Tag
MemCo提出了一种以记忆为中心的协作框架,通过维护互补的局部与全局记忆空间,帮助LLM智能体在未见过的交互环境中实现更好的泛化,实验表明其在交互式决策基准上提升了任务成功率并减少了冗余探索。
The paper introduces NP-Bench, an environment-grounded benchmark for verifying coordination among parallel LLM coding agents, and a proactive scheduling planner built into Nerveplane that partitions work into disjoint scopes and orders merges along dependency graphs. The planner lifts clean-integration success from 1/9 to 9/9 scenarios, eliminates merge conflicts, and collapses repeated-mistake rates from 1.00 to 0.00 via cross-session memory.
提出 ME-World,一种多智能体自我中心世界模型,通过在共享 token 序列中联合去噪多个 ego 流,实现细粒度交互下的跨视角一致性、共享环境一致性与状态更新传播,并引入共享世界一致性指标。
MARGIN提出了一种运行时置信度校准方法,无需重训练或预留校准集即可从观测结果中学习模型特定的置信度修正,用于异构基础模型的多智能体协调决策,在代码生成和问答任务上相比未校准的置信度加权提升了答案选择准确率。
Anthropic为Claude Code推出工业级多智能体协作架构:Haiku 5.5作为并行高速子代理处理机械任务,Sonnet 5.5作为主力生成代码,Opus 5.5仅在关键决策点以--advisor身份介入,通过'三权分立'式分层显著降低旗舰模型的上下文与token消耗。
一位用户分享如何在 Claude Code 中通过三层模型分工(Opus 5.5 负责规划与最终 diff,Sonnet 5.5 负责侦察/执行/文档,Fable 5.1 作为 advisor 仅在关键节点被咨询)大幅降低成本,并给出可直接粘贴的配置提示词,附带 Claude Code advisor 工具官方文档。
SquadRoom 推出一款多智能体 AI 团队产品:一个主智能体负责学习公司情况并规划工作,研究、写作、合作和运营等专家智能体自动执行任务,用户只需把任务放到看板上即可等待完成的结果与报告。
Docker 与 GitHub 上的 docker-agent 开源项目发布了一款 Docker CLI 插件,可通过声明式 YAML 配置构建、运行和分享 AI 智能体,支持多智能体编排、MCP 工具生态、RAG 检索以及多模型供应商。
LiveMACE is an open-source benchmark that uses live cryptocurrency and US-equity markets to evaluate LLM agent capabilities in evolving conditions. Across five frontier LLMs over 30 days, it finds an outcome–capability gap, where higher returns do not consistently reflect stronger tool use, memory, rule-following, or collaboration.
提出首个系统化的多智能体视觉-语言导航(MAVLN)形式化框架,包含11,724个跨145个场景的多机器人协作任务,并提出基于LLM子任务调度与共享拓扑记忆的协调系统TRISS作为基线。
该论文研究两个机器人智能体在二维离散环境中共同演化出的通信机制,在迁移到基于物理的三维模拟器后是否仍能保持其功能作用。结果显示通信信号主要作为辅助解堵机制而非导航引导,且更长的评估时长能持续提升性能。
Meta published how they mapped their entire 4,100-file, 4-repo pipeline codebase into a navigable graph for AI agents using 50+ agents in one session, cutting tool calls by 40% with zero hallucinated file paths. The 'compass, not encyclopedia' setup produces 59 small context files and is designed to be easily copied for other repos.
Jotbus is an end-to-end encrypted shared scratchpad that lets coding agents like Claude Code and Codex exchange notes, files, and work handoffs across machines without signup or codebase access.
介绍开源项目 Paseo(18,000+ stars),它将 Claude Code、Codex、Copilot、OpenCode 和 Pi 等 AI 编程代理整合到一个统一工作台,支持本地并行运行、代理间上下文接力、协同排错,并可在桌面、移动端、网页和 CLI 使用。
Tencent Cloud open sourced Octop, a self-hosted, multi-user AI workspace where each user gets isolated accounts and can spin up multiple agents with swappable models, memory, plugins, and built-in browser, terminal, and remote desktop, with handoff support for coding agents like Claude Code and Cursor.
An open, replayable benchmark event where 10 AI models play a single world war with pacts, betrayal and nukes, recording every decision, reasoning and observation for full replayability.
SPLATIFY 是一个多智能体框架,可将 3D Gaussian Splatting 研究论文转化为可训练的 gsplat 代码,支持复现、发现与创新,将实现时间从数周缩短至数分钟,并提出覆盖 30 篇论文的 SPLATIFY-Bench 评测基准。
论文提出 Stateless Language Agents (SLAs),采用「有状态搜索 + 无状态代理」原则,由框架持有研究状态并为每次调用重建专属上下文,在长达十亿 token 的预算下于软件工程、内核优化和算法设计任务中均取得最佳结果,且比最强基线节省超过 84% 的 token。
该论文提出 Station 系统,让 AI 代理在禁用网络访问的条件下自主进行开放式科学发现,在基于三篇 ICLR 口头报告论文构建的任务上平均恢复了 62.7% 的预定义子发现,远超 AI Scientist-v2(20.6%)和 Codex Multiagent-v2(15.4%),并开源了代码与完整研究记录。
The paper introduces Recursive Game Creator, an experience-oriented agentic framework with Designer, Builder, Player, and Reviewer roles that recursively transforms game prototypes into polished games, achieving state-of-the-art results on GameCraft-Bench and GameASG-Bench via its Coding-Native Player for gameplay evaluation.