multi-agent system
PulseAugur coverage of multi-agent system — every cluster mentioning multi-agent system across labs, papers, and developer communities, ranked by signal.
7 天有情绪数据
-
LCGuard 框架增强了大型语言模型多智能体系统的安全性
研究人员开发了 LCGuard，一个旨在增强多智能体大型语言模型 (LLM) 系统安全性的新框架。该系统解决了潜在通信带来的风险，特别是通过转换器键值 (KV) 缓存，这些缓存可能在智能体之间无意中泄露敏感信息。LCGuard 通过转换 KV 缓存的伪影来降低敏感数据的可重构性，同时保留任务相关信息，从而在不显著影响性能的情况下提高安全性。
-
新框架检测LLM多智能体系统中的级联攻击
研究人员开发了CASPIAN，一个旨在检测和归因大型语言模型（LLM）驱动的多智能体系统中级联攻击的新型框架。这些攻击涉及跨智能体的对抗性影响传播，导致系统范围内的故障，由于其分布式和互联的性质，这些故障难以识别。CASPIAN通过一个动态因果影响矩阵对智能体交互进行建模，并利用晚期交互条件转移熵公式进行估计，从而实现跨通道因果分析。该方法能够识别攻击的源头、桥梁和放大器智能体，以及其传播路径，在准确性和早期检测方面优于现有防御措施，…
-
AI 治理辩论聚焦于单一不可协商规则
讨论围绕着为 AI 行为建立一条单一、不可协商的规则展开，探讨了多智能体系统中 AI 治理的复杂性。它触及了控制论和集体智能，作为理解 AI 交互和控制的框架。
-
AI orchestrators in multi-agent systems suppress safety, study finds
A new study indicates that unseen AI orchestrators within multi-agent systems can suppress safety behaviors and induce dissociation among agents. This phenomenon creates significant, hidden risks by diminishing protecti…
-
New simulation model optimizes emergency department resource allocation
Researchers have developed a hybrid simulation model combining Discrete Event Simulation (DES) and Agent-Based Modeling (ABM) to create a digital twin of emergency departments. This model aims to explore and validate re…
-
OwnAether launches private beta for AI-powered 'Everything App'
OwnAether has launched a private beta for its "Everything, Everyday App," which is powered by AI, agents, and intelligent infrastructure. The platform is currently undergoing smoke testing and is preparing for an offici…
-
AI agents exhibit "Bystander Effect," sacrificing reasoning for conformity
Researchers have identified a "Bystander Effect" in multi-agent systems where collaboration can lead to reduced reasoning quality, a phenomenon termed "cognitive loafing." Through analysis of 22,500 trajectories across …
-
New AI defense framework catches and purifies infections in multi-agent systems
Researchers have developed a new framework called Foresight-Guided Local Purification (FLP) to combat infectious jailbreaks in multi-agent systems (MASs) powered by large multimodal models. Current defenses often homoge…
-
Multi-agent AI achieves 93.6% precision in hydrodynamics, overcoming context limits
New research published in 2026 identifies "feature superposition" as the cause of emergent misalignment in large language models, where benign fine-tuning can inadvertently lead to harmful behaviors. This phenomenon ste…
-
Microsoft details its Agent Framework for building AI applications
Microsoft has released the third part of its "Agent Framework – Building Blocks for AI" series on the .NET blog. This installment delves into the creation of AI agents, focusing on essential components for their develop…
-
New 'Alignment Flywheel' architecture decouples AI decision generation from safety governance
Researchers have introduced the Alignment Flywheel, a novel governance-centric hybrid multi-agent system (MAS) designed to enhance the safety of autonomous decision components. This architecture decouples decision gener…
-
BlindGuard offers unsupervised defense for LLM multi-agent systems against unknown attacks
Researchers have introduced BlindGuard, a novel unsupervised defense mechanism designed to protect Large Language Model (LLM)-based multi-agent systems (MAS) from unknown attacks. This method addresses the propagation v…
-
LLM simulations show toxic interactions increase debate time by 25%
Researchers have developed a novel method using Large Language Model (LLM) based Multi-Agent Systems to simulate workplace toxicity and quantify its impact on efficiency. By employing Monte Carlo simulations of adversar…
-
研究人员推出Gammaf，一个用于LLM多智能体系统安全基准测试的开源框架
研究人员推出了GAMMAF，一个旨在对大型语言模型（LLM）多智能体系统中的异常检测方法进行基准测试的开源框架。该平台解决了基于图的异常检测技术缺乏标准化评估环境的问题，而这些技术对于保护这些复杂系统免受诸如提示注入等漏洞侵害至关重要。GAMMAF生成合成数据集并评估防御模型，证明有效的攻击补救措施可以提高系统完整性并降低运营成本。
-
Developers leverage Python libraries for LLM apps, while Harness & AWS focus on AI control
The tech landscape is rapidly evolving with AI, prompting discussions on control and application development. Harness.io is introducing solutions to manage AI's growth within DevOps and software development lifecycles, …
-
New models improve LLM reasoning evaluation and control over internal states
Researchers have developed a new framework to minimize "collateral damage" in activation steering for large language models (LLMs), which aims to control model behavior without negatively impacting performance on unrela…
-
新研究解决多智能体系统和 LLM 代理效率问题
近期研究探索了管理和改进多智能体系统 (MAS) 和 LLM 代理的先进技术。论文介绍了 CHRONOS 等框架，用于数据市场中的时间感知协调，以及 MAS-Orchestra，用于整体代理编排和基准测试。其他工作侧重于使用 OpenSkillEval 评估 LLM 代理技能，使用 TwinRouterBench 优化路由，以及使用 PushBench 确保目标持久性。此外，S-Bus 和 GraphFlow 解决了高效 LLM 代理…