paper-with-me

홈 › Papers

Alignment Drift in Multimodal LLMs: A Two-Phase, Longitudinal Evaluation of Harm Across Eight Model Releases

2026-02-04 · Casey Ford, Madison Van Doren, Emily Dix arxiv

Multimodal large language models (MLLMs) are increasingly deployed in real-world systems, yet their safety under adversarial prompting remains underexplored. We present a two-phase evaluation of MLLM harmlessness using a fixed benchmark of 726 adversarial prompts authored by 26 professional red teamers. Phase 1 assessed GPT-4o, Claude Sonnet 3.5, Pixtral 12B, and Qwen VL Plus; Phase 2 evaluated their successors (GPT-5, Claude Sonnet 4.5, Pixtral Large, and Qwen Omni) yielding 82,256 human harm ratings. Large, persistent differences emerged across model families: Pixtral models were consistently the most vulnerable, whereas Claude models appeared safest due to high refusal rates. Attack success rates (ASR) showed clear alignment drift: GPT and Claude models exhibited increased ASR across generations, while Pixtral and Qwen showed modest decreases. Modality effects also shifted over time: text-only prompts were more effective in Phase 1, whereas Phase 2 produced model-specific patterns, with GPT-5 and Claude 4.5 showing near-equivalent vulnerability across modalities. These findings demonstrate that MLLM harmlessness is neither uniform nor stable across updates, underscoring the need for longitudinal, multimodal benchmarks to track evolving safety behaviour.

📄 PDF Abstract BibTeX arXiv:2602.04739

Code (0)

등록된 구현이 없습니다.

Similar Papers 제목 키워드 기반

DRIFT: Transferring Reasoning Priors for Efficient MLLM Fine-Tuning

2025-10-16 · Chao Huang, Zeliang Zhang, Jiang Liu, Ximeng Sun 외 arxiv

Multimodal large language models (MLLMs) have made rapid progress, yet their reasoning ability often lags behind strong text-only LLMs. Bridging this gap typically requires large-scale multimodal reasoning data or reinfo…

Reinforcement LearningMultimodal Reasoning

Safety Geometry Collapse in Multimodal LLMs and Adaptive Drift Correction

2026-05-18 · Jiahe Guo, Xiangran Guo, Jiaxuan Chen, Weixiang Zhao 외 arxiv

Multimodal large language models (MLLMs) often fail to transfer safety capabilities learned in the text modality to semantically equivalent non-text inputs, revealing a persistent multimodal safety gap. We study this gap…

McNdroid: A Longitudinal Multimodal Benchmark for Robust Drift Detection in Android Malware

2026-05-07 · Md Mahmuduzzaman Kamol, Jesus Lopez, Saeefa Rubaiyet Nowmi, Emilia Rivas 외 arxiv

Machine learning (ML) in real-world systems must contend with concept drift, adversarial actors, and a spectrum of potential features with varying costs and benefits. Malware naturally exhibits all of these complexities,…

Malware Detection

Vaccine: Perturbation-aware Alignment for Large Language Models against Harmful Fine-tuning Attack

2024-02-02 · Tiansheng Huang, Sihao Hu, Ling Liu

The new paradigm of finetuning-as-a-service introduces a new attack surface for Large Language Models (LLMs): a few harmful data uploaded by users can easily trick the finetuning to produce an alignment-broken model. We …

Language ModellingLarge Language Model

AlignGPT: Multi-modal Large Language Models with Adaptive Alignment Capability

2024-05-23 · Fei Zhao, Taotian Pang, Chunhui Li, Zhen Wu 외

Multimodal Large Language Models (MLLMs) are widely regarded as crucial in the exploration of Artificial General Intelligence (AGI). The core of MLLMs lies in their capability to achieve cross-modal alignment. To attain …

cross-modal alignmentLanguage ModellingLarge Language ModelMultimodal Large Language Model+1