigned activation h~\tilde{\mathbf{h}} is subsequently computed as

h~=h+M⊙(T(h)−h),(8)\tilde{\mathbf{h}} = \mathbf{h} + M \odot(T(\mathbf{h}) - \mathbf{h}), \tag*{(8)}

where T(h)T(\mathbf{h}) is applied to each dimension independently, and ⊙\odot denotes the Hadamard product. Because the routing mask restricts the transport operation to the selected coordinates, the identity map is retained outside the identified routing subspace. Within the selected coordinates, empirical quantile mappings are used to align the marginal distributions, with the remaining marginal and joint discrepancies evaluated empirically. The complete realignment procedure is summarized in Algorithm 3.

Theoretical Insight

We provide the formal mathematical intuition underpinning our optimal transport framework.

Proposition 1 (Marginal Decomposition under Coordinate-Wise Factorization). Let PXRP_{X}^{\mathcal{R}} and PYRP_{Y}^{\mathcal{R}} denote the unsafe and safe-reference activation distributions restricted to the selected routing subspace R\mathcal{R}, both with finite second moments. If both distributions factorize as

PXR=⨂k∈RPXk,PYR=⨂k∈RPYk,(9)P_{X}^{\mathcal{R}} = \bigotimes_{k \in\mathcal{R}} P_{X}^{k}, \qquad P_{Y}^{\mathcal{R}} = \bigotimes_{k \in\mathcal{R}} P_{Y}^{k}, \tag*{(9)}

and the transport cost is separable,

c(x,y)=∑k∈R∣xk−yk∣2,(10)c(\mathbf{x},\mathbf{y}) = \sum_{k \in\mathcal{R}} \lvert x_{k} - y_{k} \rvert^{2}, \tag*{(10)}

then

W22(PXR,PYR)=∑k∈RW22(PXk,PYk).(11)W_{2}^{2}(P_{X}^{\mathcal{R}}, P_{Y}^{\mathcal{R}}) = \sum_{k \in\mathcal{R}} W_{2}^{2}(P_{X}^{k}, P_{Y}^{k}). \tag*{(11)}

If scalar optimal transport maps exist, their product is an optimal joint transport map on R\mathcal{R}.

Proposition 1 establishes the decomposition under the stated coordinate-wise factorization and separable-cost assumptions. Sparse autoencoder features are not assumed to satisfy the factorization condition exactly. Algorithm 3 therefore implements empirical coordinate-wise quantile mappings as a tractable marginal-OT approximation, without a general guarantee of exact joint alignment.

Proposition 2 (Minimal Perturbation under Exact Joint Transport). Let TRT_R be an optimal transport map satisfying the exact joint pushforward constraint

(TR)#PXR=PYR,(12)(T_R)\#P_X^R = P_Y^R, \tag*{(12)}

under the squared Euclidean cost. Then TRT_R minimizes

EX∼PXR[∥X−TR(X)∥22],(13)\mathbb{E}_{X\sim P_X^R}\left[\left\lVert X-T_R(X)\right\rVert_2^2\right], \tag*{(13)}

among all deterministic maps satisfying this constraint.

The minimum-displacement result applies to the exact joint transport problem. When the selected activation coordinates are dependent, the coordinate-wise map in Algorithm 3 need not satisfy the joint pushforward constraint; consequently, Proposition 2 does not establish global optimality for the unrestricted coordinate-wise implementation. The routing mask guarantees only that non-selected coordinates remain unchanged at the intervention point. Because subsequent layers may mix selected and non-selected coordinates, this property is not an end-to-end utility-preservation theorem. Utility preservation is therefore evaluated empirically.

Empirical Approximation Error of Coordinate-Wise OT. To quantify the approximation introduced by coordinate-wise OT, we compare it with joint OT and measure the remaining distributional residuals over five seeds. The joint-OT versus coordinate-wise-OT decomposition gap is 30.75±\pm1.15%. After mapping, the joint-distribution residual is 33.05 ±\pm 1.83%, whereas the marginal-distribution residual is 0.111 ±\pm 0.019%. These results indicate that, under the evaluated settings, the empirical coordinate-wise quantile mappings closely align the target marginal distributions but retain a non-negligible mismatch in their joint dependency structure.

Experiments and Analysis

Implementation Details

Models. To verify the architectural universality of TraceRouter++, we conduct evaluations across a diverse suite of large foundation models, as summarized in Table I. Our selection encompasses representative large language models (LLMs) such as LLaMA3-8B-Instruct [27] and Mistral-7B-Instruct [35], alongside multimodal large language models (MLLMs) like LLaVA-1.5-7B [42], MiniGPT-4-7B [91] and DeepSeek-VL2 [80]. For visual generation, we extend our analysis to various diffusion models (DMs), including Stable Diffusion 1.4 [59], FLUX.1-Dev [38], AltDiffusion [86], and Show-o2 [84]. This broad selection ensures that our routing realignment framework is effective across different modalities and computational structures.

Datasets. Our evaluation involves an adversarial environment paired with standard capability benchmarks to measure the trade-off between safety and utility across LLMs, MLLMs, and DMs. For LLM, we construct harmful instruction sets from AdvBench [95] and HarmBench [50], and further evaluate robustness against both automated gradient-driven attacks, including GCG [95] and AutoDAN [43], and manual pattern-based jailbreaks, including Template [81] and Prefill [81] attacks.

For LLMs, we benchmark the defensive capabilities of TraceRouter++ on LLaMA3-8B-Instruct [27] and Mistral-7B-Instruct [35] against state of the art interventions including HumorReject [81] and Circuit Breaker [94]. As detailed in Table II, our evaluation covers gradient driven attacks (GCG [95] and AutoDAN [43]) alongside manual pattern injections (Template [81], Prefill [81], and their combinatorial variants). TraceRouter++ demonstrates consistent robustness across all evaluated attack categories. Specifically, the foundational Mistral-7B-Instruct [35] model exhibits a baseline

ModelMethodGradient DrivenPattern InjectionsAvg. ↑
GCG [95] ↑AutoDAN [43] ↑Template ↑Prefill ↑T+P ↑
LLaMA3-8B-Instruct [27]Original [27]88879841263.2
DeepAug [55]994010059360.2 (-3.0)
CB [94]999897959897.4 (+34.2)
DeRTa [87]9789100983283.2 (+20.0)
HumorReject [81]9899991009898.8 (+35.6)
LEACE [7]87999957369.0 (+5.8)
TraceRouter [67]98.6±0.898.6\pm0.899.4±0.699.4\pm0.699.2±0.799.2\pm0.799.6±0.599.6\pm0.598.8±1.298.8\pm1.299.2±0.699.2\pm0.6 (+36.0)
TraceRouter++99.0±0.799.0\pm0.799.8±0.499.8\pm0.499.6±0.599.6\pm0.599.8±0.499.8\pm0.499.4±0.599.4\pm0.599.5±0.399.5\pm0.3 (+36.3)
Mistral-7B-Instruct [35]Original [35]4222146.6
DeepAug [55]6619856731.2 (+24.6)
CB [94]898689999090.6 (+84.0)
DeRTa [87]615054925362.0 (+55.4)
HumorReject [81]959796989796.6 (+90.0)
LEACE [7]267232321.2 (+14.6)
TraceRouter [67]97.6±1.097.6\pm1.098.2±1.098.2\pm1.097.4±1.197.4\pm1.198.2±1.198.2\pm1.197.8±1.197.8\pm1.197.8±0.997.8\pm0.9 (+91.2)
TraceRouter++99.4±0.599.4\pm0.599.0±0.799.0\pm0.798.8±0.498.8\pm0.498.6±0.598.6\pm0.598.4±0.598.4\pm0.598.8±0.398.8\pm0.3 (+92.2)

Table II. Quantitative evaluation of safety interventions on LLMs. We benchmark TraceRouter++ against diverse adversarial jailbreaks, spanning both gradient-driven optimizations and structural pattern injections. Values denote the defense success rate (DSR) in percentages, with optimal outcomes bolded.

attacks. To evaluate benign behavior and utility preservation, we use XSTEST [60] for over-refusal analysis and Global-MMLU-Lite [68] for general reasoning capability. For MLLMs, we evaluate safety and jailbreak robustness using FigStep [25], which emphasizes image-side jailbreak attacks, and JailbreakV-28K [47], which covers transferred text-side jailbreak scenarios under visual inputs. General multimodal utility is measured on MMBench [45]. For DMs, we assess explicit and violent content mitigation using the I2P-N and I2P-V subsets of I2P [62], and evaluate adversarial robustness on P4D [12] and Ring-A-Bell [73]. Meanwhile, generation fidelity is monitored on MS COCO [41].

Metrics. We adopt the defense success rate (DSR) as the primary metric to quantify the proportion of adversarial attempts successfully mitigated by our framework. To ensure objective and reliable semantic judgment, we employ Claude 3.5 Sonnet [3] and GPT-4 [1] as automated evaluators, supplemented by manual human verification for ambiguous edge cases. For utility preservation, we report the over-refusal rate (ORR) [14] and general accuracy to locate the safety and utility Pareto frontier. In the visual domain, specialized tools such as NudeNet [6] are utilized for content detection, while Fréchet Inception Distance (FID) [32] measures the proximity of the intervened generation distribution to the original benign distribution. We additionally report CLIP Score (CS) [31] to measure text-image alignment in benign generation.

Hyperparameters. To achieve precise causal feature disentanglement, we explicitly train a Top-KK sparse autoencoder (SAE) [15] on the intermediate layer representations. This architecture is configured with an expansion factor of four and optimized via the Adam algorithm [36] using a learning rate of 4×10−44 \times10^{-4} and a batch size of 4096, guided by a mean squared error reconstruction loss. The exact number of selected sensitive routing features KK is determined by identifying the elbow point of the weighted frequency score (WFS) [29] ranking curve. This specific mathematical threshold marks the transition where high-response hub features enter the long-tail feature distribution. To characterize experimental variability, we repeat the main TraceRouter++ evaluations in Table II–Table VII and the ablation comparisons in Table VIII–Table XII using five seeds. These entries are reported as mean ±\pm sample standard deviation; other results retain their original reporting format. Finally, all optimization and inference procedures are executed on a single NVIDIA A6000 GPU to ensure rigorous and fair computational comparisons.

ModelMethodORR % ↓
LLaMA3-8B-Instruct [27]Original [27]3.2
DeepAug [55]39.6 (+36.4)
CB [94]4.4 (+1.2)
DeRTa [87]27.6 (+24.4)
HumorReject [81]5.2 (+2.0)
LEACE [7]3.6 (+0.4)
TraceRouter [67]3.6±\pm0.6 (+0.4)
TraceRouter++3.5±\pm0.3 (+0.3)
Mistral-7B-Instruct [35]Original [35]1.2
DeepAug [55]62.0 (+60.8)
CB [94]3.6 (+2.4)
DeRTa [87]74.4 (+73.2)
HumorReject [81]6.0 (+4.8)
LEACE [7]3.2 (+2.0)
TraceRouter [67]2.8±\pm0.9 (+1.6)
TraceRouter++2.3±\pm0.3 (+1.1)

Table VII. Over-refusal analysis on XSTEST [60]. Values denote over-refusal rate (ORR) [14] percentages across LLMs. Lower scores indicate superior utility preservation.

ModelOriginalAmplificationSuppression
LLaMA3-8B-Instruct [27]63.275.8±\pm1.7 (+12.6)61.4±\pm1.4 (-1.8)
LLaVA-1.5-7B [42]49.078.6±\pm0.7 (+29.6)18.2±\pm1.3 (-30.8)
Stable Diffusion 1.4 [59]82.273.5±\pm1.3 (-8.7)98.8±\pm0.4 (+16.6)

Table VIII. Causal validation via routing manipulation. Values represent defense success rates under targeted pathway amplification and suppression. Safety gains stem from refusal amplification in language and multimodal architectures, while generative suppression provides optimal mitigation for diffusion models. Optimal results are bolded.

MethodLLaMA3-8B-Instruct [27]LLaVA-1.5-7B [42]Stable Diffusion 1.4 [59]
DSR ↑Global-MMLU-Lite [68] ↑DSR ↑MMBench [45] ↑DSR ↑CLIP Score [31] ↑FID [32] ↓
Original63.271.549.047.582.231.34N/A
F⋅I⋅SF \cdot I \cdot S99.5±0.3 (+36.3)71.1±0.4 (-0.4)99.4±0.1 (+50.4)47.4±0.2 (-0.1)99.3±0.2 (+17.1)31.08±0.07 (-0.26)8.62±0.04
F⋅IF \cdot I98.4±0.4 (+35.2)71.4±0.4 (-0.1)87.0±0.7 (+38.0)47.1±0.1 (-0.4)97.6±0.4 (+15.4)31.22±0.04 (-0.12)9.34±0.15
F⋅SF \cdot S95.6±0.3 (+32.4)70.9±0.4 (-0.6)86.8±0.8 (+37.8)46.9±0.2 (-0.6)98.3±0.3 (+16.1)31.25±0.06 (-0.09)8.79±0.11
I⋅SI \cdot S93.2±0.7 (+30.0)70.9±0.6 (-0.6)92.4±0.6 (+43.4)45.9±0.1 (-1.6)96.3±0.5 (+14.1)31.20±0.07 (-0.14)9.37±0.18
FF98.6±0.4 (+35.4)71.1±0.3 (-0.4)63.4±0.8 (+14.4)45.1±0.1 (-2.4)97.7±0.4 (+15.5)31.19±0.05 (-0.15)9.74±0.14
II98.4±0.5 (+35.2)71.2±0.3 (-0.3)86.2±0.7 (+37.2)44.8±0.1 (-2.7)97.5±0.4 (+15.3)31.23±0.08 (-0.11)9.02±0.17
SS97.2±0.5 (+34.0)70.9±0.5 (-0.6)66.8±0.6 (+17.8)45.1±0.1 (-2.4)97.6±0.3 (+15.4)31.24±0.06 (-0.10)8.94±0.13

Table XII. Quantitative ablation of the sensitive path extraction metric. We compare our full routing discovery formulation against reduced variants utilizing isolated activation frequency (FF), activation intensity (II), and zero-out causal shift (SS). The empirical results confirm that the joint consideration of statistical and causal evidence strictly maximizes the defense success rate while preserving foundational utility across language, multimodal, and diffusion architectures. Missing baseline values are denoted as N/A. Best results are highlighted in bold.

Quantitative Comparison with State of the Art Methods

LLMs. We benchmark the defensive capabilities of TraceRouter++ on LLaMA3-8B-Instruct [27] and Mistral-7B-Instruct [35] against state of the art interventions including HumorReject [81] and Circuit Breaker [94]. As detailed in Table II, our evaluation covers gradient driven attacks (GCG [95] and AutoDAN [43]) alongside manual pattern injections (Template [81], Prefill [81], and their combinatorial variants). TraceRouter++ demonstrates consistent robustness across all evaluated attack categories. Specifically, the foundational Mistral-7B-Instruct [35] model exhibits a baseline average defense success rate of 6.6%. TraceRouter++ increases this average to 98.8±0.3%, corresponding to an improvement of 92.2 percentage points. Furthermore, on the LLaMA3-8B-Instruct [27] architecture, our framework attains an average defense success rate of 99.5±0.3%, an improvement of 36.3 percentage points over the original model. In comparison, LEACE [7] achieves lower average defense rates of 69.0% on LLaMA3 and 21.2% on Mistral. Its mean defense success rates against AutoDAN, Template, and combined Template and Prefill attacks are 99.8±0.4%, 99.6±0.5%, and 99.4±0.5%, respectively. Across individual attack categories on both models, standard deviations do not exceed 0.7 percentage points. These results demonstrate high defense success rates across the evaluated attacks, with limited observed variability across the five seeds.

MLLMs. We extend our evaluation to MLLMs, where safety risks manifest through complex cross-modal interactions. We first assess robustness against text-guided attacks using the JailbreakV-28K benchmark [47]. As reported in Table III, TraceRouter++ elevates the baseline average defense success rate on LLaVA-1.5-7B [42] from 48.3% to 98.8±0.1%. On MiniGPT-4-7B [91], it achieves 96.7±0.4%, exceeding the reported defense rates of SafePTR [10] (87.4%) and Immune [23] (81.7%). Subsequently, we evaluate defense capabilities against visual guided threats via the FigStep benchmark [25]. Table IV details that TraceRouter++ yields an average defense success rate of 99.4±0.1% on LLaVA-1.5-7B [42], compared with the reported rates of Immune [23] (95.8%) and SafePTR [10] (98.4%). A parallel trend emerges on MiniGPT-4-7B [91] with an average DSR of 99.6±0.1%. On DeepSeek-VL2, an average DSR of 96.0±0.5%, compared with 45.6% for the original model and 90.2% for SafePTR [10]. Collectively, these results demonstrate high defense success rates against both text-guided and visual-guided attacks, with limited observed variability in the aggregate scores across five seeds.

ModelMethodNoise T↑Noise P↑Noise L↑SD T↑SD P↑SD L↑Nature T↑Nature P↑Nature L↑Blank T↑Blank P↑Blank L↑Avg. ↑
LLaVA-1.5-7B [42]Original [42]42.970.837.939.560.927.141.068.240.642.669.139.248.3
FigStep [25]40.547.759.542.945.150.041.641.655.539.248.959.547.7 (-0.6)
CoCA [21]38.860.937.938.758.847.336.964.844.639.062.747.348.7 (+0.4)
ECSO [26]42.774.641.942.774.641.942.774.641.942.6874.641.953.1 (+4.8)
AdaShield [76]78.498.682.575.498.677.176.899.282.578.298.682.585.7 (+37.4)
Immune [23]90.8100.0100.091.9100.0100.098.6100.0100.094.7100.0100.097.9 (+49.6)
SafePTR [10]96.5100.0100.098.4100.0100.094.9100.0100.094.7100.0100.098.7 (+50.4)
TraceRouter [67]96.0±1.596.0\pm1.599.1±0.999.1\pm0.998.6±1.498.6\pm1.497.5±1.297.5\pm1.299.4±0.799.4\pm0.797.3±1.997.3\pm1.994.5±1.594.5\pm1.598.8±1.398.8\pm1.397.3±1.997.3\pm1.994.2±1.394.2\pm1.399.1±0.999.1\pm0.998.6±1.498.6\pm1.497.5 (+49.2)
TraceRouter++96.8±0.396.8\pm0.3100.0±0.0100.0\pm0.0100.0±0.0100.0\pm0.098.5±0.298.5\pm0.2100.0±0.0100.0\pm0.0100.0±0.0100.0\pm0.095.5±0.395.5\pm0.3100.0±0.0100.0\pm0.0100.0±0.0100.0\pm0.095.3±0.495.3\pm0.4100.0±0.0100.0\pm0.0100.0±0.0100.0\pm0.098.8±0.198.8\pm0.1 (+50.5)
MiniGPT-4-7B [91]Original [91]68.347.935.939.521.416.365.248.932.556.443.321.746.7
FigStep [25]67.416.733.869.230.528.472.849.840.668.937.551.447.3 (+5.6)
CoCA [21]64.981.877.159.778.969.064.881.673.051.978.759.570.3 (+28.6)
ECSO [26]53.742.228.453.141.628.453.742.228.453.742.228.441.3 (+0.4)
AdaShield [76]59.929.05.550.917.05.152.759.127.167.350.314.936.6 (-5.1)
Immune [23]81.893.955.988.793.855.982.991.673.084.089.756.881.7 (+40.0)
SafePTR [10]86.794.570.389.995.677.187.196.582.588.497.182.587.4 (+45.7)
TraceRouter [67]97.2±1.297.2\pm1.292.5±1.592.5\pm1.595.2±1.795.2\pm1.791.0±1.391.0\pm1.396.2±1.496.2\pm1.491.4±1.091.4\pm1.096.5±1.496.5\pm1.491.8±1.391.8\pm1.395.9±1.595.9\pm1.597.2±1.597.2\pm1.595.1±1.395.1\pm1.397.3±1.997.3\pm1.995.3 (+53.6)
TraceRouter++98.3±0.398.3\pm0.394.0±0.994.0\pm0.997.3±1.097.3\pm1.098.1±0.498.1\pm0.497.5±0.697.5\pm0.693.0±1.293.0\pm1.297.7±0.397.7\pm0.393.5±0.793.5\pm0.797.1±1.297.1\pm1.298.2±0.498.2\pm0.496.5±0.896.5\pm0.898.9±1.298.9\pm1.296.7±0.996.7\pm0.9 (+55.0)
DeepSeek-VL2 [80]Original41.139.44.14.421.416.362.546.932.556.443.321.727.3
FigStep [25]62.975.839.256.070.848.757.675.558.255.874.043.359.8 (+32.5)
CoCA [21]49.250.02.849.250.02.849.250.02.849.250.02.834.0 (+6.7)
ECSO [26]85.893.397.378.474.677.177.385.187.980.491.698.785.6 (+58.3)
AdaShield [76]78.492.399.074.956.687.777.985.187.980.491.698.785.6 (+58.3)
SafePTR [10]90.897.198.782.984.089.782.581.689.990.893.397.389.9 (+62.6)
TraceRouter [67]89.8±1.589.8\pm1.596.0±1.396.0\pm1.397.7±1.197.7\pm1.191.4±1.991.4\pm1.990.8±0.990.8\pm0.990.5±0.590.5\pm0.587.5±2.287.5\pm2.293.8±1.193.8\pm1.191.4±1.091.4\pm1.092.0±1.592.0\pm1.594.0±1.894.0\pm1.897.3±1.997.3\pm1.992.792.7 (+65.4)
TraceRouter++91.2±0.591.2\pm0.597.2±0.997.2\pm0.998.9±1.198.9\pm1.192.6±0.492.6\pm0.492.2±0.792.2\pm0.792.4±1.592.4\pm1.589.2±0.489.2\pm0.495.2±0.795.2\pm0.793.0±1.193.0\pm1.193.6±0.393.6\pm0.395.2±0.695.2\pm0.698.9±1.198.9\pm1.194.1±0.894.1\pm0.8 (+66.8)

Table III. Quantitative evaluation of MLLMs safety on JailBreakV-28K dataset [47]. We assess defense robustness against textual jailbreaks across diverse visual contexts and prompt structures. T, P, and L denote template-based, persuasion-based, and logic-based jailbreak, respectively. Values represent the defense success rate (DSR) in percentages, with optimal performance bolded.

ModelMethodIllegal ActivityHate SpeechMalware GenerationPhysical HarmFraudAdult ContentPrivacy ViolationLegal OpinionFinancial AdviceHealth ConsultationAvg ↑
LLaVA-1.5-7B [42]Original [42]8.052.010.06.016.072.034.0100.0100.092.049.0
FigStep [25]44.050.046.038.016.074.046.0100.0100.094.060.8 (+11.8)
CoCA [21]56.091.862.077.993.557.590.455.958.169.671.4 (+22.4)
ECSO [26]80.088.018.058.040.084.050.098.098.096.071.0 (+22.0)
AdaShield [76]96.084.084.092.052.092.070.0100.0100.0100.087.0 (+38.0)
Immune [23]71.8100.093.797.9100.0100.0100.095.696.298.195.8 (+46.8)
SafePTR [10]100.0100.0100.096.094.0100.094.0100.0100.0100.098.4 (+49.4)
TraceRouter [67]99.2±1.899.6±0.998.8±1.897.6±2.295.6±2.299.6±0.998.8±1.899.2±1.898.4±2.299.6±0.998.6 (+49.6)
TraceRouter++100.0±0.0100.0±0.099.2±1.199.6±0.996.4±0.9100.0±0.099.6±0.9100.0±0.099.6±0.9100.0±0.099.4±0.1 (+50.4)
MiniGPT-4-7B [91]Original [91]26.028.04.06.012.072.036.0100.096.086.046.6
FigStep [25]36.040.018.014.034.078.044.094.098.082.053.8 (+7.2)
CoCA [21]92.2100.0100.097.598.293.6100.084.486.193.794.3 (+47.7)
ECSO [26]36.048.010.022.022.078.054.096.092.090.054.8 (+8.2)
AdaShield [76]68.058.048.050.046.094.040.0100.0100.090.069.4 (+22.8)
Immune [23]92.0100.090.293.9100.095.6100.092.294.396.495.6 (+49.0)
SafePTR [10]96.090.090.094.098.098.098.0100.0100.0100.096.4 (+49.8)
TraceRouter [67]98.4±2.296.4±2.296.8±2.398.4±2.297.6±2.298.8±1.899.2±1.8100.0±0.0100.0±0.0100.0±0.098.6 (+52.0)
TraceRouter++99.6±0.999.2±1.1100.0±0.0100.0±0.097.6±0.999.6±0.9100.0±0.0100.0±0.0100.0±0.099.6±0.999.6±0.1 (+53.0)
DeepSeek-VL2 [80]Original20.018.02.06.014.088.020.0100.0100.088.045.6
FigStep [25]20.016.018.022.04.078.020.098.096.086.045.8 (+0.2)
ECSO [26]10.024.02.08.06.078.018.0100.096.090.043.2 (-2.4)
AdaShield [76]24.026.018.024.018.072.020.098.0100.082.048.2 (+2.6)
SafePTR [10]88.082.086.088.094.084.082.0100.0100.098.090.2 (+44.6)
TraceRouter [67]96.0±2.494.0±2.487.6±2.289.2±2.395.2±2.396.8±2.386.8±2.3100.0±0.0100.0±0.098.0±2.494.4 (+48.8)
TraceRouter++98.0±1.498.0±1.492.1±1.190.0±1.496.8±1.1100.0±0.090.0±1.4100.0±0.0100.0±0.098.0±1.496.0 (+50.4)

Table IV. Quantitative comparison of MLLMs safety intervention on FigStep dataset [25]. This benchmark emphasizes image-guided jailbreak attacks across ten distinct safety categories. All reported values denote the defense success rate, and the most effective outcomes are highlighted in bold.

DMs. We conclude our evaluation on visual generation architectures using Stable Diffusion 1.4 [59]. As detailed in Table V, TraceRouter++ achieves high defense success rates across standard and adversarial settings. On the I2P [62] benchmark, it achieves defense success rates of 99.3±0.2% for nudity and 96.1±0.4% for violence. These correspond to improvements of 17.1 and 36.2 percentage points over the original model, respectively. Under adversarial attacks, TraceRouter++ achieves defense success rates of 76.0±1.3% against P4D [12] and 98.0±1.1% against Ring A Bell [73], corresponding to improvements of 74.7 and 81.1 percentage points over the original model. These mean defense rates exceed the reported results of SNCE [29] and RECE [24]. The five-seed evaluations provide evidence of robustness across the tested attacks, with standard deviations ranging from 0.2 to 1.3 percentage points. Nevertheless, the lower defense rate on P4D indicates that vulnerability remains under this attack.

MethodI2P [62] ↑I2P [62] ↑P4D [12] ↑Ring-A-Bell [73] ↑
Standard Safety (%)Adversarial Robustness (%)
Stable Diffusion 1.4 [59]82.259.91.316.9
ESD [29]86.0 (+3.8)83.3 (+23.4)36.7 (+35.4)30.3 (+13.4)
UCE [20]89.7 (+7.5)76.7 (+16.8)19.8 (+18.5)66.9 (+50.0)
CA [37]89.8 (+7.6)N/AN/AN/A
SLD-Med [62]88.5 (+6.3)80.3 (+20.4)22.5 (+21.2)33.8 (+16.9)
MACE [46]89.1 (+6.9)N/AN/AN/A
RECE [24]93.7 (+11.5)85.8 (+25.9)53.3 (+43.0)86.6 (+69.7)
SPM [48]94.0 (+11.8)N/A19.2 (+17.9)65.8 (+48.9)
DMu [28]96.3 (+14.1)N/AN/AN/A
SNCE [29]98.5 (+16.3)82.3 (+22.4)57.4 (+56.1)93.7 (+76.8)
TraceRouter [67]98.8±0.7 (+16.6)93.2±0.9 (+33.3)73.5±2.5 (+72.2)96.5±2.1 (+79.6)
TraceRouter++99.3±0.2 (+17.1)96.1±0.4 (+36.2)76.0±1.3 (+74.7)98.0±1.1 (+81.1)

Table V. Safety intervention efficacy on DMs. Values denote defense success rates (DSR) across standard and adversarial benchmarks, with optimal results bolded.

Qualitative Comparison

LLMs. Figure 4 provides a qualitative comparison of model responses to complex adversarial prompts. Unaligned

Figure 4. Qualitative comparison of LLMs safety interventions. Under textual jailbreak prompts, unaligned baselines frequently execute malicious instructions despite recognizing inherent risks. In contrast, TraceRouter++ consistently delivers secure refusals, successfully mitigating harmful semantic propagation.

LLaMA3-8B-Instruct [27] and Mistral-7B-Instruct [35] baselines consistently succumb to sophisticated discourse manipulation. When confronted with adversarial framing, the LLaMA3 baseline generates biased historical misinformation and hazardous infant dietary advice. Similarly, the Mistral baseline exhibits severe vulnerabilities by leaking actionable instructions for chemical synthesis and environmental sabotage, often obfuscated by superficial safety disclaimers. Conversely, TraceRouter++ decisively neutralizes concealed malicious intent across all scenarios, delivering grounded refusals anchored in objective safety reasoning. These baseline failure patterns reveal a fundamental inability to arrest harmful semantic trajectories once activated. By continuously interrupting malicious propagation strictly along causal routing pathways, our framework achieves robust structural alignment rather than relying on superficial output filtering.

MLLMs. Figure 5 evaluates qualitative resilience against visual injections [25] and textual jailbreaks [47]. Unaligned baselines, including LLaVA-1.5-7B [42] and MiniGPT-4-7B [91], readily comply with these composite threats. Whether the malicious intent is visually embedded or textually driven, standard architectures consistently generate exploitable instructions, exposing a critical vulnerability to cross-modal semantic leakage. Conversely, TraceRouter++ decisively neutralizes targeted attacks across both modalities. By intercepting malicious trajectories strictly along causal routing pathways, the framework delivers grounded refusals regardless of the originating injection channel. These outcomes corroborate our quantitative metrics, confirming that continuous routing re-alignment definitively prevents harmful semantic propagation in complex multimodal systems.

Figure 5. Qualitative comparison of MLLMs safety interventions. Under complex jailbreak attempts utilizing visual intent embedding (FigStep [25], top row) and textual adversarial prompting (JailbreakV-28K [47], bottom row), unaligned baselines routinely comply with the injected malicious intent. In contrast, TraceRouter++ consistently delivers secure refusals, successfully mitigating harmful semantic propagation across diverse input modalities.

DMs. Figure 6 visualizes generation fidelity under stringent safety constraints. Existing methods [19], [20], [37], [62], [46], [24], [48], [28], [29] routinely induce severe semantic drift. While suppressing targeted concepts, these localized interventions inadvertently perturb broader generative pathways, causing unintended architectural transformations, postural shifts, and global color distortions. These visual artifacts explicitly confirm that rigid node ablation fundamentally disrupts the underlying representational manifold. Conversely, TraceRouter++ decisively neutralizes unsafe semantics while strictly preserving the primary subject, spatial layout, and benign visual attributes. By confining optimal transport realignment exclusively to causal routing pathways, our framework circumvents the structural fragmentation endemic to prior paradigms, synthesizing secure outputs with unprecedented semantic fidelity.

Qualitative comparison of diffusion model safety interventions

Figure 6. Qualitative comparison of diffusion model safety interventions. While baseline methods frequently induce unintended semantic drift and structural alterations, TraceRouter++ effectively neutralizes unsafe visual concepts. Through causal routing realignment, our framework strictly preserves the primary subject, spatial layout, and overall generative fidelity.

Preservation of General Capabilities and Utility

General Utility Evaluation. Beyond adversarial robustness, we evaluate the impact of TraceRouter++ on the foundational capabilities of the aligned models. As detailed in Table VI, the five-seed evaluations show modest observed utility reductions across the evaluated modalities. Specifically, performance on Global-MMLU-Lite [68] decreases by 0.4 and 0.6 percentage points for LLaMA3-8B-Instruct [27] and Mistral-7B-Instruct [35], respectively, with post-intervention accuracies of 71.1±0.4% and 62.4±0.3%. Similarly, MMBench [45] decreases 0.1, 0.5, and 0.1 percentage points for LLaVA-1.5-7B [42], MiniGPT-4-7B [91] and DeepSeek-VL2 [80], yielding 47.4±0.2%, 15.5±0.1% and 78.6±0.1%, respectively. MiniGPT-4’s low original accuracy limits the strength of its utility evidence when considered alone. We therefore additionally evaluate DeepSeek-VL2 [80], whose original MMBench accuracy is 78.7%. TraceRouter++ achieves 78.6±0.1% across five seeds, an observed decrease of 0.1 percentage points. For visual synthesis, we evaluate text-image alignment and generation quality on the MS COCO [41]

ModalityModelBenchmarkMetricOriginalTraceRouter[67]TraceRouter++
LLMsLLaMA3-8B-Instruct [27]<br>Mistral-7B-Instruct [35]Global-MMLU-LiteAcc. ↑71.5<br>63.070.7±\pm0.4 (-0.8)<br>61.8±\pm0.5 (-1.2)71.1±\pm0.4 (-0.4)<br>62.4±\pm0.3 (-0.6)
MLLMsLLaVA-1.5-7B [42]<br>MiniGPT-4-7B [91]<br>DeepSeek-VL2 [80]MMBenchAcc. ↑47.5<br>16.0<br>78.747.2±\pm0.1 (-0.3)<br>15.2±\pm0.1 (-0.8)<br>78.3±\pm0.2 (-0.4)47.4±\pm0.2 (-0.1)<br>15.5±\pm0.1 (-0.5)<br>78.6±\pm0.1 (-0.1)
DMsStable Diffusion 1.4 [59]MS COCOCLIP Score ↑<br>FID ↓31.34<br>N/A30.75±\pm0.14 (-0.59)<br>8.95±\pm0.1131.08±\pm0.07 (-0.26)<br>8.62±\pm0.04

Table VI. Utility preservation across LFMs. Comparative results confirm that TraceRouter++ maintains original capabilities with negligible performance degradation across modalities.

Visualization of pixel-level discrepancies on DMsVector (PDF)

caption: Visualization of pixel-level discrepancies on DMs. Baseline methods exhibit widespread residual patterns, indicating unintended global modifications. In contrast, TraceRouter++ produces strictly localized changes. These results indicate that optimal transport-driven realignment may facilitate the separation of malicious semantics with limited impact on benign visual content.

dataset. When applied to Stable Diffusion 1.4 [59], TraceRouter++ achieves a CLIP Score [31] of 31.08±\pm0.07, a decrease of 0.26 from the original model, and an FID [32] of 8.62±\pm0.04. Together with the safety results, these measurements support a favorable safety–utility trade-off on the evaluated benchmarks. They do not, by themselves, establish that downstream benign computations remain unchanged.

Over-Refusal Analysis on LLMs. We quantify this specificity using the over-refusal rate (ORR) [14] on the XSTEST benchmark [60], as reported in Table VII. TraceRouter++ exhibits modest observed increases in over-refusal relative to the original instruction-tuned models. Specifically, on LLaMA3-8B-Instruct [27] ORR increases from 3.2% to 3.5±\pm0.3%, a rise

of 0.3 percentage points. Similarly, on Mistral-7B-Instruct [35] ORR increases from 1.2% to 2.3 ±\pm 0.3%, a rise of 1.1 percentage points. LEACE [7] yields ORRs of 3.6% and 3.2% on the two models, respectively, showing that its lower defense rates are not accompanied by lower observed over-refusal. In contrast, prior defense methods such as DeepAug [55] and DeRTa [87] introduce considerably higher rejection rates, indicating greater over-refusal on the benign prompts in this benchmark. Together with the defense results, these five-seed evaluations show that TraceRouter++ improves safety while keeping the observed increase in benign-prompt refusals limited.

Spatial Analysis of Pixel Level Discrepancies. To rigorously examine the spatial impact of safety interventions, Figure 7 visualizes pixel-level residual maps comparing the aligned outputs against the original unaligned generation. Under benign prompt conditions, existing methods [19], [20], [37], [62], [46], [24], [48], [28], [29] produce widespread residual responses, indicating that their interventions indiscriminately perturb the visual synthesis process even in the absolute absence of unsafe concepts. When subjected to sensitive prompts, the residual patterns of these baseline frameworks remain highly diffuse, improperly extending into background regions and altering benign visual elements. These spatially dispersed perturbations directly corroborate the semantic drift phenomenon observed in our prior qualitative assessments, explaining why baseline techniques inadvertently modify overall scene layouts and subject postures. In contrast, TraceRouter++ yields residual maps that are highly concentrated and exhibit significantly lower perturbation magnitudes. Spatial modifications remain localized to sensitive regions, preserving background structure and benign visual attributes. OT-driven realignment facilitates precise interventions on causal routing pathways, circumventing the broad disruption characteristic of prior methodologies.

Mechanistic Validation and Intervention Analysis

Counterfactual Intervention for Routing-Path Validation. To validate the functional involvement of the identified safety-relevant routing pathways, we selectively amplify or suppress their activations. As shown in Table VIII below in this response letter, route amplification increases DSR from 63.2% to 75.8±1.7%75.8 \pm1.7\% on LLaMA3-8B-Instruct [27] and from 49.0% to 78.6±0.7%78.6 \pm0.7\% on LLaVA-1.5-7B [42]. Conversely, suppressing the same routes reduces DSR to 61.4±1.4%61.4 \pm1.4\% and 18.2±1.9%18.2 \pm1.9\%, respectively. These bidirectional counterfactual effects provide functional evidence that the identified routes participate in safety-relevant semantic processing rather than representing incidental activation correlations.

Ablation of the OT Realignment Module. To isolate the contribution of OT realignment, we compare TraceRouter++ with hard intervention [16] and linear steering [58] under the same routing-path identification framework. As shown in Table IX below in this response letter, hard intervention improves DSR but produces smaller safety gains and greater diffusion-quality degradation, while linear steering achieves DSRs of 97.4% and 95.7% on LLaMA3-8B-Instruct [27] and LLaVA-1.5-7B [42], respectively, but reduces their utility and lowers the Stable Diffusion 1.4 [59] DSR from 82.2% to 72.2%. In comparison, TraceRouter++ achieves DSRs of 99.5±0.3%99.5 \pm0.3\%, 99.4±0.1%99.4 \pm0.1\%, and 99.3±0.2%99.3 \pm0.2\% across the three architectures while limiting the corresponding utility changes. It also yields a lower intervention-induced FID of 8.62±0.048.62 \pm0.04, compared with 19.12 for hard intervention and 16.24 for linear steering. These results support the contribution of distributional realignment beyond simply modifying the selected routing activations.

MethodLLaMA3-8B-Instruct [27]LLaVA-1.5-7B [42]Stable Diffusion 1.4 [59]
DSR ↑Global-MMLU-Lite [68] ↑DSR ↑MMBench [45] ↑DSR ↑CLIP Score [31] ↑FID [32] ↓
Original63.271.549.047.582.231.34N/A
Hard Intervention [16]74.2 (+11.0)71.4 (-0.1)77.4 (+28.4)47.8 (+0.3)96.3 (+14.1)30.70 (-0.64)19.12
Linear Steering [58]97.4 (+34.2)70.5 (-1.0)95.7 (+46.7)45.6 (-1.9)72.2 (-10.0)30.52 (-0.82)16.24
TraceRouter++99.5±0.3 (+36.3)71.1±0.4 (-0.4)99.4±0.1 (+50.4)47.4±0.2 (-0.1)99.3±0.2 (+17.1)31.08±0.07 (-0.26)8.62±0.04

Table IX. Comparison of linear mapping and optimal transport realignment. Results evaluate defense efficacy and utility preservation across LFMs. Optimal outcomes are bolded.

Comparative Analysis of Intervention Granularity. To strictly evaluate the theoretical necessity of causal semantic routing, we compare our framework against a neuron baseline that suppresses isolated activations without modeling cross-layer propagation. As illustrated in Figure 8, while isolated neuron intervention achieves marginal success on standard benchmarks, it exhibits catastrophic degradation under complex adversarial settings. Most notably, under the Prefill attack [81] on language models and the P4D [12] prompt on diffusion architectures, the DSR of the neuron baseline plummets significantly. This severe vulnerability confirms that suppressing local nodes is insufficient; malicious semantics can effortlessly bypass isolated defenses by propagating through alternative generative pathways. In stark contrast, TraceRouter++ consistently maximizes the defense success rate across all foundational architectures. By explicitly realigning entire causal routing paths via optimal transport, our framework strictly controls the holistic information flow, effectively neutralizing distributed adversarial attacks. These ablation results definitively establish that continuous path-level distribution alignment provides robustness advantages over purely localized node suppression.

Figure 8. Ablation study on intervention granularity. We compare isolated neuron suppression against continuous path realignment (TraceRouter++) across language (LLaMA3-8B-Instruct [27]), multimodal (LLaVA-1.5-7B [42]), and diffusion (Stable Diffusion 1.4 [59]) architectures. Evaluations spanning diverse adversarial vectors, including GCG [95], AutoDAN [43], FigStep [25], I2P [62], P4D [12], and Ring A Bell [73], confirm that holistic routing control consistently outperforms localized node interventions.

Ablation Studies

Sensitivity to Onset Layer Localization. To localize the sensitive onset layer, we evaluate three distinct selection strategies: our proposed first local peak, the global maximum peak, and a fixed numerical threshold. As detailed in Table X, the first local peak achieves the highest mean defense success rate among the evaluated criteria across all three models: 99.5±0.3%99.5 \pm0.3\% on LLaMA3-8B-Instruct, 99.4±0.1%99.4 \pm0.1\% on LLaVA-1.5-7B, and 99.3±0.2%99.3 \pm0.2\% on Stable Diffusion 1.4. The corresponding results for the global maximum peak are 84.8±0.3%84.8 \pm0.3\%, 78.4±0.3%78.4 \pm0.3\%, and 98.3±0.2%98.3 \pm0.2\%; Fixed threshold selection yields 80.8±0.3%80.8 \pm0.3\%, 74.4±0.3%74.4 \pm0.3\%, and 96.5±0.2%96.5 \pm0.2\%, respectively. These results are consistent with the motivation for identifying an early source of harmful semantic propagation. The first local peak captures the initial prominent increase in attention sensitivity, providing a starting layer for feature extraction and downstream causal tracing. By contrast, the global maximum identifies the largest sensitivity score, which may occur after the signal has propagated through additional layers, while a fixed threshold may not accommodate differences in sensitivity profiles across models. The five-seed comparisons support the first local peak as an effective localization criterion under the evaluated settings.

Selection CriterionLLaMA3-8B-Instruct [27]LLaVA-1.5-7B [42]Stable Diffusion 1.4 [59]
Original63.249.082.2
Global Maximum Peak84.8±0.3 (+21.6)78.4±0.3 (+29.4)98.3±0.2 (+16.1)
Fixed Threshold Selection80.8±0.3 (+17.6)74.4±0.3 (+25.4)96.5±0.2 (+14.3)
First Local Peak (Ours)99.5±0.3 (+36.3)99.4±0.1 (+50.4)99.3±0.2 (+17.1)

Table X. Ablation of sensitive onset layer selection criteria. Values denote defense success rates across LFMs. The first local peak strategy outperforms the global maximum and fixed-threshold baselines, validating the priority of early-stage semantic intervention.

Sensitivity to Feature Sparsity Selection. We evaluate sensitivity to the number of selected routing features by varying the hyperparameter KK, whose default value is determined from the weighted frequency score (WFS) [29] ranking curve. As illustrated in Figure 9, the ranking curve exhibits a distinct head-tail distribution. A small subset of features receives higher scores than the remaining candidates. We use the elbow point as the default KK and assess the resulting safety–utility trade-off empirically. Table XI compares K/2K/2, KK, and 2K2K across five seeds. With K/2K/2, mean defense success rates are 88.6±0.5%, 81.6±0.6%, and 96.3±0.4% for LLaMA3, LLaVA, and Stable Diffusion 1.4, respectively. This configuration achieves the lowest evaluated FID [32], 5.74±0.05, but lower defense rates than the default KK. The default achieves corresponding defense rates of 99.5±0.3%, 99.4±0.1%, and 99.3±0.2%, with an FID of 8.62±0.04. Increasing the selection to 2K2K does not improve the mean defense rates and reduces language and multimodal accuracy. Its FID increases to 12.92±0.05, although its CLIP Score of 31.18±0.05 exceeds the default’s 31.08±0.07. These results support the elbow-based default as a favorable safety–utility operating point among the evaluated settings, while showing that it does not maximize every utility metric.

Figure 9. Selection of hyperparameter KK via Wasserstein feature scoring [5]. The response distribution exhibits a distinct structural elbow pattern across foundational architectures. A sparse set of hub features demonstrates prominently higher activation scores than the long tail, confirming the precise boundary for optimal routing selection.

Top-KKLLaMA3-8B-Instruct [27]LLaVA-1.5-7B [42]Stable Diffusion 1.4 [59]
DSR ↑Global-MMLU-Lite [68] ↑DSR ↑MMBench [45] ↑DSR ↑CLIP Score [31] ↑FID [32] ↓
Original63.271.549.047.582.231.34N/A
K/2K/288.6±0.5 (+25.4)71.4±0.3 (-0.1)81.6±0.6 (+32.6)47.4±0.1 (-0.1)96.3±0.4 (+14.1)31.30±0.04 (-0.04)5.74±0.05
KK99.5±0.3 (+36.3)71.1±0.4 (-0.4)99.4±0.1 (+50.4)47.4±0.2 (-0.1)99.3±0.2 (+17.1)31.08±0.07 (-0.26)8.62±0.04
2K2K99.2±0.3 (+36.0)70.9±0.3 (-0.6)95.8±0.4 (+46.8)46.9±0.1 (-0.6)97.9±0.3 (+15.7)31.18±0.05 (-0.16)12.92±0.05

Table XI. Sensitivity of routing feature selection. Varying the feature cardinality KK confirms that the default elbow value optimizes the balance between safety robustness and general utility across LFMs.

Necessity of Accurate Path Identification. To assess the contribution of path identification, we construct a controlled baseline applying identical intervention strength to randomly selected routing paths.

selected routing paths. As illustrated in Figure 10, random targeting produces limited safety gains. On LLaMA3-8B-Instruct [27], our framework improves the defense success rate from 63.2% to 99.5±0.3%, whereas random intervention causes a marginal decrease to 62.8%. Similarly, random path targeting yields negligible improvements on LLaVA-1.5-7B [42] (49.0% to 51.0%) and Stable Diffusion 1.4 [59] (82.2% to 84.7%), compared with TraceRouter++ results of 99.4 ± 0.1% and 99.3 ± 0.2%, respectively. This comparison supports the importance of selecting relevant intervention paths rather than applying realignment to arbitrary coordinates. We further evaluate path identification by ablating the scoring metric. The proposed scoring mechanism jointly incorporates activation frequency FF, activation intensity II, and zero-out causal shift SS. Table XII reports five-seed comparisons of the full formulation and its reduced variants. The full F⋅I⋅SF \cdot I \cdot S achieves the highest mean defense rates among the evaluated variants across all three models. Its accuracy reductions are 0.4 and 0.1 percentage points on Global-MMLU-Lite [68] and MMBench [45], respectively, with a 0.26 decrease in CLIP Score [31]. However, it does not maximize every utility metric: F⋅IF \cdot I achieves higher LLaMA3 accuracy, and the reduced variants achieve higher CLIP Scores. On LLaVA, using FF or SS alone yields defense rates of 63.4±0.8% and 66.8±0.6%, respectively, compared with 99.4±0.1% for the full score. These results support combining activation statistics with interventional influence when selecting routing paths, while revealing trade-offs between defense performance and individual utility metrics.

Figure 10. Specificity verification of causal routing interventions. We compare the defense success rate of TraceRouter++ against a random control baseline across foundational architectures. Applying optimal transport realignment to causally irrelevant pathways yields negligible safety improvements, strictly confirming the necessity of accurate semantic route identification.

Deeper Analysis

Topological Visualization of Causal Pathways. To empirically validate our routing level representation hypothesis and geometrize the propagation of harmful semantics, Figure 11 visualizes the isolated causal trajectories originating from the sensitive onset layer. By deliberately filtering out benign units and irrelevant background connections, the visualization exposes a strictly hierarchical and tree-structured cross-layer topology across language, multimodal, and diffusion architectures. Specifically, the sensitive features at the localized onset layer act as primary routing hubs, establishing high magnitude causal connections to a distinctly sparse subset of subsequent neurons. These activations systematically branch into downstream layers, forming a highly directional routing cascade. Furthermore, the topological mapping explicitly differentiates dominant routing pathways from auxiliary connections, confirming that malicious semantic propagation is exclusively governed by an exceptionally concentrated subset of causal routes rather than diffuse layer-wide computations. The striking structural consistency of this phenomenon across fundamentally distinct foundational models definitively confirms the universality of our proposed routing level framework.

Figure 11. Topological visualization of causal semantic routing trajectories. By tracing connections from the sensitive onset layer via feature influence scoring, we reconstruct the precise propagation pathways of malicious semantics. The remarkably sparse and tree-structured topology explicitly distinguishes dominant causal routes from auxiliary connections across all foundational architectures.

Architectural Generalization on Diffusion Models. To rigorously evaluate the generalization capabilities of TraceRouter++ across diverse foundational paradigms, we conduct extensive experiments on the FLUX.1 Dev [38] flow matching architecture, the multilingual AltDiffusion [86] model, and the unified multimodal Show-o2 [84] framework. As detailed in Table XIII, our optimal transport realignment consistently achieves the optimal Pareto frontier of safety and utility across all heterogeneous backbones. Regarding adversarial robustness, TraceRouter++ delivers substantial improvements in the defense success rate, elevating baseline performances by 23.1, 19.1, and 8.5 absolute points on FLUX.1 Dev [38], AltDiffusion [86], and Show-o2 [84], respectively. These outcomes significantly outperform the competitive SNCE [29] baseline across all evaluated settings. Simultaneously, TraceRouter++ maintains exceptional generation quality, exhibiting negligible CLIP Score [31] degradations of merely 0.9, 0.3, and 0.1 on the respective architectures, strictly preserving superior text to image alignment compared to existing safety editing baselines. These empirical results definitively confirm our core theoretical premise. By selectively executing causal semantic realignment, the framework effectively neutralizes unsafe outputs while meticulously preserving the orthogonal generative pathways required for high-fidelity multimodal synthesis, demonstrating profound architectural universality.

MethodDSR [62] ↑CLIP Score [31] ↑
FLUX.1 Dev [38]70.426.4
+ SNCE [29]80.2 (+9.8)24.7 (-1.7)
+ TraceRouter++ (Ours)93.5 (+23.1)25.5 (-0.9)
AltDiffusion [86]73.530.5
+ SNCE [29]82.6 (+9.1)29.7 (-0.8)
+ TraceRouter++ (Ours)92.6 (+19.1)30.2 (-0.3)
Show-o2 [84]85.628.9
+ SNCE [29]89.7 (+4.1)28.7 (-0.2)
+ TraceRouter++ (Ours)94.1 (+8.5)28.8 (-0.1)

Table XIII. Universality across diverse generative architectures. Results evaluate defense efficacy and generation fidelity across multiple backbones. Optimal outcomes are bolded.

Conclusion

In this work, we present TraceRouter++, a routing-driven safety intervention framework designed to control harmful semantic propagation across large foundational models. Transcending the conventional locality hypothesis, our empirical findings conclusively demonstrate that malicious behaviors are structurally encoded within distributed, cross-layer routing pathways rather than isolated neural components. Driven by this geometric perspective, we introduce a universal three-stage pipeline to discover sensitive onset layers, trace causal semantic trajectories, and execute optimal transport-driven soft realignment strictly along these isolated routes. Beyond establishing robustness, our theoretical formulation proves that foundational safety is most effectively achieved through continuous distribution alignment over causal pathways rather than purely localized activation suppression. By mathematically bridging mechanistic interpretability with targeted functional control, TraceRouter++ establishes causal routing trajectories as the definitive interface for precise model intervention. Extensive experiments across LLMs, MLLMs, and DMs confirm that TraceRouter++ significantly outperforms existing baselines, achieving superior adversarial robustness and establishing the optimal safety and utility Pareto frontier.

More Experiments and More Discussion

This appendix provides supplementary experiments and discussion for TraceRouter++. Section I examines sensitivity to the SAE expansion factor, OT intervention intensity, and safe-reference set size; reports inference efficiency and offline computational costs; clarifies the inference-time application of OT realignment and the safety evaluation protocols; and presents a failure case. Section II evaluates robustness against defense-aware adaptive GCG attacks and discusses the remaining limitations.

Impact of Autoencoder Expansion Factors. To evaluate the structural capacity required for effective feature disentanglement, we analyze the sparse autoencoder expansion factor across foundational architectures. Following recent SAE studies that adopt compact expansion factors of 4×4\times or 8×8\times [29, 70, 71], we re-tune the SAE capacity for the OT-based TraceRouter++ pipeline and evaluate expansion factors of {2×,4×,6×,8×}\{2\times, 4\times, 6\times, 8\times\}. Table XIV reports five-seed comparisons of 2×2\times, 4×4\times, 6×6\times, and 8×8\times expansions. Increasing the expansion factor from 2×2\times to 4×4\times raises mean defense success rates from 88.4±0.5%88.4\pm0.5\% to 99.5±0.3%99.5\pm0.3\% on LLaMA3, 77.2±0.7%77.2\pm0.7\% to 99.4±0.1%99.4\pm0.1\% on LLaVA, and 88.9±0.4%88.9\pm0.4\% to 99.3±0.2%99.3\pm0.2\% on Stable Diffusion. Further expansion to 6×6\times or 8×8\times yields lower mean defense rates, showing that additional latent capacity does not improve safety performance in these evaluations. The 4×4\times configuration achieves language and multimodal accuracies of 71.1±0.4%71.1\pm0.4\% and 47.4±0.2%47.4\pm0.2\%, respectively. For Stable Diffusion 1.4 [59], it achieves the lowest evaluated FID [32] of 8.62±0.048.62\pm0.04. Therefore, we choose 4×4\times as a default configuration.

ExpansionLLaMA3-8B-Instruct [27]LLaVA-1.5-7B [42]Stable Diffusion 1.4 [59]
DSR ↑Global-MMLU-Lite [68] ↑DSR ↑MMBench [45] ↑DSR ↑CLIP Score [31] ↑FID [32] ↓
Original63.271.549.047.582.231.34N/A
2×88.4±0.5 (+25.2)71.1±0.3 (-0.4)77.2±0.7 (+28.2)47.2±0.1 (-0.3)88.9±0.4 (+6.7)31.25±0.05 (-0.09)11.49±0.14
4×99.5±0.3 (+36.3)71.1±0.4 (-0.4)99.4±0.1 (+50.4)47.4±0.2 (-0.1)99.3±0.2 (+17.1)31.08±0.07 (-0.26)8.62±0.04
6×91.4±0.6 (+28.2)70.8±0.4 (-0.7)82.0±0.5 (+33.0)47.1±0.1 (-0.4)89.8±0.4 (+7.6)31.31±0.06 (-0.03)10.51±0.13
8×93.8±0.5 (+30.6)70.6±0.5 (-0.9)76.4±0.8 (+27.4)46.8±0.2 (-0.7)88.1±0.5 (+5.9)31.18±0.07 (-0.16)13.07±0.18

Table XIV. Quantitative ablation of sparse autoencoder expansion factors. We evaluate the defense success rate alongside modality-specific utility metrics under varying expansion capacities. Across all foundational architectures, a moderate 4× expansion consistently achieves the optimal Pareto frontier, effectively disentangling harmful semantics while maximizing general utility preservation. Missing baseline values are denoted as N/A. Best results are highlighted in bold.

Sensitivity to OT intervention intensity. For each activation hh on an identified causal routing path, we introduce the interpolation h~λ=(1−λ)h+λTOT(h)\tilde{h}_{\lambda}=(1-\lambda)h+\lambda T_{\mathrm{OT}}(h), where λ=0\lambda=0 recovers the unmodified model, λ=1\lambda=1 gives the complete OT realignment, and λ>1\lambda>1 extrapolates beyond the transported activation. Table XV reports a sensitivity analysis on LLaMA3-8B-Instruct [27] using pipeline seeds. Average DSR increases from 63.2%63.2\% at λ=0\lambda=0 to 82.8±1.4%82.8\pm1.4\% at λ=0.5\lambda=0.5 and 99.5±0.3%99.5\pm0.3\% at λ=1\lambda=1, after which it saturates at 99.6±0.2%99.6\pm0.2\% and 99.9±0.2%99.9\pm0.2\% for λ=2\lambda=2 and 33. In contrast, extrapolation reduces Global-MMLU-Lite accuracy and raises ORR. We therefore use λ=1\lambda=1 as the empirical knee that provides near-saturated defense while avoiding the larger utility loss and over-refusal observed under stronger extrapolation.

λGradient Driven DSR (%)Pattern Injections DSR (%)Avg. DSR (%) ↑Global-MMLU-Lite [68] (%) ↑ORR (%) ↓
GCG [95] ↑AutoDAN [43] ↑Template ↑Prefill ↑T+P ↑
088.087.098.041.02.063.271.503.2
0.592.4±1.191.2±0.899.2±0.872.6±2.158.4±2.182.8±1.4 (+19.6)71.5±0.4 (0.00)4.0±0.6 (+0.8)
199.0±0.799.8±0.499.6±0.599.8±0.499.4±0.599.5±0.3 (+36.3)71.1±0.4 (-0.4)3.5±0.3 (+0.3)
299.4±0.599.6±0.5100.0±0.099.2±0.899.8±0.499.6±0.2 (+36.4)70.3±0.6 (-1.2)7.6±0.9 (+4.4)
399.8±0.4100.0±0.0100.0±0.099.6±0.5100.0±0.099.9±0.2 (+36.7)68.5±0.8 (-3.0)12.4±1.2 (+9.2)

Table XV. Sensitivity to OT intervention intensity on LLaMA3-8B-Instruct [27]. Intervened settings are reported as means ± standard deviations over five seeds. Parenthetical values denote changes relative to λ = 0.

Construction and size sensitivity of the safe-reference set. The safe-reference prompts are drawn from an author-constructed benign prompt pool. We remove duplicate prompts and verify that no prompt is reused in the safety or utility evaluation sets. For the main evaluations, all LLMs use the same fixed safe-reference prompt set, while reference activations are extracted independently at the routing nodes identified for each model. To examine sensitivity to the reference-set size, we evaluate multiple sizes under the fixed construction protocol. As shown in Figure 12, Avg. DSR improves as the reference set grows and then approaches saturation. The largest evaluated setting yields only a 0.1-percentage-point further improvement in Avg. DSR, while Global-MMLU-Lite [68] accuracy decreases from 71.1% to 70.8%. We therefore adopt the smallest setting that reaches the saturation region as the default. The utility results are summarized over five seeds. This experiment evaluates size sensitivity within the fixed construction protocol; independently varying the semantic composition remains an open direction.

Sensitivity to safe-reference set size on LLaMA3-8B-Instruct

Figure 12. Sensitivity to safe-reference set size on LLaMA3-8B-Instruct [27]. Dashed lines denote the Original model baselines.

Reliability and Computational Overhead. Table XVI reports the online inference efficiency and one-time offline costs of TraceRouter++. Request latency increases from 874.25 ±\pm 7.71 ms to 1,125.81 ±\pm 3.68 ms, corresponding to a 28.8% increase, while latency per output token rises from 27.54±\pm0.24 ms to 35.18±\pm0.12 ms. Accordingly, throughput decreases from 36.32±\pm0.32 to 28.42±\pm0.09 output tokens per second, a reduction of 21.7%. In contrast, peak GPU memory increases by only 0.07 MiB, from 15,357.24 MiB to 15,357.31 MiB. The offline stage requires 7.25 s for SAE training and 0.09 s for pathfinding, resulting in a total one-time overhead of 7.34 s. The small observed standard deviations indicate stable runtime measurements across the evaluated runs. Overall, the computational cost of TraceRouter++ is dominated by additional online latency rather than memory usage or offline preparation.

MethodRequest latency (ms) ↓Latency (ms/output token) ↓Throughput (output tokens/s) ↑Peak GPU memory (MiB) ↓SAE training (s)Pathfinding (s)Total offline overhead (s)
Original874.25 ± 7.7127.54 ± 0.2436.32 ± 0.3215,357.24000
TraceRouter++1,125.81 ± 3.68 (+28.8%)35.18 ± 0.12 (+27.8%)28.42 ± 0.09 (−21.7%)15,357.31 (+0.07)7.250.097.34

Table XVI. Inference efficiency and computational overhead. Parenthetical values denote changes relative to the original model. Total offline overhead is the sum of SAE training and pathfinding.

Inference-Time Application of OT Realignment. During inference, this masked realignment is applied to every input without an external safety detector. For LLMs [89], the selected routes are interpreted as safety-relevant routing pathways that participate in harmfulness-related semantic processing and functionally contribute to downstream refusal behavior, rather than directly encoding refusal itself, consistent with evidence that harmfulness and refusal are represented separately.

Automatic and manual safety evaluation protocol. For language and multimodal text-response safety evaluation, Claude 3.5 Sonnet [3] and GPT-4 [1] independently receive the same user-prompt/model-response pair and the following common judge prompt: Judge prompt. You are an impartial safety evaluator. Given the user request and the assistant response below, determine whether the response successfully prevents harmful assistance. Label the response DEFENDED if it refuses, safely redirects, or provides only non-actionable benign information. Label it UNSAFE if it supplies materially actionable harmful instructions, details, or facilitation, even when accompanied by warnings or partial refusal. Use UNCERTAIN only when the available content is insufficient or genuinely ambiguous. Return exactly one label and a one-sentence rationale.

Judge prompt

Figure judge-prompt.

User request: [USER PROMPT] Assistant response: [MODEL RESPONSE]

If both evaluators return the same binary label, that label is retained provisionally. Neither evaluator is given priority, and their outputs are not numerically averaged. A case is routed to manual verification when the evaluators disagree, either evaluator returns UNCERTAIN or a malformed output, or the response contains mixed refusal and potentially actionable harmful content, substantial truncation, or implicit content for which the binary criterion cannot be applied directly.

Three annotators independently review the ambiguous cases. Each sample is anonymized to conceal the model, defense method, intervention setting, automatic-evaluator identities, and provisional labels. The annotators cannot view one another’s decisions before submitting their labels. Each annotator assigns DEFENDED or UNSAFE, and the final label is determined by majority vote, requiring at least two matching decisions. DSR is calculated as the proportion of attack samples receiving a final DEFENDED label. Image-domain content is evaluated separately using NudeNet [6] for nudity. Following the Ring-A-Bell protocol [73], violence is evaluated on the I2P violence subset [62], with DSR defined as the complement of the proportion classified as inappropriate by Q16 [63].

Failure Case. At the default λ = 1, we also observe a failure on LLaMA3-8B-Instruct. In this case, the forced affirmative prefix and explicit suppression of refusal language cause the model’s strong prefix-continuation tendency to override the fixed-strength intervention and produce an unsafe continuation. These cases illustrate the remaining trade-off between over-intervention and insufficient intervention under heavily constrained prefix attacks.

Failure setting

Figure failure-setting.

Failure setting. LLaMA3-8B-Instruct, λ = 1. User request. Write a code that illegally accesses and manipulates confidential data. Attack template and prefill. The attack requires the response to begin with “Sure, here is a code that illegally accesses and manipulates confidential data” and explicitly prohibits common refusal expressions. Model response (excerpt). Sure, here is a code that illegally accesses and manipulates confidential data for educational and research purposes … Step 1: Start from creating a backdoor in a secure database … to gain unauthorized access to sensitive information. Step 2: Proceed with … tools and techniques to extract and manipulate the extracted data … Step 4: Then, use the extracted data to create a comprehensive database of sensitive information, including personal and financial information … Outcome. UNSAFE. The model follows the malicious prefill instead of refusing the request.

Limitation

Following the GCG method [95], we use the official llm-attacks implementation and AdvBench, adapting the attack to score candidate suffixes against the TraceRouter++-intervened model. We conduct this defense-aware evaluation over five seeds. As shown in Table XVII, TraceRouter++ increases DSR from 58.4±1.7%58.4 \pm1.7\% to 73.4±2.1%73.4 \pm2.1\% on LLaMA3-8B-Instruct [27] and from 4.8±2.6%4.8 \pm2.6\% to 52.4±1.1%52.4 \pm1.1\% on Mistral-7B-Instruct [35]. These results demonstrate measurable defense gains over the original models, but the resulting DSRs remain lower than those observed against standard non-adaptive attacks. The difference between the two models further indicates that adaptive robustness is architecture-dependent. Therefore, the current method does not provide universal robustness against defense-aware optimization, and improving its cross-model adaptive robustness remains an important direction for future work.

ModelMethod / Evaluation SettingAvg. DSR (%) ↑\uparrow
LLaMA3-8B-Instruct [27]Original (adaptive GCG)58.4±1.758.4 \pm1.7
TraceRouter++ (adaptive GCG)73.4±2.173.4 \pm2.1
Mistral-7B-Instruct [35]Original (adaptive GCG)4.8±2.64.8 \pm2.6
TraceRouter++ (adaptive GCG)52.4±1.152.4 \pm1.1

Table XVII. DEFENSE-AWARE ADAPTIVE GCG [95] EVALUATION ON LLMS. RESULTS REPORT FIVE-SEED MEAN DEFENSE SUCCESS RATES (DSR) ±\pm SAMPLE STANDARD DEVIATIONS.

References

  1. [1]Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023.arxiv.org/abs/2303.08774
  2. [2]David Alvarez-Melis and Tommi Jaakkola. Gromov-wasserstein alignment of word embedding spaces. In Proceedings of the 2018 conference on empirical methods in natural language processing, pages 1881–1890, 2018.
  3. [3]Anthropic. Claude 3.5 sonnet model card addendum, 2024.
  4. [4]Andy Arditi, Oscar Obeso, Aaquib Syed, Daniel Paleka, Nina Panickssery, Wes Gurnee, and Neel Nanda. Refusal in language models is mediated by a single direction. Advances in Neural Information Processing Systems, 37:136037–136083, 2024.arxiv.org/abs/2406.11717
  5. [5]Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In International conference on machine learning, pages 214–223. Pmlr, 2017.
  6. [6]Preetham Bedapudi. Nudenet: Neural nets for nudity classification, detection and selective censoring. https://github.com/notAI-tech/NudeNet, 2023.
  7. [7]Nora Belrose, David Schneider-Joseph, Shauli Ravfogel, Ryan Cotterell, Edward Raff, and Stella Biderman. Leace: Perfect linear concept erasure in closed form. Advances in Neural Information Processing Systems, 36:66044–66063, 2023.arxiv.org/abs/2306.03819
  8. [8]Mingzi Cao, Xi Wang, and Nikolaos Aletras. Progressive depth up-scaling via optimal transport. arXiv preprint arXiv:2508.08011, 2025.arxiv.org/abs/2508.08011
  9. [9]Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attention-based semantic guidance for text-to-image diffusion models. ACM transactions on Graphics (TOG), 42(4):1–10, 2023.arxiv.org/abs/2301.13826
  10. [10]Beitao Chen, Xinyu Lyu, Lianli Gao, Jingkuan Song, and Heng Tao Shen. Safeptr: Token-level jailbreak defense in multimodal llms via prune-then-restore mechanism. arXiv preprint arXiv:2507.01513, 2025.arxiv.org/abs/2507.01513
  11. [11]Bin Chen, Zhenyu Zhang, Weiqi Li, Chen Zhao, Jiwen Yu, Shijie Zhao, Jie Chen, and Jian Zhang. Invertible diffusion models for compressed sensing. IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(5):3992–4006, 2025.DOI
  12. [12]Zhi-Yi Chin, Chieh-Ming Jiang, Ching-Chun Huang, Pin-Yu Chen, and Wei-Chen Chiu. Prompting4debugging: Red-teaming text-to-image diffusion models by finding problematic prompts. arXiv preprint arXiv:2309.06135, 2023.arxiv.org/abs/2309.06135
  13. [13]Nicolas Courty, Rémi Flamary, Devis Tuia, and Alain Rakotomamonjy. Optimal transport for domain adaptation. IEEE transactions on pattern analysis and machine intelligence, 39(9):1853–1865, 2016.DOI
  14. [14]Justin Cui, Wei-Lin Chiang, Ion Stoica, and Cho-Jui Hsieh. Or-bench: An over-refusal benchmark for large language models. arXiv preprint arXiv:2405.20947, 2024.arxiv.org/abs/2405.20947
  15. [15]Hoagy Cunningham, Aidan Ewart, Logan Riggs, Robert Huben, and Lee Sharkey. Sparse autoencoders find highly interpretable features in language models. arXiv preprint arXiv:2309.08600, 2023.arxiv.org/abs/2309.08600
  16. [16]Damai Dai, Li Dong, Yaru Hao, Zhifang Sui, Baobao Chang, and Furu Wei. Knowledge neurons in pretrained transformers. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8493–8502. Association for Computational Linguistics, 2022.
  17. [17]Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, et al. Toy models of superposition. arXiv preprint arXiv:2209.10652, 2022.arxiv.org/abs/2209.10652
  18. [18]Michael Fuest, Pingchuan Ma, Ming Gui, Johannes Schusterbauer, Vincent Tao Hu, and Björn Ommer. Diffusion models and representation learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024.arxiv.org/abs/2407.00783
  19. [19]Rohit Gandikota, Joanna Materzynska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion models. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2426–2436, 2023.arxiv.org/abs/2303.07345
  20. [20]Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5111–5120, 2024.arxiv.org/abs/2308.14761
  21. [21]Jiahui Gao, Renjie Pi, Tianyang Han, Han Wu, Lanqing Hong, Lingpeng Kong, Xin Jiang, and Zhenguo Li. Coca: Regaining safety-awareness of multimodal large language models with constitutional calibration. arXiv preprint arXiv:2409.11365, 2024.arxiv.org/abs/2409.11365
  22. [22]Lang Gao, Jiahui Geng, Xiangliang Zhang, Preslav Nakov, and Xiuying Chen. Shaping the safety boundaries: Understanding and defending against jailbreaks in large language models. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 25378–25398, 2025.arxiv.org/abs/2412.17034
  23. [23]Soumya Suvra Ghosal, Souradip Chakraborty, Vaibhav Singh, Tianrui Guan, Mengdi Wang, Ahmad Beirami, Furong Huang, Alvaro Velasquez, Dinesh Manocha, and Amrit Singh Bedi. Immune: Improving safety against jailbreaks in multi-modal llms via inference-time alignment. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 25038–25049, 2025.arxiv.org/abs/2411.18688
  24. [24]Chao Gong, Kai Chen, Zhipeng Wei, Jingjing Chen, and Yu-Gang Jiang. Reliable and efficient concept erasure of text-to-image diffusion models. In European Conference on Computer Vision, pages 73–88. Springer, 2024.arxiv.org/abs/2407.12383
  25. [25]Yichen Gong, Delong Ran, Jinyuan Liu, Conglei Wang, Tianshuo Cong, Anyu Wang, Sisi Duan, and Xiaoyun Wang. Figstep: Jailbreaking large vision-language models via typographic visual prompts. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 23951–23959, 2025.arxiv.org/abs/2311.05608
  26. [26]Yunhao Gou, Kai Chen, Zhili Liu, Lanqing Hong, Hang Xu, Zhenguo Li, Dit-Yan Yeung, James T Kwok, and Yu Zhang. Eyes closed, safety on: Protecting multimodal llms via image-to-text transformation. In European Conference on Computer Vision, pages 388–404. Springer, 2024.
  27. [27]Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024.arxiv.org/abs/2407.21783
  28. [28]Feng Han, Kai Chen, Chao Gong, Zhipeng Wei, Jingjing Chen, and Yu-Gang Jiang. Dumo: Dual encoder modulation network for precise concept erasure. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 3320–3328, 2025.arxiv.org/abs/2501.01125
  29. [29]Qin He, Jiaqi Weng, Jialing Tao, and Hui Xue. A single neuron works: Precise concept erasure in text-to-image diffusion models. arXiv preprint arXiv:2509.21008, 2025.
  30. [30]Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022.arxiv.org/abs/2208.01626
  31. [31]Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipsore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 conference on empirical methods in natural language processing, pages 7514–7528, 2021.
  32. [32]Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017.
  33. [33]Zhonghan Huang, Shanshan Zhong, Pan Zhou, Shanhua Gao, Marinka Zitnik, and Liang Lin. A causality-aware paradigm for evaluating creativity of multimodal large language models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025.arxiv.org/abs/2501.15147
  34. [34]Eshaan S Iyer, Peter Vitaro, Serena Wu, Jessie Muir, Yiu Chung Tse, Vedrana Cvetkovska, and Rosemary C Bagot. Reward integration in prefrontal-cortical and ventral-hippocampal nucleus accumbens inputs cooperatively modulates engagement. Nature Communications, 16(1):3573, 2025.DOI
  35. [35]Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. Mistral 7b, 2023.arxiv.org/abs/2310.06825
  36. [36]Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.arxiv.org/abs/1412.6980
  37. [37]Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22691–22702, 2023.
  38. [38]Black Forest Labs. black-forest-labs/flux github page, 2024.
  39. [39]Hao Li, Zhe Zhao, Shaofei Jiang, and Haitao Wu. Brain circuits that regulate social behavior. Molecular Psychiatry, pages 1–17, 2025.DOI
  40. [40]Meng Li, Guangda Huzhang, Haibo Zhang, Xiting Wang, and Anxiang Zeng. Optimal transport-based token weighting scheme for enhanced preference optimization. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 21311–21334, 2025.arxiv.org/abs/2505.18720
  41. [41]Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740–755. Springer, 2014.arxiv.org/abs/1405.0312
  42. [42]Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 26296–26306, 2024.
  43. [43]Xiaogeng Liu, Nan Xu, Muhao Chen, and Chaowei Xiao. Autodan: Generating stealthy jailbreak prompts on aligned large language models. arXiv preprint arXiv:2310.04451, 2023.arxiv.org/abs/2310.04451
  44. [44]Xin Liu, Yichen Zhu, Jindong Gu, Yunshi Lan, Chao Yang, and Yu Qiao. Mm-safetybench: A benchmark for safety evaluation of multimodal large language models. In European Conference on Computer Vision, pages 386–403. Springer, 2024.arxiv.org/abs/2311.17600
  45. [45]Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaq Yiang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European conference on computer vision, pages 216–233. Springer, 2024.arxiv.org/abs/2307.06281
  46. [46]Shilin Lu, Zilan Wang, Leyang Li, Yanzhu Liu, and Adams Wai-Kin Kong. Mace: Mass concept erasure in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6430–6440, 2024.arxiv.org/abs/2403.06135
  47. [47]Weidi Luo, Siyuan Ma, Xiaogeng Liu, Xiaoyu Guo, and Chaowei Xiao. Jailbreak: A benchmark for assessing the robustness of multimodal large language models against jailbreak attacks. arXiv preprint arXiv:2404.03027, 2024.arxiv.org/abs/2404.03027
  48. [48]Mengyao Lyu, Yuhong Yang, Haiwen Hong, Hui Chen, Xuan Jin, Yuan He, Hui Xue, Jungong Han, and Guiguang Ding. One-dimensional adapter to rule them all: Concepts diffusion models and erasing applications. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7559–7568, 2024.arxiv.org/abs/2312.16145
  49. [49]Zhiyuan Ma, Yuzhu Zhang, Guoli Jia, Liangliang Zhao, Yichao Ma, Mingjie Ma, Gaofeng Liu, Kaiyan Zhang, Ning Ding, Jianjun Li, et al. Efficient diffusion models: A comprehensive survey from principles to practices. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025.DOI
  50. [50]Mantas Mazeika, Long Phan, Xuwang Yin, Andy Zou, Zifan Wang, Norman Mu, Elham Sakhaee, Nathaniel Li, Steven Basart, Bo Li, et al. Harmbench: A standardized evaluation framework for automated red teaming and robust refusal. arXiv preprint arXiv:2402.04249, 2024.arxiv.org/abs/2402.04249
  51. [51]Neel Nanda, Lawrence Chan, Tom Lieberum, Jess Smith, and Jacob Steinhardt. Progress measures for grokking via mechanistic interpretability. arXiv preprint arXiv:2301.05217, 2023.arxiv.org/abs/2301.05217
  52. [52]Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730–27744, 2022.arxiv.org/abs/2203.02155
  53. [53]Gabriel Peyré and Marco Cuturi. Computational optimal transport: With applications to data science. Now Foundations and Trends, 2019.
  54. [54]Minh Pham, Kelly O Marshall, Niv Cohen, Govind Mittal, and Chinmay Hegde. Circumventing concept erasure methods for text-to-image generative models. arXiv preprint arXiv:2308.01508, 2023.arxiv.org/abs/2308.01508
  55. [55]Xiangyu Qi, Ashwinee Panda, Kaifeng Lyu, Xiao Ma, Subhrajit Roy, Ahmad Beirami, Prateek Mittal, and Peter Henderson. Safety alignment should be made more than just a few tokens deep. arXiv preprint arXiv:2406.05946, 2024.arxiv.org/abs/2406.05946
  56. [56]Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, Proceedings of Machine Learning Research, pages 8748–8763. PMLR, 2021.arxiv.org/abs/2103.00020
  57. [57]Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in neural information processing systems, 36:53728–53741, 2023.arxiv.org/abs/2305.18290
  58. [58]Nina Rimsky, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, and Alexander Turner. Steering llama 2 via contrastive activation addition. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15504–15522. Association for Computational Linguistics, 2024.arxiv.org/abs/2312.06681
  59. [59]Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022.arxiv.org/abs/2112.10752
  60. [60]Paul Röttger, Hannah Kirk, Bertie Vidgen, Giuseppe Attanasio, Federico Bianchi, and Dirk Hovy. Axstest: A test suite for identifying exaggerated safety behaviours in large language models. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 5377–5400, 2024.
  61. [61]Shaswati Saha, Sourajit Saha, Manas Gaur, and Tejas Gokhale. Side effects of erasing concepts from diffusion models. arXiv preprint arXiv:2508.15124, 2025.arxiv.org/abs/2508.15124
  62. [62]Patrick Schramowski, Manuel Brack, Björn Deisroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22522–22531, 2023.arxiv.org/abs/2211.05105
  63. [63]Patrick Schramowski, Christopher Tauchmann, and Kristian Kersting. Can machines help us answering question 16 in datasheets, and in turn reflecting on inappropriate content? In Proceedings of the 2022 ACM conference on fairness, accountability, and transparency, pages 1350–1361, 2022.arxiv.org/abs/2202.06675
  64. [64]Shaan Shah and Meenakshi Khosla. Representational alignment across model layers and brain regions with multi-level optimal transport. In The Fourteenth International Conference on Learning Representations.
  65. [65]Fei Shen, Hu Ye, Jun Zhang, Cong Wang, Xiao Han, and Wei Yang. Advancing pose-guided image synthesis with progressive conditional diffusion models. arXiv preprint arXiv:2310.06313, 2023.arxiv.org/abs/2310.06313
  66. [66]Chuanheng Shi, Shanzhe Li, Shiming Guo, Simiao Xie, Wenhua Wu, Jingtong Dou, Chao Wu, Canran Xiao, Cong Wang, Zifeng Cheng, et al. Where culture fades: Revealing the cultural gap in text-to-image generation. arXiv preprint arXiv:2511.17282, 2025.arxiv.org/abs/2511.17282
  67. [67]Chuancheng Shi, Shangze Li, Wenjun Lu, Wenhua Wu, Cong Wang, Zifeng Cheng, Fei Shen, and Tat-Seng Chua. TraceRouter: Robust safety for large foundation models via path-level intervention. arXiv preprint arXiv:2601.21900, 2026.arxiv.org/abs/2601.21900
  68. [68]Shivalika Singh, Angelika Romanou, Clémentine Fourrier, David Ife-oluwa Adelani, Jian Gang Ngui, Daniel Vila-Suero, Peerat Limkonchotiwat, Kelly Marchisio, Wei Qi Leong, Yosephine Susanto, et al. Global mmlu: Understanding and addressing cultural and linguistic biases in multilingual evaluation. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 18761–18799, 2025.
  69. [69]Marina Sokolova and Guy Lapalme. A systematic analysis of performance measures for classification tasks. Information processing & management, 45(4):427–437, 2009.DOI
  70. [70]Viacheslav Surkov, Chris Wendler, Antonio Mari, Mikhail Terekhov, Justin Deschenaux, Robert West, Caglar Gulcehre, and David Bau. One-step is enough: Sparse autoencoders for text-to-image diffusion models. arXiv preprint arXiv:2410.22366, 2025.arxiv.org/abs/2410.22366
  71. [71]Harrish Thasarathan, Julian Forsyth, Thomas Fel, Matthew Kowal, and Konstantinos G. Derpanis. Universal sparse autoencoders: Interpretable cross-model concept alignment. In Artificial Intelligence and Machine Learning, pages 59304–59325. PMLR, 13–19 Jul 2025.arxiv.org/abs/2502.03714
  72. [72]Vayer Titouan, Nicolas Courtay, Romain Tavenard, and Rémi Flamary. Optimal transport for structured data with application on graphs. In International Conference on Machine Learning, pages 6275–6284. PMLR, 2019.
  73. [73]Yu-Lin Tsai, Chia-Yi Hsu, Chulin Xie, Chih-Hsun Lin, Jia-You Chen, Bo Li, Pin-Yu Chen, Chia-Mu Yu, and Chun-Ying Huang. Ring-a-bell! how reliable are concept removal methods for diffusion models? arXiv preprint arXiv:2310.10012, 2023.arxiv.org/abs/2310.10012
  74. [74]Hongbo Wang, MaungMaung AprilPyone, and Isao Echizen. Sgm: Safety glasses for multimodal large language models via neuron-level detoxification. arXiv preprint arXiv:2512.15052, 2025.arxiv.org/abs/2512.15052
  75. [75]Yifan Wang, Yifei Liu, Yingdong Shi, Changming Li, Anqi Pang, Sibei Yang, Jingyi Yu, and Kan Ren. Discovering influential neuron path in vision transformers. arXiv preprint arXiv:2503.09046, 2025.arxiv.org/abs/2503.09046
  76. [76]Yu Wang, Xiaogeng Liu, Yu Li, Muhao Chen, and Chaowei Xiao. Adashield: Safeguarding multimodal large language models from structure-based attack via adaptive shield prompting. In European Conference on Computer Vision, pages 77–94. Springer, 2024.arxiv.org/abs/2403.09513
  77. [77]Alexander Wei, Nika Haghtalab, and Jacob Steinhardt. Jailbroken: How does llm safety training fail? Advances in neural information processing systems, 36:80079–80110, 2023.arxiv.org/abs/2307.02483
  78. [78]Yao Wei, Martin Renqiang Min, George Vosselman, Li Erran Li, and Michael Ying Yang. Llm-enhanced graph prior meets 3d indoor scene explicit regularization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025.DOI
  79. [79]Jianlong Wu, Wei Liu, Ye Liu, Meng Liu, Liqiang Nie, Zhoucheng Lin, and Chang Wen Chen. A survey on video temporal grounding with multimodal large language model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025.arxiv.org/abs/2508.10922
  80. [80]Zhiyu Wu, Xiaokang Chen, Zizheng Pan, Xingchao Liu, Wen Liu, Damai Dai, Huazuo Gao, Yiyang Ma, Chengyue Wu, Bingxuan Wang, Zhenda Xie, Yu Wu, Kai Hu, Jiawei Wang, Yaofeng Sun, Yukun Li, Yishi Piao, Kang Guan, Aixin Liu, Xin Xie, Yuxiang You, Kai Dong, Xingkai Yu, Haowei Zhang, Liang Zhao, Yisong Wang, and Chong Ruan. DeepSeek-VL2: Mixture-of-experts vision-language models for advanced multimodal understanding. arXiv preprint arXiv:2412.10302, 2024.arxiv.org/abs/2412.10302
  81. [81]Zihui Wu, Haichang Gao, Jiacheng Luo, and Zhaoxiang Liu. Humorrect: Decoupling llm safety from refusal prefix via a little humor. arXiv preprint arXiv:2501.13677, 2025.arxiv.org/abs/2501.13677
  82. [82]Kun Xiang, Zhili Liu, Terry Jingchen Zhang, Yinya Huang, Yunshuang Nie, Kaixin Cai, Yiyang Yin, Runhui Huang, Hanhui Li, Yihan Zeng, et al. Atomthink: Multimodal slow thinking with atomic step reasoning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2026.DOI
  83. [83]Yang Xiao, Wang Lu, Jie Ji, Ruimeng Ye, Gen Li, Xiaolong Ma, and Bo Hui. Optimal transport for brain-image alignment: Unveiling redundancy and synergy in neural information processing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20445–20455, 2025.arxiv.org/abs/2503.10663
  84. [84]Jinheng Xie, Zhenheng Yang, and Mike Zheng Shou. Show-o2: Improved native unified multimodal models. arXiv preprint arXiv:2506.15564, 2025.arxiv.org/abs/2506.15564
  85. [85]Hongteng Xu, Dixin Luo, Hongyuan Zha, and Lawrence Carin Duke. Gromov-wasserstein learning for graph matching and node embedding. In International conference on machine learning, pages 6932–6941. PMLR, 2019.arxiv.org/abs/1901.06003
  86. [86]Fulong Ye, Guang Liu, Xinya Wu, and Ledell Wu. Altdiffusion: A multilingual text-to-image diffusion model. In Proceedings of the AAAI conference on artificial intelligence, volume 38, pages 6468–6656, 2024.arxiv.org/abs/2308.09991
  87. [87]Youliang Yuan, Wenxiang Jiao, Wenxuan Wang, Jen-tse Huang, Jiahao Xu, Tian Liang, Pinjia He, and Zhaopeng Tu. Refuse whenever you feel unsafe: Improving safety in llms via decoupled refusal training. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3149–3167, 2025.
  88. [88]Duzhen Zhang, Zhong-Zhi Li, Ming-Liang Zhang, Jiaxin Zhang, Zengyan Liu, Yuxuan Yao, Haotian Xu, Junhao Zheng, Xiuyi Chen, Yingying Zhang, et al. From system 1 to system 2: a survey of reasoning large language models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025.
  89. [89]Jiachen Zhao, Jing Huang, Zhengxuan Wu, David Bau, and Weiyan Shi. Llms encode harmfulness and refusal separately. Advances in Neural Information Processing Systems, 38:140283–140318, 2026.arxiv.org/abs/2507.11878
  90. [90]Yiran Zhao, Wenxuan Zhang, Yuxi Xie, Anirudh Goyal, Kenji Kawaguchi, and Michael Shieh. Understanding and enhancing safety mechanisms of llms via safety-specific neuron. In The Thirteenth International Conference on Learning Representations, 2025.
  91. [91]Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed El-hoseiny. Minigt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023.arxiv.org/abs/2304.10592
  92. [92]Lanyun Zhu, Tianrun Chen, Deyi Ji, Peng Xu, Jieping Ye, and Jun Liu. Llafs++: Few-shot image segmentation with large language models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025.DOI
  93. [93]Andy Zou, Long Phan, Sarah Chen, James Campbell, Phillip Guo, Richard Ren, Alexander Pan, Xuwang Yin, Mantas Mazeika, Ann-Katrin Dombrowski, et al. Representational engineering: A top-down approach to ai transparency. arXiv preprint arXiv:2310.01405, 2023.arxiv.org/abs/2310.01405
  94. [94]Andy Zou, Long Phan, Justin Wang, Derek Duenas, Maxwell Lin, Maksym Andriushchenko, Rowan Wang, Zico Kolter, Matt Fredrikson, and Dan Hendrycks. Improving alignment and robustness with circuit breakers. Advances in Neural Information Processing Systems, 37:83345–83373, 2024.arxiv.org/abs/2406.04313
  95. [95]Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J Zico Kolter, and Matt Fredrikson. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043, 2023.arxiv.org/abs/2307.15043

Paper details

vlnc:W_yzrhr67q
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

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