This paper introduces MemJack, a memory-augmented multi-agent framework designed to systematically expose visual-semantic vulnerabilities in Vision-Language Models (VLMs). It orchestrates automated jailbreak attacks using unmodified natural images by dynamically mapping visual entities to malicious intents and leverages a persistent memory to transfer successful strategies across different images.

This paper introduces a novel deep reinforcement learning framework for safe multi-agent control in highway merging scenarios, integrating partial attention mechanisms into a QMIX architecture. It proposes both spatial and temporal attention to focus on relevant neighboring vehicles and their historical states, combined with a comprehensive reward signal to balance global traffic objectives and individual agent interests. The approach demonstrates significant improvements in safety, driving speed, and overall reward compared to baseline models in SUMO simulations.

This paper proposes viewing large language model (LLM) teams through the lens of distributed systems to create a principled framework for their design and evaluation. It reveals that many established advantages and challenges from distributed computing, such as scalability limits and coordination issues, directly apply to and explain the behavior of LLM teams.

This paper introduces an active inference computational framework for empathy in AI agents, enabling explicit perspective-taking through a self-other model transformation. It demonstrates that empathic perspective-taking can induce robust cooperation in strategic dilemmas like the Iterated Prisoner's Dilemma, highlighting empathy as a structural prior for socially aligned AI.