Two new research papers explore vulnerabilities and defenses in large language models (LLMs) and large vision-language models (LVLMs). The first paper introduces Robustness of Prompting (RoP), a strategy designed to enhance LLM resilience against adversarial perturbations by correcting input errors and generating optimal guidance prompts. The second paper details a Multi-Turn Adaptive Prompting Attack (MAPA) that targets LVLMs by alternating text-vision attacks and iteratively refining the attack trajectory to amplify malicious responses, outperforming existing methods on several benchmarks. AI
IMPACT New research highlights vulnerabilities in LLMs and LVLMs, suggesting a need for more robust prompting strategies and defenses against sophisticated attacks.
RANK_REASON Two academic papers published on arXiv detailing new methods for LLM robustness and LVLM attacks.
- GPT-4o-mini
- Large Language Models
- Large Vision-Language Models
- Llama-3.2-Vision-11B-Instruct
- LLaVA-v1.6-Mistral-7B
- Multi-Turn Adaptive Prompting Attack
- Qwen2.5-VL-7B-Instruct
- Robustness of Prompting
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
