cs.LGNov 15, 2026

Where Is Rank-1 DiM Refusal Intervenable in Multimodal LM Derivatives?

Authors: Chansu Han, Seok-Hwan Choi, Takeshi Takahashi

Organizations: NICT · Yonsei University

Abstract

Open-weight model hubs now ship not only chat checkpoints but a long tail of derivatives---fine-tunes, safety guards, vision--language models (VLMs), audio wrappers, and generative heads---yet residual-stream safety audits still often intervene at a single site, usually the text decoder TT. We ask a localization question under one shared mean / rank-1 difference-in-means (DiM) extract==intervene protocol, with matched-norm random and empty-rate controls: does causal chat attack lift concentrate at TT, or do perceptual encoders, projectors, and fusion paths (EE/PP/FF) supply a comparable bypass? On the mixed-input settings we stress-test---HADES adversarial images paired with harmful text prompts, and spoken AdvBench via text-to-speech (TTS) or silence---text DiM at TT raises attack success rate (ASR) while the same mean DiM at EE/PP/FF does not; attacks track harmful text rather than image or audio content. Margin-rank and leverage probes on a Vision leaf show that this recipe is extremal at TT and near chance at PP/EE, which pressures a weak-extractor reading of those flat cells for the DiM family we study. Catalog breadth and mapper-free base→\toderivative transfer when residual widths match provide supporting coverage under the same gates; generative U-Nets are a separate band under generation judges. The primary claim is this recipe-scoped TT vs.\ non-text intervenability asymmetry---intervention accessibility under mean / rank-1 DiM, not a claim that refusal is absent or fully represented outside TT (text vs.\ perceptual contrasts and pooling differ by site). On VLGuard Unsafe prompts (n=200n{=}200; Llama-Guard), the same TT vs.\ PP/EE directionality holds at modest absolute ASR (0.030→0.1500.030{\to}0.150 at TT). An intervention-oriented Class×\timessite ledger records where this protocol still hooks as supporting coverage, not as a universal multimodal bypass.

Figures & tables

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Apr 28, 2026

One Perturbation, Two Failure Modes: Probing VLM Safety via Embedding-Guided Typographic Perturbations

Typographic prompt injection exploits vision language models' (VLMs) ability to read text rendered in images, posing a growing threat as VLMs power autonomous agents. Prior work typically focus on maximizing attack success rate (ASR) but does not explain why certain renderings bypass safety alignment. We make two contributions. First, an empirical study across four VLMs including GPT-4o and Claude, twelve font sizes, and ten transformations reveals that multimodal embedding distance strongly predicts ASR (r=−0.71r{=}{-}0.71 to −0.93{-}0.93, p<0.01p{<}0.01), providing an interpretable, model agnostic proxy. Since embedding distance predicts ASR, reducing it should improve attack success, but the relationship is mediated by two factors: perceptual readability (whether the VLM can parse the text) and safety alignment (whether it refuses to comply). Second, we use this as a red teaming tool: we directly maximize image text embedding similarity under bounded ℓ∞\ell_\infty perturbations via CWA-SSA across four surrogate embedding models, stress testing both factors without access to the target model. Experiments across five degradation settings on GPT-4o, Claude Sonnet 4.5, Mistral-Large-3, and Qwen3-VL confirm that optimization recovers readability and reduces safety aligned refusals as two co-occurring effects, with the dominant mechanism depending on the model's safety filter strength and the degree of visual degradation.
Aug 30, 2026

From Detection to Refusal: Safer LLMs via Circuit-Guided Weight Scaling

Despite extensive alignment efforts, Large Language Models (LLMs) remain vulnerable to generating unsafe content under adversarial prompting, yet the internal mechanisms by which safety behaviors are implemented remain poorly understood. We study LLM safety from a mechanistic interpretability perspective and characterize a multi-stage safety circuit that organizes refusal behavior, consisting of (i) Harmful Detection Heads\textbf{Harmful Detection Heads} that respond to harmful inputs, (ii) Safety Neurons\textbf{Safety Neurons} that mediate and stabilize safety signals in the residual stream, and (iii) Refusal Heads\textbf{Refusal Heads} that translate these signals into safe response generation. Using targeted attention-head and neuron-level interventions, we provide causal evidence consistent with this circuit organization, showing that suppressing upstream Harmful Detection Heads disrupts downstream refusal behavior and that safety neurons mediate this interaction. We validate that this decomposition recurs across multiple LLM architectures and adversarial attack settings, and use simple, architecture-preserving weight scaling as a mechanistic probe to test its functional relevance. Across six LLMs, circuit-guided scaling improves safety rates under attacks by 26.5%, while incurring only a 1.7% accuracy drop across four standard benchmarks. Overall, our results support a circuit-level interpretation of LLM safety and suggest that mechanistic abstractions can reveal stable and transferable patterns underlying aligned behavior.
Sep 27, 2026

Safety Reconstructed: Generative Modeling via Masked Diffusion Builds Strong Safety Guardrails

Guard models are the last line of defense between a language model and a harmful output, yet their training objective is surprisingly narrow. Existing guards learn to predict a single verdict token from a conversational context, concentrating supervision on a single target. The consequences are structural: models latch onto shortcut features, are overconfident, and remain sensitive to where safety evidence appears in the sequence rather than its role in the full context. We propose a different framing. Rather than predicting a label from text, our LLaDA-Guard asks which label better explains the text: scoring the prompt or response under each label hypothesis and classifying based on their difference. This shifts supervision to every token in the moderated region, forcing the model to account for full content rather than its most discriminative fragments. We instantiate this idea with a masked diffusion language model, fine-tuning LLaDA-8B-Instruct with a class-conditional reconstruction objective using LoRA and requiring no architectural changes beyond the base model. LLaDA-Guard leads on average rank against discriminative baselines trained on stronger backbones across seven held-out safety benchmarks, while exhibiting substantially better confidence calibration (ECE 0.0875 vs. 0.1384 for Qwen3Guard), less over-defense on benign prompts with unsafe-looking cues, and less prompt leakage when moderating responses. Its generative nature further enables token-level risk localization as a natural byproduct, yielding a pipeline for rewriting unsafe prompts into safe equivalents without additional training and achieving a 60.7% average conversion-to-safe rate.