When Behavioral Safety Evaluation Fails: A Representation-Level Perspective

• URL: http://arxiv.org/abs/2606.08044v1
• Date: Sat, 06 Jun 2026 08:10:56 GMT
• Title: When Behavioral Safety Evaluation Fails: A Representation-Level Perspective
• Abstract summary: Large Language Model (LLM) safety has often been evaluated at the behavior level, which provides limited evidence of internal robustness.<n>We formalize this discrepancy as the audit gap: the difference between behavioral safety and robustness under intervention.<n>We introduce an intervention-based evaluation framework to test model robustness through soft interventions in parameter and latent spaces.
• Score: 23.598318016787147
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Large Language Model (LLM) safety has often been evaluated at the behavior level, which provides limited evidence of internal robustness, as these evaluations target outputs rather than representation-level vulnerability under intervention. We formalize this discrepancy as the audit gap: the difference between behavioral safety and robustness under intervention. To study this gap, we construct dissociated models that preserve safe outward behavior while remaining vulnerable in the latent space. We introduce an intervention-based evaluation framework to test model robustness through soft interventions in parameter and latent spaces, including harmful fine-tuning and layer-wise latent perturbations. To formalize the evaluation, we propose the Latent Vulnerability Score (LVS) to measure how easily harmful behavior can be elicited by bounded latent perturbations. Using this evaluation framework, we show that behavioral safety metrics are insufficient measures of representation-level robustness across multiple safely and unsafely aligned state-of-the-art models. Notably, dissociated models show substantially elevated LVSs despite comparable refusal behavior under harmful intervention, with intermediate representations being the most sensitive to intervention. Our results suggest that behavioral safety evaluation alone provides an incomplete picture of model robustness, motivating representation-aware audits of latent vulnerability and observable behavior.

Related papers

• When Confidence Lacks Concepts: Interpretable OOD Detection via Representation Perturbations [6.4945321478117455]
  We propose an interpretable OOD detection framework that probes the stability of model predictions under class-conditioned semantic perturbations.<n>By framing OOD detection as a concept conditioned stability analysis, our approach provides both a discriminative OOD signal and an interpretable lens into the internal mechanisms driving model uncertainty.
  arXiv  Detail & Related papers  (2026-06-15T04:11:29Z)
• RefusalGuard: Geometry-Preserving Fine-Tuning for Safety in LLMs [5.100622189286672]
  Fine-tuning safety-aligned language models for downstream tasks can lead to substantial degradation of refusal behavior.<n>We introduce REFUSALGUARD, a representation-level fine-tuning framework that preserves safety-relevant structure during model adaptation.
  arXiv  Detail & Related papers  (2026-05-03T14:48:18Z)
• OOD-MMSafe: Advancing MLLM Safety from Harmful Intent to Hidden Consequences [64.01706941950489]
  Current safety paradigms primarily target malicious intent or situational violations.<n>We propose shifting the safety frontier toward consequence-driven safety, a paradigm essential for the robust deployment of autonomous and embodied agents.<n>We develop the Consequence-Aware Safety Policy Optimization (CASPO) framework, which integrates the model's intrinsic reasoning as a dynamic reference for token-level self-distillation rewards.
  arXiv  Detail & Related papers  (2026-03-10T14:16:43Z)
• When Evaluation Becomes a Side Channel: Regime Leakage and Structural Mitigations for Alignment Assessment [0.0]
  Safety evaluation for advanced AI systems assumes that behavior observed under evaluation predicts behavior in deployment.<n>We recast alignment evaluation as a problem of information flow under partial observability.<n>We study regime-blind mechanisms, training-time interventions that restrict access to regime cues.
  arXiv  Detail & Related papers  (2026-02-09T10:00:24Z)
• Steering Externalities: Benign Activation Steering Unintentionally Increases Jailbreak Risk for Large Language Models [62.16655896700062]
  Activation steering is a technique to enhance the utility of Large Language Models (LLMs)<n>We show that it unintentionally introduces critical and under-explored safety risks.<n>Experiments reveal that these interventions act as a force multiplier, creating new vulnerabilities to jailbreaks and increasing attack success rates to over 80% on standard benchmarks.
  arXiv  Detail & Related papers  (2026-02-03T12:32:35Z)
• Character as a Latent Variable in Large Language Models: A Mechanistic Account of Emergent Misalignment and Conditional Safety Failures [70.48661957773449]
  Emergent Misalignment refers to a failure mode in which fine-tuning large language models on narrowly scoped data induces broadly misaligned behavior.<n>Across multiple domains and model families, we find that fine-tuning models on data exhibiting specific character-level dispositions induces substantially stronger and more transferable misalignment than incorrect-advice fine-tuning.
  arXiv  Detail & Related papers  (2026-01-30T15:28:42Z)
• SafeRBench: A Comprehensive Benchmark for Safety Assessment in Large Reasoning Models [60.8821834954637]
  We present SafeRBench, the first benchmark that assesses LRM safety end-to-end.<n>We pioneer the incorporation of risk categories and levels into input design.<n>We introduce a micro-thought chunking mechanism to segment long reasoning traces into semantically coherent units.
  arXiv  Detail & Related papers  (2025-11-19T06:46:33Z)
• Embedding Poisoning: Bypassing Safety Alignment via Embedding Semantic Shift [23.0914017433021]
  This work identifies a novel class of deployment phase attacks that exploit a vulnerability by injecting imperceptible perturbations directly into the embedding layer outputs without modifying model weights or input text.<n>We propose Search based Embedding Poisoning, a practical, model agnostic framework that introduces carefully optimized perturbations into embeddings associated with high risk tokens.
  arXiv  Detail & Related papers  (2025-09-08T05:00:58Z)
• Probing the Robustness of Large Language Models Safety to Latent Perturbations [30.16804362984161]
  Safety alignment is a key requirement for building reliable Artificial General Intelligence.<n>We observe that minor latent shifts can still trigger unsafe responses in aligned models.<n>We introduce Layer-wise Adversarial Patch Training(LAPT), a fine-tuning strategy that injects controlled perturbations into hidden representations during training.
  arXiv  Detail & Related papers  (2025-06-19T07:03:05Z)
• Extreme Miscalibration and the Illusion of Adversarial Robustness [66.29268991629085]
  Adversarial Training is often used to increase model robustness. We show that this observed gain in robustness is an illusion of robustness (IOR) We urge the NLP community to incorporate test-time temperature scaling into their robustness evaluations.
  arXiv  Detail & Related papers  (2024-02-27T13:49:12Z)
• Exploring Robustness of Unsupervised Domain Adaptation in Semantic Segmentation [74.05906222376608]
  We propose adversarial self-supervision UDA (or ASSUDA) that maximizes the agreement between clean images and their adversarial examples by a contrastive loss in the output space. This paper is rooted in two observations: (i) the robustness of UDA methods in semantic segmentation remains unexplored, which pose a security concern in this field; and (ii) although commonly used self-supervision (e.g., rotation and jigsaw) benefits image tasks such as classification and recognition, they fail to provide the critical supervision signals that could learn discriminative representation for segmentation tasks.
  arXiv  Detail & Related papers  (2021-05-23T01:50:44Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.