🤖 LLM Agent Tech News

The HIP kernel uses one-instruction asm wrappers and an eight-wave pipeline to outperform AMD's AITER v3 on MI300X.The postMoonMath AI Open-Sources a HIP Attention Kernel for AMD MI300X That Beats AIT...

该开源 HIP 内核采用单指令 asm 封装和八波流水线技术，在 MI300X 显卡上的各种形状和舍入模式下，性能均超越了 AMD 官方的 AITER v3 基准。

LLMs are stateless by default. Agent memory fixes that. This guide breaks down all 7 types — working, semantic, episodic, procedural, retrieval, parametric, and prospective. It covers what each stores...

大模型默认是无状态的，而智能体记忆解决了这一问题。本指南拆解了工作记忆、语义、情景、程序、检索、参数化和前瞻性共 7 种记忆模块的存储机制与架构应用。

Cisco Foundation AI has open-sourced FAPO (Fully Automated Prompt Optimization), a Claude Code-driven system that autonomously optimizes multi-step LLM pipelines from baseline prompts to target accura...

思科 AI 基金会开源了 FAPO（全自动提示词优化），该系统由 Claude Code 驱动，能够通过步骤级失败归因机制，自主将多步骤大模型流水线从基线提示词优化至目标准确率。

Nous Research has added a Blank Slate setup mode to its open-source Hermes Agent. It starts an agent with everything off except provider, model, File Operations, and Terminal. You opt in to the rest.T...

Nous Research 为开源智能体 Hermes Agent 引入了“白板模式”。该模式默认禁用所有功能，仅保留提供商、模型、文件和终端权限，其余工具需开发者手动开启。

SpatialClaw is a training-free agent that writes Python in a persistent kernel, composing perception tools for 3D spatial reasoningThe postNVIDIA AI Introduce SpatialClaw: A Training-Free Agent That T...

SpatialClaw 是一个无需训练的智能体，通过在持久化内核中编写 Python 代码，组合各类视觉感知工具来实现精准的 3D 空间推理能力。

VibeThinker-3B, a 3B MIT-licensed reasoning model matching DeepSeek V3.2 and Kimi K2.5 on verifiable benchmarks.The postVibeThinker-3B: A 3B Dense Reasoning Model Built on Qwen2.5-Coder-3B With the Sp...

VibeThinker-3B 是一个采用 MIT 开源协议的 3B 参数稠密推理模型。通过 Spectrum-to-Signal 后训练流水线，它在可验证基准上的表现媲美 DeepSeek V3.2 和 Kimi K2.5。

Liquid AI's LFM2.5 Retrievers combine a dense bi-encoder and ColBERT late-interaction model for multilingual search on edge devices.The postLiquid AI Introduces LFM2.5-Embedding-350M and LFM2.5-ColBER...

Liquid AI 推出包含稠密双编码器和 ColBERT 延迟交互模型（晚期交互机制）的 LFM2.5 检索套件，专为边缘设备上的多语言搜索优化，支持 11 种语言。

Perplexity has launched Brain, a self-improving memory system for its Computer agent. Instead of remembering the user, Brain remembers the agent's work — what worked, what failed, and what corrections...

Perplexity 为其 Computer 智能体推出了 Brain 记忆系统。该系统不再记忆用户信息，而是记忆智能体自身的工作历程（成功、失败与纠正），并在夜间进行离线自学习与改进。

随着智能体在深度搜索和研究中被广泛使用，它们不可避免地会接触并处理敏感信息。MosaicLeaks 基准测试旨在评估这些研究型 Agent 是否会无意间泄露机密数据。

World models are increasingly regarded as a decisive step toward artificial general intelligence, yet modeling the physical world demands more than rendering convincing frames on demand: it requires a...

世界模型被视为迈向通用人工智能（AGI）的关键，但模拟物理世界不能仅靠按需生成逼真视频帧，还需要持久的状态跟踪。研究指出，当前主流世界模型从根本上缺乏这种持久的状态记忆核心。

Mainstream LLM serving systems reuse prefix work mainly through paged or radix key-value (KV) caches. This is highly effective for high-throughput, high-concurrency serving, but it manages only one po...

主流 LLM 服务系统主要通过分页 KV 缓存复用前缀，这适合高并发服务，但难以处理需要复杂状态管理的请求。该文提出执行状态胶囊，为端侧物理 AI 提供低延迟的检查点与恢复方案。

Policy-adherent tool-calling agents in customer-service domains must maintain task states across turns while calling tools and obeying domain policies. Task states consist of relevant facts, identifie...

在客服领域，严格遵守业务策略的工具调用智能体必须在跨轮次调用工具时维持任务状态。LedgerAgent 采用结构化状态账本，集中管理相关事实和身份信息，确保智能体合规执行。

Multimodal large language models (MLLMs) are increasingly deployed in personally and societally consequential settings, yet the visual cues that shape how these models judge people remain poorly under...

多模态大模型越来越多地部署在重要社会场景中，但塑造其判断的视觉线索尚不清楚。StylisticBias 研究指出，模型对人物的判断往往由极少数表面风格化视觉线索（如穿着或发型）所主导，引发严重社会偏见。

Autonomous agents are increasingly connected to cloud, deployment, and data-control workflows, but production mutation authority should not reside inside non-deterministic reasoning processes. Existin...

接入云和数据工作流的自治智能体日益增多，但生产环境的核心修改权限不应存在于不确定的推理过程中。该架构引入主权执行代理，通过绑定证书在控制平面实施严格的权限管控。

LiveCodeBench (LCB) has recently become a widely adopted benchmark for evaluating large language models (LLMs) on code-generation tasks. By curating competitive programming problems, constantly adding...

LiveCodeBench (LCB) 是评估大模型代码生成能力的基准。Multi-LCB 将其扩展到多种编程语言环境，通过持续更新的竞赛编程问题，全面测试大模型在跨语言场景下的泛化与代码生成能力。

Real-world spatial intelligence requires reasoning over a continuous and evolving 3D world, yet existing VLMs and tool-augmented agents largely remain tied to static, stateless inference from isolated...

现实世界的空间智能要求模型能在连续演化的 3D 环境中进行推理。S-Agent 摆脱了现有 VLM 只能进行静态孤立推理的限制，通过空间工具的调用实现对动态三维环境的感知与逻辑演算。

LLM-based coding agents need higher-level operational knowledge about a repository (which files house which subsystems, how to run the test suite, which workflows have historically led to wrong fixes)...

基于大模型的代码智能体在操作代码库时需要掌握高层级知识（如核心子系统分布、测试运行方式等）。本研究提出一种探查与微调机制，帮助智能体更好地理解并利用代码仓库的全局上下文信息。

Securing AI agents that operate in complex digital environments has become a critical need, and runtime monitoring approaches that formulate and enforce policies expressed in a formal language like Da...

在复杂数字环境中运行的 AI 智能体面临严峻的安全挑战。该研究提出一种基于运行时监控的方法，通过形式化语言（如 DSL）制定并执行策略，以概率模型高效验证智能体行为的安全性。

Prior work has shown that in-context demonstrations can jailbreak language models, but it remains unclear how models interpret different types of compliance demonstrations. We study this by mixing ben...

已有研究表明上下文示例可对大模型进行越狱，但模型如何解析不同类型的合规演示尚不明确。本研究通过混合良性指令与越狱示例，深入探究了安全对齐的大模型在面对混合合规演示时的内部机制。

When large language models serve as evaluators in multi-agent systems, their systematic evaluation biases propagate through the agent network. We introduce Contagion Networks, a formal framework for m...

当大模型在多智能体系统中担任评估者时，其固有的系统性评估偏见会通过网络传播并放大。研究提出了“传染网络”形式化框架来量化和控制这种偏差蔓延。

Real-world computer-use tasks often span multiple applications and devices, requiring agents to coordinate heterogeneous environments under dynamic runtime failures. Existing multi-device agent system...

现实中的跨设备计算任务常面临动态运行失败。现有系统多采用全局重规划导致效率低下。该研究提出一种层次化恢复机制，让智能体能在局部失败时进行针对性调整而非推倒重来。

Vision-language navigation agents achieve competitive average success on benchmark tasks, yet failures often arise through predictable trajectory-level breakdowns such as oscillation, stagnation, or i...

视觉语言导航智能体常出现震荡、停滞等轨迹级崩溃。GroundControl 通过估计轨迹一致性不确定性，能够在智能体彻底失败前预测并干预这些可预见的崩溃点。

In batch-style trace distillation, the same memory operation may receive contradictory feedback across different batches. Existing methods lack a cross-batch, operation-level evidence accumulation mec...

在批量轨迹蒸馏中，相同的记忆操作在不同批次可能收到矛盾反馈。现有方法缺乏跨批次操作级别的证据累积机制。本文提出边际优势累积机制，解决记忆演化过程中的冲突评估问题。

Context-heavy agents place unusual pressure on the key-value (KV) cache: long prefixes are reused across many short turns, while concurrency determines whether the serving system can keep GPUs utilize...

高上下文负载的智能体对键值（KV）缓存造成巨大压力。UltraQuant 提出一种 4-bit 量化方案，专门针对长前缀复用和短轮次并发的推理场景，有效提升 GPU 利用率。

Agentic AI systems increasingly rely on language-model components to interpret instructions, process external data, invoke tools, and coordinate with other agents. These capabilities make prompt-injec...

随着智能体系统广泛使用大模型处理指令、调用工具，提示词注入攻击风险激增。该研究探讨了如何利用防御性误导策略，对抗基于模型引导的自动化攻击，保障智能体系统的运行安全。

Mathematicians understand a PDE solution through mathematical structures rather than tables of computed values. Historically, this has been the product of mathematical analysis, carried out by hand fo...

数学家通过数学结构而非数值表理解偏微分方程（PDE）的解。本研究提出一种智能体驱动的符号搜索方法，无需依赖手工表达式或传统神经网络，自动发现并表征 PDE 的数学解。

Continual learning (CL) systems often forget previously acquired knowledge, yet the mechanisms driving forgetting remain hard to isolate in practice because real datasets entangle many factors. We pre...

持续学习（CL）系统常常遗忘旧知识，但其内在机制难以在复杂数据集中剥离分析。本研究通过合成数据集深入剖析了表征留存过程，揭示了稀疏性、叠加特征与灾难性遗忘之间的机制联系。

Vision-language models (VLMs) often generate fluent but visually unsupported descriptions, especially by mentioning objects absent from the image. We propose QK Product Steering, a data-free, training...

视觉语言模型（VLMs）常生成未经图像支持的幻觉描述。本文提出 QK Product Steering，这是一种无需数据、免训练的推理期干预方法，通过调整注意力机制缓解幻觉。

Large language model (LLM) agents are increasingly proposed as supervisory components for safety-critical systems, yet their robustness under sustained, adaptive adversarial pressure remains poorly ch...

大模型智能体正逐步被引入安全关键系统，但在持续、自适应的对抗攻击下的鲁棒性仍未知。研究全面评估了智能体在多轮红队测试和越狱基准下的表现，揭示了现有安全防线的脆弱性。

Large Language Model (LLM)-based agents are increasingly being used to automate multi-step engineering work flows by interacting with software tools, interpreting intermediate results, and autonomousl...

基于大语言模型的智能体正被用于自动化多步工程工作流。PowerAgentBench-Dyn 推出了一个专门针对电力系统动态研究的新基准，测试智能体在解释中间结果、自主调用专业软件方面的表现。

Spacecraft guidance, navigation, and control functions are increasingly realized as learned policies distilled from expert solvers. Developing such a policy is itself a research process: an investigat...

航天器的制导、导航与控制功能越来越依赖从专家求解器中蒸馏出的学习策略。Agentic AutoResearch 提出了一种完全可审计的大模型研究智能体，旨在自动化探索和生成这类复杂的航空航天控制策略。

Safety is a core concern for deploying reinforcement learning (RL) agents in real-world domains such as robotics and autonomous driving. While benchmarks have been central to progress in RL, existing ...

安全性是强化学习（RL）在机器人和自动驾驶等领域落地的核心阻碍。CRAX 提供了一个全新基准，专门针对快速且安全的强化学习进行评估，填补了现有基准在安全约束考量上的空白。

Large Language Models (LLMs) show promise for code compilation tasks, but applying them to runtime performance tuning is difficult due to complex microarchitectural effects and noisy runtime measureme...

尽管大语言模型在代码编译任务上表现出色，但复杂的微架构效应和嘈杂的运行时测量导致性能调优困难。AutoPass 引入证据引导机制，使 LLM 智能体能够更精准地优化编译器性能。

Explicit skill libraries make computer-using agents easier to inspect, but it remains unclear whether such libraries can be mined from interaction data in a way that improves downstream policies. We s...

显式的技能库能让计算机智能体更易被审查，但如何从交互数据中提取有效技能库尚不明确。该研究表明，通过挖掘交互轨迹自动生成 SKILL.md，能够显著提升下游任务策略的执行效果。

Agent skills are commonly deployed as natural-language Markdown files that encode answer policies, evidence-use habits, and task procedures. These files are readable and portable, but they are consume...

智能体技能通常以自然语言 Markdown 文件形式部署，虽然易读但消耗大量 Token。SoftSkill 提出行为压缩技术，将这些技能文件压缩为短前缀或连续表征，优化上下文适配并降低成本。

Large language models (LLMs) have achieved strong performance across a wide range of language-based tasks by leveraging both extensive parametric knowledge and in-context learning ability, enabling th...

大模型依赖参数化知识与上下文学习能力处理任务。当用户提供的外部上下文与模型固有知识相悖时，推理会受影响。该研究提出显式知识冲突消解机制，提升模型在矛盾信息下的鲁棒性。

We present Phoenix, a multi-agent LLM system that resolves GitHub issues from triage through pull-request creation, combining seven layered safety controls with a baseline-aware test evaluation strate...

Phoenix 是一个多智能体大模型系统，旨在解决从分类到创建 Pull Request (PR) 的 GitHub issue。该系统结合了七层安全控制机制和基线感知的测试评估策略，保障代码修复的安全性与准确性。

Multimodal Large Language Models (MLLMs) have demonstrated remarkable success in various Remote Sensing (RS) tasks. However, their ability to comprehend negation remains underexplored, limiting deploy...

多模态大语言模型在遥感（RS）任务中表现出色，但对“否定”语义的理解能力仍未被充分挖掘。这限制了模型在复杂遥感场景中准确排除特定目标的部署能力。

AI agents may soon search for and use their own tools at runtime, thanks to a new open standard backed by Microsoft and Google.

得益于微软和谷歌支持的一项全新开放标准，AI 智能体未来或能在运行时自主搜索并调用合适的工具 API。

Large language models (LLMs) are increasingly applied to computer-aided design (CAD) to generate design artifacts from textual instructions. In engineering practice, this requires more than creating n...

大模型正越来越多地应用于 CAD 以生成设计工件，但在工程实践中还需要修改现有模型。BIM-Edit 基准专门评测 LLM 理解和编辑基于 IFC 标准的建筑信息模型的能力。

The KV cache now outweighs model weights at long context. Here's how TurboQuant, OSCAR, and EpiCache each attack that memory bottleneck — and why they're more complementary than competitive.The postTh...

在处理长上下文时，KV 缓存的内存占用已经超过了模型权重。文章对比了 TurboQuant、OSCAR 和 EpiCache 三种前沿技术如何攻克这一内存瓶颈，并指出它们更多是互补而非纯竞争关系。

MLLM-based mobile GUI agents have made substantial progress in UI understanding and action execution, but adapting them to real target apps remains costly because mobile apps are numerous, frequently ...

基于多模态大模型的移动端 GUI 智能体在 UI 理解上进步显著，但适配海量且频繁更新的目标 App 成本高昂。MobileForge 提出一种无标注适配框架，利用分层反馈引导策略优化，实现低成本迁移。

MLLM-based mobile GUI agents have made substantial progress on short-horizon tasks, yet remain unreliable on long-horizon tasks that require retaining intermediate facts across many steps and app tran...

现有的移动端 GUI 智能体在短程任务上表现良好，但面对跨越多个步骤和 App 转换的长程任务时往往失效。MemGUI-Agent 引入了端到端的主动上下文管理机制，有效保留中间事实，提升长程可靠性。

OpenAI's LifeSciBench evaluates whether frontier AI can handle real life-science research across 750 expert-authored tasks, seven workflows, and seven biological domains. Built by 173 PhD scientists w...

OpenAI 发布的 LifeSciBench 旨在评估前沿 AI 在真实生命科学研究中的能力。该基准包含 750 个由 173 位博士科学家编写的任务，覆盖七大生物领域和工作流。

In this tutorial, we use NVIDIA SkillSpector to evaluate AI skills for security risks before deployment. We build a corpus of benign and deliberately vulnerable skills, then scan them through SkillSpe...

该指南使用 NVIDIA SkillSpector 工具，通过静态分析扫描 AI 技能包，在部署前发现潜在的安全漏洞，并输出标准化的 SARIF 格式安全报告。

如何评估开源模型是否具备接入自有工具链执行任务的能力至关重要。该测试方法专注于评估模型在特定业务环境和工具集下的智能体化表现。

LoRA 是目前最流行的大模型参数高效微调（PEFT）技术。本文探讨并测试了是否有其他替代方案能够在微调效率或效果上超越 LoRA。

Passive models for long video understanding typically rely on a "watch-it-all" paradigm, processing frames uniformly regardless of query difficulty, causing computational cost to grow with video durat...

针对传统长视频理解模型“全盘观看”导致计算成本随时长线性暴增的问题，研究提出主动感知机制，根据查询难度动态调整需要处理的视频帧。

Progress in legal AI increasingly depends on access to authoritative legal text at scale. Yet one of the most consequential layers of American law remains largely absent from existing machine-readable...

法律人工智能的发展严重依赖于权威法律文本的大规模获取，但美国法律体系中最关键的地方法规层目前缺乏机器可读的格式。LOCUS 填补了这一空白，提供了一个全面的美国地方法规语料库。

Post-training of reasoning language models is commonly driven by supervised distillation and reinforcement learning with verifiable rewards. Distillation often relies on chain-of-thought annotations t...

推理语言模型的训练常依赖思维链标注进行蒸馏，或使用带可验证奖励的强化学习。该研究提出一种基于评分表的条件自蒸馏技术，以优化这两类训练范式的结合效果。

Production data integration is bottlenecked by repeated, lossy handoffs between data owners, engineers, and analysts who must collaboratively discover, structure, and query enterprise data. We present...

为解决数据所有者、工程师和分析师之间反复低效交接导致的生产数据集成瓶颈，研究提出自主编码的数据智能体，实现企业数据的自动化发现、结构化与查询。

Enhancing the formal math reasoning capabilities of Large Language Models (LLMs) has become a key focus in both mathematical and computer science communities in recent years. While significant progres...

提升大语言模型在形式化数学推理方面的能力一直是难题。该研究跳出传统的自回归生成框架，探索扩散模型在形式化定理证明任务中的配方与潜力。

Large language model (LLM)-based multi-agent systems (MAS) have demonstrated great potential in solving tasks with execution complexity, by distributing subtasks across cooperative agents. However, th...

基于大模型的多智能体系统在分配协作任务时，容易因利益冲突或信息不对称陷入次优解。研究引入虚拟博弈机制，提升多智能体复杂任务下的决策鲁棒性。

Vercel has open-sourced eve, an Apache-2.0 agent framework now in public preview. An agent is a directory of files, with durable execution, sandboxes, approvals, connections, channels, and evals built...

Vercel 开源了 Apache-2.0 协议的 Agent 框架 eve。它将智能体定义为一系列文件的目录，内置了持久执行、沙箱、审批、连接和评测能力。

Embodied Vision-Language-Action (VLA) models are typically obtained by fine-tuning powerful pretrained VLMs on robotics data, yet it is unclear how much commonsense and factual knowledge they retain a...

具身视觉-语言-动作（VLA）模型通常基于强大的预训练 VLM 微调而来，但研究发现它们在机器人数据上微调后，保留常识和事实知识的能力存在疑问。

The knowledge encoded in large language models (LLMs) can serve as a substrate for structured reasoning over variables describing a complex world, but accessing this knowledge in a probabilistically c...

大语言模型编码的知识可作为复杂世界变量进行结构化推理的基础，但在概率上一致性访问这些知识面临挑战。该研究引入大型语言模型吉布斯采样，以实现可靠的结构化推理。

Block diffusion language models accelerate decoding through parallel block-wise denoising, yet whether they can be reliably scaled for long chain-of-thought (CoT) reasoning remains unresolved. To this...

块扩散语言模型能通过并行去噪加速解码，但其能否可靠地扩展至长链推理任务尚不明确。DreamReasoner-8B 引入分块课程学习，成功提升了扩散模型在长逻辑链上的稳定性。

Artificial intelligence (AI) agents promise to accelerate drug discovery by compressing interpretation and decision-making loops, but practical deployment requires trusted evaluation on realistic prog...

虽然 AI 智能体有望通过加速解释和决策闭环来缩短药物发现周期，但在实际部署前需要可信的评估。TxBench-PP 旨在评估智能体在真实临床前药理学任务中的表现。

AI agents now handle personal data through tool use, function calls, and multi turn dialogue, which can create obligations under the General Data Protection Regulation (GDPR). Current testing practice...

随着 AI 智能体通过工具调用和多轮对话处理个人数据，这触发了 GDPR 等法规的合规义务。本研究提出针对智能体运行时的合规性自动化验证机制。

Reinforcement Learning with Verifiable Rewards algorithms like GRPO have emerged as the dominant post-training paradigm for complex reasoning in LLMs, yet commonly suffer from policy entropy collapse ...

像 GRPO 这样基于可验证奖励的强化学习算法在训练复杂推理 LLM 时，常出现策略熵迅速下降的问题。STARE 通过 token 级的优势重加权机制来稳定熵值，防止模型过早失去探索能力。

Code large language models increasingly retrieve external code context from repositories, documentation, issue threads, and coding-agent environments, creating an indirect prompt-injection surface whe...

针对代码大模型从代码库、文档或 Agent 环境中读取外部上下文时面临的间接提示词注入攻击面，研究提出 CodeSentinel 三层防御架构以确保代码安全。

Agent skills allow LLM-based coding agents to acquire domain-specific capabilities from third-party packages, but they also introduce a new supply-chain attack surface. We present PhantomSkill, an att...

Agent 技能允许大模型从第三方包获取能力，但也引入了新的供应链攻击面。PhantomSkill 展示了黑客如何通过恶意代码注入对 Agent 生态系统发起攻击。

Personal memory in a language model is two problems: content and reasoning skill. The brain keeps the two apart (a sparse, local engram in the hippocampus for each episode, a slow neocortex for the sh...

语言模型的个人记忆包含内容和推理能力两个层面。受大脑海马体局部稀疏记忆印迹启发，该研究通过局部参数编辑的方式将用户专属记忆与模型通用能力解耦。

To achieve deeper safety alignment for large language models (LLMs), recent efforts have studied how to push safety interventions earlier into the pretraining stage, primarily by filtering unsafe data...

为了使大语言模型实现更深度的安全对齐，研究者开始将安全干预推前至预训练阶段。该工作摒弃了单纯过滤有害数据的做法，转而采用定期安全反思的机制来构建更安全的基座模型。

AudioLLMs enable speech recognition conditioned on textual prompts such as domain descriptions or entity lists. However, it remains unclear whether these models genuinely utilise such context or rely ...

音频大模型能够根据领域描述或实体列表等文本提示来调整语音识别结果。该基准测试评估 AudioLLMs 在 8 种印度语中是否真正有效利用了上下文信息。

As large language models (LLMs) advance and multi-agent systems aim to overcome the limits of standalone agents, robust communication protocols are becoming essential infrastructure for distributed ag...

随着大模型多智能体系统试图突破单智能体限制，建立稳健的通信协议正成为分布式智能的关键基础设施。本文对现有的智能体通信协议进行了系统性的技术分类与梳理。

Current image editing methods excel at static attributes but fail at complex Human-Object Interactions (HOI), a critical challenge unaddressed by existing benchmarks that conflate HOI with static attr...

现有图像编辑方法在静态属性上表现出色，但无法处理复杂的人机交互（HOI）。研究提出了一种认知基准和自纠正智能体框架，利用 I2V 模型解决复杂 HOI 图像编辑挑战。

Multi-turn tool-use RL is bottlenecked by the rapid depletion of informative samples in static datasets. We observe that the gradient signal in GRPO concentrates on tasks with the highest rollout rewa...

多轮工具调用强化学习常因静态数据集中有效样本耗尽而遇到瓶颈。RODS 技术通过观察 GRPO 算法中的梯度信号，在线动态合成高价值的新训练数据，以维持模型的学习效率。

Agents are increasingly deployed in document-intensive workflows where sensitive private information is not an edge case but a routine input, e.g., an agent booking a flight needs passport numbers. In...

在处理机票预订等密集文档工作流时，智能体常接触护照号等敏感隐私。TRAP 基准重点测试智能体在完成复杂任务的同时，能否抵抗针对隐私数据的主动提取攻击。

OpenAI and Molecule.one show how a near-autonomous AI chemist using GPT-5.4 improved a key drug-making reaction, advancing medicinal chemistry research.

OpenAI 和 Molecule.one 展示，基于 GPT-5.4 构建的近乎自主的 AI 化学家，成功优化了药物合成中的一项关键反应，推进了药物化学研究。

GLM-5.2 发布，该模型专为处理需要多步骤、长时间跨度规划与执行的长程任务而设计。

MiniMax released MSA, a sparse attention built on Grouped Query Attention. A lightweight Index Branch selects Top-k key-value blocks per query and GQA group; the Main Branch attends only to those bloc...

MiniMax 发布 MSA 稀疏注意力机制，基于分组查询注意力（GQA），通过轻量级索引分支筛选 Top-k KV 块，主分支只关注这些块，从而大幅降低计算量。

OpenAI introduced Deployment Simulation on June 16, 2026. The method replays past conversations through a new candidate model before release. It then grades the completions to estimate deployment-time...

OpenAI 引入“部署模拟”机制，在正式发布前用候选模型重放历史对话，并模拟工具调用，以评估并打分其在智能体编码场景下的实际部署风险。

AI web agents can perform complex, multi-step tasks such as searching for products, comparing options, and making purchases on behalf of users. However, verifying the correctness of an agent's output ...

AI Web 智能体能代用户执行比价、购物等复杂多步任务，但很难验证其输出正确性。HANSEL 通过在智能体的操作轨迹中提取“面包屑”线索，实现对执行过程的交互式验证。

We implement xFormers, a practical toolkit for fast, memory-efficient Transformer models on GPUs. We validate memory-efficient attention against a standard implementation, then compare speed and memor...

文章介绍了如何使用 xFormers 工具包在 GPU 上构建快速且节省内存的 Transformer 模型。通过对比标准实现，验证了高效注意力机制在速度和内存上的优化效果。

Introducing LifeSciBench, an expert-authored, expert-reviewed benchmark for evaluating how AI systems handle real-world life science research tasks and decisions.

LifeSciBench 是一个由专家撰写并审查的全新基准测试，专门用于评估 AI 系统在处理真实世界生命科学研究任务和复杂决策时的表现。

Unified Multimodal Modeling aims to integrate visual understanding and generation within a single system. However, existing approaches typically rely on two disparate visual tokenizers, which splits t...

针对现有统一多模态模型在视觉理解与生成上使用不同分词器导致割裂的问题，研究提出使用单一的共享上下文视觉分词器，以实现真正的多模态大一统自回归建模。

Current world models face a fundamental tension: faithful long-horizon simulation demands deep computation, but deeper models are expensive to deploy and prone to compounding errors. We resolve this b...

当前世界模型面临根本性矛盾：长期的忠实模拟需要深度计算，但深层模型昂贵且易产生累积误差。本文通过引入循环架构来解决这一矛盾。

Looped architectures provide an inductive bias toward learning step-by-step procedures for tasks that require compositional reasoning. The number of effective layers reached by looping determines the ...

循环架构为需要组合推理的任务提供了逐步学习的归纳偏置。本文提出“定点推理器”，通过寻找稳定的不动点来决定循环层数，实现自适应深度的推理。

The LLM-empowered personal health agents with user health (sensor) metrics have offered a promising pathway to alleviate global disparities in healthcare access. However, large-scale clinical deployme...

结合用户健康传感器数据的 LLM 个人健康智能体有望缓解医疗资源不均。然而，为了实现大规模临床部署，本文提出了 RubricsTree，一种可扩展且不断演进的开放式评估方法。

Agent skills are emerging as an important attack surface in LLM-based systems. Through an empirical study of existing skill scanners, we find that current defenses primarily rely on textual descriptio...

Agent 技能（插件）正成为 LLM 系统的重要攻击面。研究发现现有防御主要依赖文本检测，多模态隐藏指令攻击可以通过在图像中藏匿指令绕过扫描器。

On-policy self-distillation (OPSD) has proven effective for post-training large language models (LLMs), yet its application to diffusion LLMs (dLLMs) remains unexplored. Existing OPSD methods are inhe...

研究将在线策略自蒸馏（OPSD）引入扩散大模型，解决了现有自蒸馏方法内在的非在线策略缺陷，提升了模型的后训练效果。

We evaluate the adversarial robustness of two frontier large language models (LLMs) developed by Anthropic, Fable 5 and Opus 4.8, against four families of automated jailbreak attack across 7 826 harmf...

研究评估了 Anthropic 开发的两款前沿大模型 Fable 5 和 Opus 4.8 在面对 7826 种有害提示和四类自动越狱攻击时的对抗鲁棒性。

Deep research (DR) systems are increasingly used for complex information-seeking tasks, but existing works mainly focus on generating reports and summaries. In contrast, many enterprise tasks instead ...

深度研究系统多用于生成报告，但许多企业级任务需要预测和执行个性化的研究工作流。DRFLOW 基准专门评估模型在这方面的能力。

Illegal, unreported, and unregulated fishing (IUU) traditionally refers to fishing activities that violate applicable laws or occur in areas that lack applicable laws. We propose the term IUU+ to capt...

研究提出 IUU+ 概念，利用大语言模型从非结构化文本中提取信息，追踪传统上缺乏有效监管的非法、未报告和无管制（IUU）捕捞、海鲜欺诈及劳工滥用问题。

Chain-of-thought generation can turn a multi-step computation into a sequence of locally checkable state updates, but the training dynamics by which transformers acquire such updates remain poorly und...

研究通过构建可解的 Transformer 模型，深入探讨了模型如何通过思维链将多步计算转化为局部可检查的状态更新，并揭示了模型在训练过程中掌握这种能力的内在动态机制。

Large language models now produce legal text of at least median quality, yet no existing benchmark can evaluate whether they perform doctrinal legal reasoning, which forms the interpretive core of leg...

尽管现有大模型生成的法律文本质量已达中位数水平，但目前缺乏评估其是否真正执行教义法律推理的基准。本研究针对《欧盟AI法案》提出了全新的评估基准。

Recent incidents involving LLMs used for mental-health support reveal a critical evaluation gap: surface-level safety scores do not capture how models behave across realistic, emotionally sensitive in...

针对大模型在心理健康支持场景中的评估盲区，研究指出表面安全得分无法反映模型在真实敏感情绪交互中的表现，揭示了模型在长时间运行后可能出现的认知能力下降问题。

Modern AI systems increasingly solve a task not with a single model call but with several imperfect agents working together: some propose pieces of a solution, others verify them, and the results are ...

现代 AI 系统越来越多地使用多个不完美的 Agent 协同工作（提出方案、验证结果）。本文研究了这类 Agent 网络的可靠性，引入了密度演化和停止集等概念进行架构优化。

State Space Models (SSMs) such as Mamba-2 offer linear-time inference but their memory footprint limits edge deployment. Prior ternary SSM work (Slender-Mamba) trains from scratch on 150B tokens; we s...

针对 Mamba 等状态空间模型（SSM）显存占用大、难以在边缘设备部署的问题，本研究提出三值化分组量化感知训练方案（W1.58A16），显著降低模型体积。

Agentic navigation systems require a base navigation model whose observation strategy can be externally reconfigured at inference time, because instruction following, object search, target tracking, a...

阿里 Qwen 团队发布 RobotNav 报告，指出导航系统需要能在推理时外部重配置观测策略的基础模型，以适应指令跟随、目标追踪等动态任务。

Recent advances in Large Language Models (LLMs) and multi-agent systems have driven the rise of Agentic AI, showing promise for medical reasoning. However, open-ended conversational agents remain pron...

尽管大模型在医疗推理中很有潜力，但开放式对话 Agent 容易产生幻觉或过早得出结论。本文提出一种 Agentic AI 框架来缓解这些问题。

As LLM-driven multi-agent systems (MAS) transition from lab to production, system behavior exhibits nonlinear degradation. We introduce the Intelligence Entropy Principle: probability-driven systems s...

随着 LLM 驱动的多智能体系统（MAS）走向生产，系统行为表现出非线性退化。本文提出了“智能熵原则”以及 ADE 框架来解决概率驱动系统的稳定性问题。

LLM agents increasingly rely on external skills -- reusable tool specifications -- but real-world tasks often require composing multiple skills, not just selecting one. We formalize this as the Compos...

现实任务往往需要组合多个外部工具（技能），而非单一调用。本文将此形式化为“组合式技能路由”问题，提出了分解、检索和组合的解决路径。

Vision-language-action models (VLAs) combine vision-language backbones with expressive generative action heads trained via flow matching on large-scale robotic datasets. Despite their strong empirical...

视觉-语言-动作模型（VLAs）结合生成式动作头，在机器人任务中表现出色。但在关键场景中，量化其输出的不确定性至关重要。本文提出了一种针对 Flow-based VLAs 的不确定性量化方法。

Tool-using LLM agents increasingly use the Model Context Protocol (MCP) to answer from heterogeneous evidence sources, including search, APIs, databases, clinical records, and formulary tools. Standar...

使用工具的 LLM Agent 常通过模型上下文协议（MCP）从异构证据源（如 API、数据库、病历）回答问题。标准验证方法难以应对，ProvenanceGuard 提供了一种具备溯源能力的验证机制。

AI systems deployed in legal workflows hallucinate at rates that aggregate metrics report at ~52%, but this average conceals where errors concentrate and in which direction they run, leaving complianc...

部署在法律工作流中的 AI 系统幻觉率高达 52%。LegalHalluLens 提出了类型化幻觉审计和多 Agent 校准辩论机制，以确保法律 AI 的可信度。

Visual attribution is a fundamental tool for interpreting modern vision and vision-language models, particularly when their decisions must be inspected, diagnosed, or audited. Its goal is to explain h...

视觉归因是解释现代视觉和视觉语言模型的重要工具，旨在解释模型为何做出某种决策。PhaseWin 算法提供了一种高效的搜索方案，用于更精准地进行视觉归因。

Semi-supervised medical image segmentation has emerged as a dominant research problem in medical image analysis, mitigating annotation scarcity by leveraging consistency regularization on unlabeled da...

半监督医学图像分割通过利用未标记数据的一致性正则化来缓解标注稀缺问题。本文提出利用思维链（Chain-of-Thought）增强推理来提升模型表现。

When vision contradicts text, multimodal large language models (MLLMs) consistently favor text, even when images provide clear evidence otherwise. This bias poses risks for applications requiring visu...

当视觉信息与文本矛盾时，多模态大语言模型（MLLMs）往往会忽略图像证据而盲目相信文本。这种偏见对需要视觉敏锐度的应用构成风险，研究提出了一种追踪和纠正深层文本偏见的方法。

Most Vision-Language-Action (VLA) models map observations directly to actions without explicit reasoning, limiting their capacity for reasoning-intensive long-horizon tasks. To address this, existing ...

现有的视觉-语言-动作（VLA）模型大多直接将观测映射为动作，缺乏明确的推理过程。ThinkingVLA 引入了交织的视觉和语言推理，以解决长时序复杂任务的能力瓶颈。

Foundation models in language and multimodality achieve strong generalization by aligning heterogeneous data under a unified formulation and training at scale. In this report, we investigate whether t...

阿里 Qwen 团队发布了机器人操作技术报告，探讨了将语言和多模态基础模型的对齐技术应用于异构机器人数据，验证了规模化训练在机器人操作中的效果。

Humanoid robots promise whole-body interaction in human-centered environments, but scalable policy learning remains difficult because task-level decision-making and whole-body dynamic execution are ti...

人形机器人有望在以人为中心的环境中进行全身交互，但任务级决策和全身动态执行的脱节阻碍了策略学习的扩展。HumanoidArena 为此提供了一套全身学习的基准。

As an increasing majority of global video content is consumed on social platforms for interactive social purposes, video generation models built for social worlds are important but largely overlooked ...

随着全球视频内容在社交平台上的互动消费激增，专为社交世界构建的视频生成模型变得重要却常被忽视。MaineCoon 旨在打造一个实时的视听社交世界模型。

Despite the remarkable progress of Video Large Language Models (Video-LLMs), current online architectures still struggle to simultaneously process continuous video streams, decide autonomously when to...

目前的视频大语言模型在同时处理连续视频流、决定何时响应以及提取长时序记忆方面存在困难。LiveStarPro 提出了一种具备分层记忆的主动式流媒体视频理解架构。

Generalist embodied agents require more than object recognition: they must reason about spatial relations, actions, procedures, human intentions, environmental constraints, and commonsense consequence...

通用具身智能体不仅需要识别物体，还需要对空间关系、动作程序和人类意图进行推理。ERQA-Plus 提供了一套诊断基准，评估这些复杂推理能力。

OpenAI introduces Deployment Simulation, a method to predict AI model behavior before deployment using real conversation data to improve safety and evaluation accuracy.

OpenAI 推出“部署模拟”方法，利用真实对话数据在实际部署前预测 AI 模型的行为，从而提升安全性和评估的准确性。

Personalized presentation generation requires more than conditioning on a current prompt or template: agents must preserve stable user preferences across tasks, retain newly introduced preferences and...

为解决个性化 PPT 生成难题，MemSlides 框架引入分层记忆机制，使智能体不仅能记住用户跨任务的固定偏好，还能处理新偏好并进行局部多轮修改。