The field of large language models is moving towards improving safety and control, with a focus on developing innovative methods to mitigate potential risks and vulnerabilities. Recent research has explored various approaches, including multimodal prompt decoupling attacks, adaptive subspace steering, and backdoor attribution, to name a few. These advancements aim to enhance the reliability and trustworthiness of large language models, ensuring their safe deployment in real-world applications. Noteworthy papers in this area include Multimodal Prompt Decoupling Attack, which proposes a novel attack method to bypass safety filters, and Backdoor Attribution, which introduces a framework to elucidate and control backdoor mechanisms in language models. Additionally, papers like SafeSteer and ASGuard have made significant contributions to the development of efficient defense mechanisms against jailbreak attacks.
Advances in Safety and Control of Large Language Models
Sources
Multimodal Prompt Decoupling Attack on the Safety Filters in Text-to-Image Models
SafeSteer: Adaptive Subspace Steering for Efficient Jailbreak Defense in Vision-Language Models
JaiLIP: Jailbreaking Vision-Language Models via Loss Guided Image Perturbation
Backdoor Attribution: Elucidating and Controlling Backdoor in Language Models
The Rogue Scalpel: Activation Steering Compromises LLM Safety
Jailbreaking on Text-to-Video Models via Scene Splitting Strategy
Enhancing LLM Steering through Sparse Autoencoder-Based Vector Refinement
Toward Preference-aligned Large Language Models via Residual-based Model Steering
TokenSwap: Backdoor Attack on the Compositional Understanding of Large Vision-Language Models
EasySteer: A Unified Framework for High-Performance and Extensible LLM Steering
VISOR++: Universal Visual Inputs based Steering for Large Vision Language Models
ASGuard: Activation-Scaling Guard to Mitigate Targeted Jailbreaking Attack
Alignment-Aware Decoding
Reinforcement Learning-Based Prompt Template Stealing for Text-to-Image Models
Drones that Think on their Feet: Sudden Landing Decisions with Embodied AI
CHAI: Command Hijacking against embodied AI
Backdoor Attacks Against Speech Language Models
A Comparative Analysis of Sparse Autoencoder and Activation Difference in Language Model Steering
Towards Interpretable and Inference-Optimal COT Reasoning with Sparse Autoencoder-Guided Generation