🤖 LLM Agent Tech News

New AI toolbars and prompts are showing up in Google Docs and Gmail. If you don’t want Gemini’s help in writing documents and emails, here’s how to turn that stuff off.

谷歌在 Docs 和 Gmail 中加入了新的 AI 工具栏和提示词。如果你不需要 Gemini 协助编写文档和邮件，本文提供了关闭这些功能的具体步骤。

Two security researchers bought cheap domains—including noreply.net and deleteduser.com—and set up email listening services. Hundreds of companies are sending them corporate secrets.

两名安全研究人员购买了 noreply.net 等廉价废弃域名并设置邮件监听服务，结果发现有数百家公司误将企业机密发到了这些域名对应的邮箱中。

Dive into the advanced visualization capabilities of the Reflex XY Python library. This tutorial guides you through building high-performance, interactive charts—from handling million-point datasets a...

该教程深入讲解了 Reflex XY Python 库的高级可视化功能，指导开发者如何处理百万级数据点、流式数据，并构建自定义标记的高性能交互式图表。

Mistral AI has released Shieldstral 1.0 3B, an open-weights, policy-adaptive multimodal safety classifier that frames content moderation as a single yes/no question instead of a fixed harm taxonomy. O...

Mistral AI 发布了 Shieldstral 1.0 3B，这是一个开放权重、支持自适应策略的多模态安全分类器。它将复杂的伤害类型分类简化为简单的“是/否”问题判断。

Tencent Cloud has open-sourced TencentDB Agent Memory v2.0, a team-level memory hub that turns conversations, documents and code into four governed, reusable assets — Chat Memory, Skill, LLM-Wiki and ...

腾讯云开源了 TencentDB Agent Memory v2.0，这是一个团队级记忆中心，能将对话、文档和代码转化为四种受管治的可复用资产：对话记忆、技能库、LLM-Wiki 和经验库。

In this tutorial, we build an advanced multimodal retrieval-augmented generation pipeline with NVIDIA NeMo Retriever. We begin by configuring a Python 3.12 environment, installing the required package...

该教程介绍了如何利用 NVIDIA NeMo Retriever 等工具搭建高级的多模态检索增强生成（RAG）管道，涵盖环境配置、重排序（Reranking，即对初步召回的结果进行二次精准打分排序）和基础生成。

NVIDIA Labs has open-sourced NOOA (NVIDIA Object-Oriented Agents), a model-agnostic Python framework for building AI agents. Agent development today is split across prompt templates, tool schemas, cal...

NVIDIA 实验室开源了 NOOA（NVIDIA 面向对象 Agent），这是一个与模型无关的 Python Agent 构建框架。它将散落在提示词模板、工具配置和回调中的逻辑，统一封装进一个 Python 类中。

Cloudflare has introduced Kitesurf, a cloud-hosted browser designed for AI agents instead of people. The company says the browser uses less computing power than Chromium for common automation tasks, h...

Cloudflare 推出了云托管的 Kitesurf 浏览器，专为 AI Agent 而非人类设计。官方称在常见自动化任务中，它比 Chromium 消耗的算力更少。

Microsoft has open sourced code-testing-generator, a polyglot unit-test agent shipping in the MIT-licensed dotnet/skills repository. It reads a repository before writing anything — detecting the langu...

微软开源了 code-testing-generator，这是一个支持多语言的单元测试智能体，已在 MIT 许可下的 dotnet/skills 仓库中发布。它在编写代码前会先读取整个代码库，检测编程语言和框架。其任务完成率达到 92.1%，而原版 Copilot 为 78.9%。

Liquid AI released LFM2.5-2.6B, an agentic model that plans, calls tools, and completes multi-step tasks entirely on-device. The 2.69B parameter model pairs 22 double-gated short convolution blocks wi...

Liquid AI 发布了 LFM2.5-2.6B 模型，这是一个完全在端侧运行、可进行规划、调用工具并完成多步任务的智能体。该模型拥有 26.9 亿参数，结合了 22 个双门控短卷积块，专为端侧设备优化。

Cloudflare has released Kitesurf, a stateless web browser built specifically for AI agents that runs entirely in V8 isolates on Cloudflare Workers, with no Chromium underneath. The browser drops human...

Cloudflare 发布了 Kitesurf，这是一款专为 AI 智能体设计的无状态网页浏览器。它完全运行在 Cloudflare Workers 的 V8 隔离区中，底层不依赖 Chromium，抛弃了人类交互界面的冗余，专为机器读取和操作网页而设计。

Language models increasingly condition their answers on external signals, and a single misleading one can turn a correct answer wrong. The obvious remedy, training models to resist such signals, hides...

语言模型越来越多地基于外部信号（如 RAG 提供的上下文）生成答案，而单个误导性信号可能导致正确答案变错。显而易见的补救措施是训练模型抵抗这些信号，但这会削弱模型获取新信息的能力。该研究提出通过偏好优化让模型学会判断何时该信任外部上下文。

Humanoid household tasks often require concurrent loco-manipulation, where the robot must move, adjust posture, maintain balance, and manipulate objects as a single coordinated behavior. Yet existing ...

人形机器人执行家务任务通常需要并发移动操作（边走边拿东西等），机器人必须将移动、姿态调整、保持平衡和物体操作作为一个协调行为。该研究提出了 $ω$-0 模型以解决这一难题。

Vision-Language-Action (VLA) models have become a powerful paradigm for robot manipulation, but training a single generalist policy for heterogeneous robot embodiments remains an open problem. Existin...

视觉-语言-动作（VLA）模型已成为机器人操作的强大范式，但为异构机器人本体训练单一通用策略仍是一个未解难题。DyPES-VLA 提出分离共享的动力学先验和特定本体的控制策略，从而提升跨具身设备的泛化能力。

Tool use transforms LLMs into agents that act beyond their training data, and for code-capable models, programmatic tool calling extends this further by replacing rigid JSON calls with scripts that ch...

工具使用将 LLM 转变为超越其训练数据行动的智能体。对于具备代码能力的模型，程序化工具调用通过用动态脚本替代死板的 JSON 调用来扩展这一能力。文章探讨了过度依赖结构化 JSON 调用工具的局限性。

Deciding which of two agents is stronger means playing games until skill outweighs luck, and every game costs money, model inference, or expert time. Since the number of games needed is unknown, fixed...

判断两个智能体哪个更强需要进行多局博弈，直到技巧胜过运气，而每局游戏都有算力和时间成本。由于不知道具体需要多少局，固定数量的测试往往造成浪费。AV-AIVAT 提出了一种提前停止机制，将评估成本降低了 74 倍。

We give a formal mechanism design model for the continuous participatory governance of a deployed AI agent. The mechanism is built on the principle that governance should control an AI agent through r...

研究提出了一个正式的机制设计模型，用于持续参与式治理已部署的 AI 智能体。该机制建立在一个原则上：治理应通过资源授权（如审批权限的分配）来控制 AI 智能体，而不是单纯依靠预设规则。

Multilingual reasoning transfer is crucial for extending reasoning capabilities of large language models (LLMs) beyond high-resource languages. On-policy self-distillation (OPSD) and its variants have...

多语言推理迁移对于将 LLM 的推理能力扩展到高资源语言之外至关重要。策略自蒸馏（OPSD）严重依赖高质量的推理轨迹。该研究提出 RP-OPSD，利用推理枢轴来引导策略自蒸馏，提升低资源语言的推理能力。

LLM-based agentic systems have shown remarkable capabilities in complex domains, while suffering from cascading errors and difficulty in debugging. Critical error detection aims to locate the earliest...

基于 LLM 的智能体系统在复杂场景中展现出强大的能力，但容易产生级联错误且难以调试。错误检测旨在定位导致失败的最早偏差，TRAJDEBUG 提供了一种追踪错误生命周期的方法来识别长周期任务中的关键故障。

From natural-language query interfaces to automated report generation, data analysis tools need a description of the data: the real-world entities it contains, which columns function as measures or id...

数据分析工具需要了解数据的描述（实体、度量、维度等）才能工作。Tytan 提出了一种交互式神经符号方法，能够从关系型数据库中自动构建分析的语义模式，提升自动化数据分析的准确性。

Task-oriented conversational agents are evaluated using curated or automatically generated benchmarks, yet benchmark quality is rarely assessed. Poor benchmarks may contain inconsistent tasks, simplis...

任务导向的对话智能体通常使用精心策划或自动生成的基准进行评估，但基准本身的质量却很少被评估。糟糕的基准可能包含不一致的任务或过于简单的测试，该研究提出了一种评估现有测试基准质量的方法。

Large language models (LLMs) increasingly support complex professional tasks, yet their capabilities in rule-intensive document review remain insufficiently evaluated. National standard documents, suc...

大语言模型（LLM）越来越多地支持复杂的专业任务，但它们在规则密集型文档审查方面的能力仍未得到充分评估。该研究针对国家标准文件（如 lengthy 技术规范）对 LLM 进行了测试，并提出了增强其审查能力的方法。

Recent advances in reward modeling show a paradigm shift from discriminative reward models to generative reward models. However, despite their strong capabilities in response ranking, generative rewar...

奖励建模的最新进展显示出从判别式奖励模型向生成式奖励模型转变的趋势。然而，尽管生成式模型在响答排序方面具有很强的能力，但在强化学习（RL）中构建密集奖励仍面临挑战。RRC 提出了一种基于排序的方法来解锁这一限制。

Retrieval-augmented generation over long documents is dominated by one design: chunk the text, embed the chunks, and surface the top-k nearest neighbours of the query. We argue that for an important c...

长文档上的检索增强生成（RAG）被一种模式主导：分块、向量化，然后返回与查询最相近的前 K 个结果。研究者认为，对于长文档的复杂查询，单纯依赖向量相似度的黑盒检索是不够的，提出使用可解释的智能体操作来替代。

As LLMs are increasingly deployed within agentic systems, their capabilities depend not only on the model weights but also on the harness: the prompts, tools, control flow, memory, and orchestration c...

随着 LLM 越来越多地部署在智能体系统中，其能力不仅取决于模型权重，还取决于外围框架（Harness）：提示词、工具、控制流、记忆和编排代码。该基准专门评估 LLM 优化这些框架组件的能力。

On-policy (Self-)Distillation (OPD / OPSD) has shown strong potential for post-training large language models (LLMs). However, existing methods still rely heavily on external supervision, including gr...

策略（自）蒸馏（OPD / OPSD）在大型语言模型（LLM）的后训练中展现出强大潜力。然而，现有方法仍严重依赖外部监督，包括人工标注或更强模型的输出。该研究提出一种完全无需外部监督信号的策略自蒸馏方法。

In this tutorial, we explore adaptive experimentation using Meta’s Ax with the modern Client API. We work through a complete workflow where we tune a RandomForest model on a synthetic classification d...

本文提供了一个实用教程，介绍如何使用 Meta 开发的 Ax 平台及其现代 Client API 进行自适应实验。教程通过在合成分类数据上微调 RandomForest 模型的完整工作流，展示了如何高效地进行超参数优化和贝叶斯优化。

Retrieval-augmented generation (RAG) improves question answering by grounding large language models (LLMs) in external knowledge such as text corpora. However, its reasoning process remains largely op...

检索增强生成（RAG）通过将大型语言模型（LLM）锚定在外部知识库（如文本语料库）来改善问答。然而，其推理过程很大程度上仍是一个黑盒。NeSy-RAG 结合了神经计算与符号逻辑，使得 RAG 的推理链条变得可解释。

The "thinking-with-images" paradigm equips multimodal LLMs with active visual operations such as crop-and-zoom. However, models using these operations often achieve only marginal or negative gains ove...

“图像思考”范式为多模态大模型配备了裁剪和缩放等主动视觉操作。但因果审计发现，使用这些操作往往只能带来边际甚至负面的增益，模型可能只是学会了走捷径。

Synthetic clinical benchmarks for enterprise AI agents can pass existing utility checks and still remain structurally unrealistic, especially in privacy-sensitive healthcare settings where operational...

面向企业级 AI 智能体的合成临床基准测试，即使通过了常规效用检查，在隐私敏感的医疗场景中仍可能存在结构不真实的问题。本文探讨了如何在满足数据脱敏和效用约束的前提下，生成更具真实业务代表性的临床测试基准。

Current video world models struggle in multiplayer environments because they entangle world state with view-dependent visual latents, leading to redundant compute, view inconsistencies, and poor scala...

现有视频世界模型在多玩家环境中表现不佳，因为它们将世界状态与视角相关的视觉特征耦合在一起。本文提出了 MASS，通过引入权威共享状态机制，解耦了全局状态与玩家视角，解决了多人场景下的计算冗余和视角不一致问题。

Post-training adaptation has become central to modern machine learning practice and includes techniques such as retraining, fine-tuning, parameter-efficient adaptation, alignment, retrieval augmentati...

训练后适配已成为现代机器学习实践的核心，包括微调、参数高效适配、对齐、检索增强等技术。该研究提出了一个六维度分类法来系统化这些技术，并探讨了其在 AI 合规与治理中的应用。

Reinforcement learning with verifiable rewards (RLVR) improves the reasoning capabilities of large language models using automatically verifiable outcome signals, but these signals are typically spars...

使用可验证奖励的强化学习（RLVR）依赖稀疏的结果信号，这在复杂任务中训练效率低下。DASH提出了一种发散自适应监督视野方法，用于推理模型的在线自蒸馏。

Despite advances in artificial intelligence (AI) across multiple sectors, today's AI tools, including deep learning and generative AI, still fail when embedded into physical systems, such as robots an...

尽管人工智能发展迅速，但现有的 AI 工具在嵌入机器人和工业控制等物理系统时仍易失效。本文提出“整子数字孪生”概念，将传统的被动镜像转变为能够主动参与物理网络协同的智能体，以解决复杂系统中的实时控制难题。

While deep learning models, particularly transformer-based architectures, have shown impressive performance in time series forecasting, the application of retrieval-augmented generation (RAG) in this ...

虽然深度学习在时间序列预测中表现出色，但检索增强生成（RAG）的应用仍较少。TS-RAG探索了如何将RAG技术有效应用于时间序列预测任务中。

We study Muon, a recently proposed matrix-aware optimization method, in the context of the Stiefel manifold. This manifold consists of matrices with orthonormal columns and is ubiquitous in machine le...

研究者在 Stiefel 流形（由具有标准正交列的矩阵构成，在机器学习中很常见）的背景下研究了最近提出的矩阵感知优化方法 Muon，并推导出了精确的闭式更新公式。

Large language model (LLM) benchmark evaluations are routinely used to support claims about model safety, reliability, and deployment readiness. Yet most evaluations rely on a single access modality (...

大型语言模型（LLM）的基准测试常被用来证明模型的安全性，但大多数测试仅依赖单一的访问模态（如纯文本API）。本文探讨了模态、搜索和引用缺失对安全评估的影响。

Training large language model agents for long-horizon tool use typically relies on interactions with real or synthesized executable environments, whose construction and verification are costly, or on ...

训练大语言模型智能体进行长时序工具使用时，依赖真实或合成执行环境的成本极高。EnvACE 提出通过“世界推演”机制，让模型在内部记忆中内化环境的动态变化规律，从而减少对外部真实环境的高频依赖。

Agents backed by large skill libraries must decide which skills to load and in what order. Loading the entire library into context is expensive and provides no structure for autonomous sequencing. We ...

拥有庞大技能库的智能体必须决定加载哪些技能以及按何顺序执行。将整个库塞入上下文不仅成本高且缺乏自主排序能力。本文对比了不同的检索策略，探讨了如何高效地从大型代码/技能库中提取并结构化编排执行动作。

Large-language-model inference is a fast-growing electricity load whose marginal carbon intensity varies by more than an order of magnitude across grid regions and across the day, making request place...

大语言模型（LLM）推理是一个快速增长的电力负载，其在不同电网区域和不同时间的边际碳强度差异超过一个数量级。该研究提出通过请求路由分配来降低碳排放。

Web agents observe a browser through text, pixels, or both, and the choice is usually fixed once for all tasks. We measure six observation modes across eight site-model combinations (cells) on VisualW...

Web 智能体通过文本、像素或两者的结合来观察浏览器，而这种选择通常是全局固定的。研究测量了不同观测模式在多个网站和模型组合下的表现，发现在任务最复杂、价值最高的地方，静态观测路由往往最难学习。

Synthetic 3D scene generation is increasingly used as a data source for computer vision and embodied AI, but existing generators often optimize perceptual realism without reliably satisfying task-crit...

合成 3D 场景广泛用于计算机视觉和具身 AI 的训练，但现有生成器往往只顾视觉真实性，无法严格满足任务特定的空间约束。iARCS 引入迭代式的智能体强化学习，让生成过程可控，确保产出的 3D 场景精准符合任务需求。

Automated skill evolution enables Large Language Model (LLM) agents to continuously improve without expensive retraining. However, existing approaches typically treat skill evolution as a sequence of ...

该研究提出了一种自动化的技能演进方法，使大型语言模型（LLM）智能体能够在不进行昂贵重新训练的情况下持续提升编码能力，将技能演进视为一个全局的、可复用的过程。

Most agent benchmarks evaluate tasks independently and cannot measure whether experience from one task helps with later tasks. Existing self-evolution benchmarks do not jointly cover professional work...

FinEvo-Bench 是一个新型基准测试，用于评估AI智能体在专业金融工作流中的长期经验积累能力，填补了现有基准无法衡量跨任务经验复用性的空白。

AI agents performing cryptographic operations (signing Git commits, authenticating API calls, issuing certificates) currently store private keys in software-accessible locations: plaintext files, envi...

针对AI智能体在执行加密操作（如签署Git提交、API认证）时的私钥存储安全问题，本文提出了一种零信任的MCP（模型上下文协议）执行架构，利用硬件密钥库进行防护。

Search agents extend large language models beyond static parametric memory by enabling them to acquire and use ex ternal evidence during multi-step reasoning. For knowledge intensive tasks involving c...

针对搜索智能体在处理知识密集型多步推理任务时的问题，该研究提出了一种上下文信息策略优化方法，以更好地在推理过程中获取和利用外部证据。

CLI-based software-engineering agents have matured rapidly, yet the open ecosystem has converged on a single training environment: trajectory datasets used to fine-tune open models are collected almos...

为了打破命令行（CLI）软件工程智能体训练环境的单一性，DCAS提出了解耦脚手架的方法，使模型能够将规划能力内化，从而在不同的CLI执行环境中具备泛化能力。

Hospitals are rapidly adopting artificial intelligence for triage, imaging, scheduling etc., yet most deployments remain isolated point solutions locked inside departmental silos, resulting in duplica...

面对医院内部AI应用各自为政、数据孤岛严重的现状，本文提出了一种合规优先的多层级Agent平台架构，旨在统一管理分诊、影像和排班等AI工作流。

This paper presents ECHO (Enhanced Care \& Health Observer), a locally-deployable conversational health assistant for long-term chronic care management. ECHO integrates three complementary software mo...

ECHO是一款专注于长期慢性病管理的本地化健康对话助手。它集成了时间记忆、安全护栏以及语音评估三大模块，强调数据隐私和长期照护。

Investment competence is inherently personalized: the same market evidence can justify different actions for investors with different goals, horizons, portfolios, and risk boundaries. Yet financial LL...

由于投资者的目标、投资期限、投资组合和风险承受能力各不相同，对相同市场证据的合理反应也不同。该研究建立了一个评估LLM个性化投资逻辑的真实基准。

Commercial vision-language models are reshaping computer vision, with visual priors broad enough to rival task-specific systems. This raises a natural question: do they reduce the need for classic, ph...

随着通用视觉语言模型（VLM）的崛起，它们广泛的视觉先验已经能媲美特定的专用系统。本文探讨了在Agent进行图像编辑（如阴影去除）时，是否还需要经典的物理特征处理方法。

GUI agents are commonly trained offline from successful interaction trajectories. Standard training decomposes each trajectory into prefix-action pairs: the agent predicts an action from the current s...

针对移动端图形界面（GUI）智能体离线训练效率低下的问题，本文提出了一种门控事后见蒸馏方法，利用下一屏幕的状态信息来优化当前动作的预测。

Tool-calling agents infer task state from accumulated dialogue and tool traces. In persistent interactions, however, historical traces may remain structurally valid and semantically plausible after th...

智能体通常从对话和工具历史轨迹中推断任务状态，但在持久交互中，历史记录在任务状态改变后可能依然保持结构有效和语义合理，从而产生误导。本文对此进行了评估与改进。

AI-agent workflows often involve remote calls to models, memory stores, and tools distributed across a network. As execution progresses, these dependency calls collectively form an agentic service gra...

AI智能体工作流通常涉及对网络中分布式模型、内存和工具的远程调用。随着执行推进，这些依赖调用形成了一个服务图，ASGE-RR提出使用带可修改预留的图嵌入来优化这些动态调用。

In India, almost every social benefit starts with a form, yet the people who need these benefits most are often unable to read or write. Reaching them requires a spoken conversation. Today that work f...

在印度，几乎所有社会福利申请都需要填表，但最需要这些福利的人往往不识字。FormBharo设计了一个语音助手，通过对话方式帮助农村用户完成表格填写。

Economic World Models (EWMs) are generative economic models that simulate how economies evolve from within by modeling heterogeneous agents, their beliefs and actions, and the market and institutional...

经济世界模型（EWMs）是一种生成式经济模型，通过模拟异构智能体、其信念和行为以及市场和制度结构，从内部模拟经济体的演变。本文提出了一种系统蓝图。

Search-agent rewards mix answer quality, citation grounding, tool cost, and anti-hacking terms; a high score therefore need not imply that cited evidence was retrieved, and added penalties can cancel....

搜索智能体的奖励机制通常混合了答案质量、引用依据和工具成本等，导致高分不一定代表真正检索到了引用证据。HERALD提出了一种反事实审计和最小修复机制。

Modern LLM infrastructure increasingly manages tensors not only as computation data, but also as persistent states shared across distributed components. Existing systems optimize individual tensor man...

现代大型语言模型（LLM）基础设施不仅将张量作为计算数据，还将其作为分布式组件间共享的持久状态。TensorCast填补了现有系统在张量管理层面的空白。

Vision-language-action (VLA) models have demonstrated remarkable capabilities in robotic manipulation by leveraging pretrained vision-language models. However, existing post-training methods predomina...

尽管视觉-语言-动作（VLA）模型在机器人操作上表现优异，但现有的后训练方法多为扁平化的策略。本文提出了一种分层后训练方法，以提升实体智能体的长时序任务处理能力。

Autonomous agents choose actions using scores that may not reflect experimental success. We developed OPERA, an operator-residual framework for optical experiments. It represents experimental actions ...

自主智能体通常依据打分选择动作，但这不能反映实验的真实成功率。研究团队开发了OPERA框架，利用算子残差来表示实验动作，以实现更可靠的自主光学实验。

Reinforcement learning (RL) with verifiable rewards constructs trajectory-level advantage estimates, yet it often fails to credit the few pivotal decisions that determine outcomes in long-horizon, mul...

在具有可验证奖励的强化学习（RL）中，通常进行轨迹级别的优势估计，但很难在长时序多步操作中归因少数关键决策。AgentOPSD引入了递归自蒸馏来解决这一问题。

Large language models (LLMs) excel in structured tasks but struggle with dynamic social interactions, where success requires long-term goal coordination and rapid adaptation. Current methods often app...

大型语言模型（LLM）在结构化任务中表现出色，但在动态社会互动中表现挣扎，这需要长期目标协调和快速适应。当前方法往往采用短视策略，该研究提出通过分层推理结合话语级目标奖励来提升模型的社会智能。

Prime Intellect has open-sourced Prime Agent, a coding and research harness built on two abstractions: the Recursive Language Model, which turns sub-agent calls into functions inside a persistent IPyt...

Prime Intellect 开源了 Prime Agent 编码与研究框架。其核心是将子智能体的调用转变为持久 IPython 内核内的函数调用（递归语言模型），从而大幅减少上下文窗口的消耗并提升复杂任务的执行效率。

Vision-Language-Action (VLA) models have become the dominant recipe for generalist manipulation, yet they are almost universally trained by behavior cloning: a policy imitates expert action chunks con...

针对现有视觉-语言-动作（VLA）模型大多通过行为克隆进行训练的局限，本文提出通过上下文后训练和智能体工具使用，赋予VLA模型更强的语言理解与交互能力。

As capabilities rapidly increase, AI agents can move from running inside one app to acting across a user's devices over time. Yet existing agent systems still fall short in this scenario. This is beca...

随着 AI 智能体能力的快速提升，它们正从单一应用内向跨用户设备协作演进。现有系统在跨设备场景下仍存在明显不足，主要原因是缺乏统一的上下文管理和跨端调度机制。

World models enable agents to perform forward rollout and planning without real-world interaction. However, their application in open-world embodied intelligence remains limited by the high cost of ac...

世界模型使智能体无需真实世界交互即可进行前瞻推演。LAWM-3D通过从人类视频中学习3D感知的潜在动作，降低了数据获取成本，提升了机器人世界模型在开放环境中的泛化能力。

Smart-home assistants increasingly use multimodal large language models (MLLMs) that perceive video and audio directly. This raises a safety question specific to the home: can the agent tell a genuine...

智能家居助手越来越多地采用多模态大模型（MLLMs）直接感知音视频。这引发了特定的安全问题：智能体能否区分真实用户指令与隐藏在视频/音频流中的恶意提示注入？PromptShield Home 旨在提供一种环境防御机制。

Most coverage of Microsoft's SkillOpt centers on its 52/52 result. The more consequential finding is in Section 4.3: the exported best_skill.md keeps working in environments it was never trained on. A...

微软的 SkillOpt 表明，智能体导出的最优技能文件（如 best_skill.md）可以在未经训练的新环境（如从 Codex 迁移到 Claude Code）中继续生效。这证明了 Agent 技能具备跨模型和跨框架的通用性。

Procedural generators produce useful verifiable reasoning problems at scale, but have received less attention as data for completion-supervised fine-tuning. We introduce Reasoning Core, a collection o...

程序化生成器可以大规模生成可验证的推理问题，但很少被用于补全监督微调。本文推出了 Reasoning Core 数据集，利用程序化生成的数据结构，专门用于训练模型在已知前文的情况下补全后续严密的推理逻辑。

Long-horizon reasoning requires an agentic runtime that can persist when evidence supports its current approach and pivot when measurements reveal failure, hidden constraints, or a misspecified object...

Argus 是一种通用的智能体运行时环境，专为长周期推理设计。它能够支持智能体在当前策略有证据支持时坚持执行，而在测量结果显示失败、隐藏约束或目标错误时灵活转向，具备高鲁棒性。

Large language model agents are increasingly being developed to control a wide range of scientific characterization tools including microscopes and synchrotron beamlines. Research into agentic control...

大语言模型智能体正越来越多地被开发用于控制包括显微镜和同步加速器光束线在内的科学仪器。研究表明，智能体控制系统在特定资质认证上表现良好，但并不一定能泛化到未见过的科研任务中。

Context lengths of language models (LMs) have dramatically increased, driven by the demands for in-context learning, self-improvement, and long-horizon agentic workflows. Existing long-context corpora...

OctoLong 通过在跨代码库的代码上下文上进行中期训练，增强了语言模型的长上下文建模能力。由于现有长上下文语料库缺乏复杂的项目级代码关联，该方法填补了模型在处理大型复杂代码库时的性能空白。

Long-horizon reasoning in recent LLMs demands that the model switch between distinct skills inside a reasoning chain, such as first doing a math derivation, then using the result to plan a schedule. W...

长程推理要求模型在推导过程中切换不同技能（例如先算数学再用结果做日程规划）。本文提出了“技能熵”指标来评估模型在推理链中的技能切换能力，并介绍了如何通过训练提升模型管理多种底层技能的能力。

Modern Greek is absent from NVIDIA's Nemotron retrieval models and from major multilingual retrieval benchmarks, despite being important for retrieval-augmented generation (RAG) in legal, energy, fina...

本文详细介绍了如何为 NVIDIA 的 Nemotron 检索模型添加现代希腊语支持。研究涵盖了从法律、能源和金融等垂直领域的语料挖掘，到 RAG 检索适配和生成落地的全过程，填补了主流多语言基准的空白。

Spoken Language Understanding (SLU) is the core component of task-oriented dialogue systems and a pivotal link in achieving seamless human-agent interaction. While traditional SLU can effectively extr...

口语理解（SLU）是对话系统的核心，传统方法需要先将语音转录为文本。本文为大型音频语言模型提出了“语音函数调用”的新范式，允许模型直接从音频流中提取意图并触发外部 API 调用，实现更无缝的人机交互。

Long context reasoning in large language models (LLMs) is usually constrained by the fact that a single inference trajectory has to simultaneously explore the context, store intermediate state, verify...

为突破单次推理轨迹在探索上下文、存储中间状态和验证结果上的限制，本文提出了链式递归语言模型。它将长上下文推理拆解为多轮迭代，通过递归调用模型来解决单一推理过程难以应对复杂逻辑的问题。

Prompt injection poses significant security risks to LLM agents. Efficient and effective red-teaming is therefore critical, both for evaluating these risks and for collecting training data to improve ...

提示词注入对大语言模型智能体构成巨大安全威胁。为了高效评估风险，本文构建了一个对抗性智能体系统，能够自动化生成并执行各种复杂的提示词注入攻击，用于收集训练数据以提升模型的安全防御能力。

Long-horizon search agents must make multiple sequential actions (steps) to search, retrieve, verify, and integrate evidence to reach a final answer. However, existing methods for training these agent...

ABSeeker 针对需要搜索、检索、验证和整合证据的长周期搜索智能体，提出了一种基于答案回溯的信用分配方法。该方法有效解决了现有训练方法难以对多步骤顺序搜索行为进行准确奖励分配的问题。

Agents for long term reasoning require a memory that can be efficiently and effectively updated over time, as new facts and external feedback continue to arrive. Recently, graph memory has been adopte...

本文提出了一种面向长期推理智能体的层级图记忆机制。随着新事实和外部反馈的不断涌入，该机制允许对图谱进行路径级的高效更新和重写，克服了传统图记忆难以应对长周期动态变化的缺陷。

Muon has recently emerged as a promising alternative to AdamW for language model pretraining by orthogonalizing momentum matrices using Newton-Schulz iterations. Although Muon mitigates gradient aniso...

Muon 近期作为 AdamW 的替代方案用于语言模型预训练，它通过牛顿-舒尔茨迭代对动量矩阵进行正交化。尽管缓解了梯度各向异性，但计算开销大。MALT 引入轻量级的对角预条件机制，在保持优势的同时大幅降低了计算成本。

Systems that automate scientific discovery must repeatedly decide which experiment to run, which hypothesis to test, which tool to build, and when to stop. Many systems make these decisions by maximiz...

自动化科学发现系统必须反复决定运行哪个实验、测试哪个假设以及何时停止。当前许多系统通过最大化即时奖励来做决策。该研究指出了这种“短视”策略的局限性，并提出了一种基于“能力门控”的规划方法，通过评估走向目标的成本来更合理地选择实验。

Critic-free group-based reinforcement learning has become a scalable approach for post-training large language models. However, most existing methods allocate the same number of rollouts to every task...

无 Critic（价值网络）的强化学习是后训练大语言模型的主流方式，但现有方法往往为所有任务分配相同数量的探索次数。本文提出恢复性感知的干预学习，动态调整分配给不同难度任务的 Rollout（策略推演）次数，提升数据利用效率。

Small open-weight language models increasingly run in private, offline, and cost-sensitive settings, where the key deployment question is not only what a model answers but when it should defer to a hu...

在私有化或离线部署的小型开源语言模型时，系统不仅需要知道模型何时能给出答案，还要知道何时该把问题转交给人类。该研究为模型“语言化表达不确定性”提供了理论界限，并提出了一种认证延迟机制，确保模型在不自信时安全移交给人类。

The semiconductor industry is undergoing a dual revolution: the shift toward heterogeneous 2.5D chiplet systems and the integration of Large Language Models (LLMs) into Electronic Design Automation (E...

半导体行业正经历双重变革：向2.5D异构芯片系统转移，以及将大语言模型（LLM）集成到电子设计自动化（EDA）工具中。本文探讨了在这一背景下硬件设计流程的重构以及潜在的安全挑战与应对策略。

Released aligned large language models remain vulnerable to malicious downstream finetuning. Existing defenses are largely designed for the fine-tuning-as-a-service (FTaaS) paradigm or rely on downstr...

开源的对齐大语言模型极易受到恶意下游微调的攻击。现有的防御手段大多依赖服务器端或高成本算法。本文提出“梯度免疫”方法，在模型发布前将其参数映射到特定零空间，使得恶意微调的梯度无法破坏模型原有的安全对齐。

Sparse matrix kernels are fundamental to scientific computing, graph analytics, and machine learning. Their GPU performance depends strongly on the input sparsity pattern and execution strategy. For t...

SparseDitto 利用基于 LLM 的智能体系统，为机器学习和图分析中的稀疏矩阵自动定制 GPU 算子。由于 GPU 性能高度依赖于输入的稀疏模式和执行策略，该系统能有效解决传统手工调优难以覆盖所有模式的痛点。

High-quality annotation of artworks is essential for computational art research, yet extracting implicit semantics remains challenging due to the reliance on culturally grounded meanings and deep cont...

ArtAnno 是一个艺术品语义标注框架，通过大模型 Agent 驱动双向人机协同，提取艺术品中深层的、依赖文化背景的隐式语义，从而为计算艺术研究提供高质量数据。

Analog circuit design automation using reinforcement learning (RL) has emerged as a promising approach for reducing manual effort. However, many existing RL-based methods focus on single-objective opt...

强化学习（RL）已被用于减少模拟电路设计的手动工作量，但现有方法多局限于单目标优化。本文提出了 ORACLE 优化器，结合大语言模型（LLM）引导的探索机制，解决了模拟电路设计中需要同时权衡功耗、面积等多个冲突目标的问题。

Large language model (LLM) agents repeatedly process long, privacy-sensitive contexts, while cloud-only deployment exposes user data beyond the trusted endpoint and fully local deployment often requir...

RAC（Reference-Aware Activation Compression）针对隐私敏感的长上下文 LLM 智能体，提出在云和端之间拆分推理。通过参考感知激活压缩技术，在保护本地用户数据的同时，大幅降低了端云通信的数据量。

We show that tiny transformers can profitably employ a simple form of Chain of Thought, which we call protoreasoning, allowing us to study step-by-step reasoning on ~1M-parameter models and opening up...

研究表明，微型 Transformer（约100万参数）能够有效利用一种简单的思维链形式，研究称之为“原初推理”。这为在小模型上研究逐步推理机制打开了大门，使得研究人员无需庞大算力即可探究模型内部的逻辑推演过程。

SciCode is the standard measure of the scientific-coding ability of language models: research-level problems that demand both frontier scientific theory and its implementation as working numerical cod...

SciCode 是衡量大语言模型编写科学研究级数值代码能力的标准基准。该研究发现原基准测试存在答案校验缺陷，导致低估了模型的实际编程能力。研究团队修复了这些缺陷，并重新评估了当前主流模型的科学编码实力。

The capabilities of an LLM agent depend not only on its model but on the harness: the executable program that constructs context, invokes tools, verifies results, and recovers from failure. Recent wor...

EvolveNet 提出，LLM Agent 的能力不仅取决于底层模型，还取决于构建上下文、调用工具的“框架”。该系统通过让 Agent 自动进化其执行框架，从而实现自我改进，提升从失败中恢复的能力。

Interactive video world models are essential for long-horizon planning and exploration, yet they suffer from compounding errors. Post-training methods such as reinforcement learning (RL) can improve t...

交互式视频世界模型对于长时序规划至关重要，但容易产生复合误差。本文提出了 WorldCycle，一种基于自验证机制的强化学习后训练方法，能够有效减少模型在生成长周期视频或进行多步环境推演时的误差累积。

CheMLFlow is an open-source platform for building and executing end-to-end, high-throughput, and agentic workflows for scientific and technological applications. CheMLFlow targets a common bottleneck ...

CheMLFlow 是一个开源平台，旨在为化学和材料科学应用构建端到端、高通量的智能体工作流。它主要解决了这些垂直科学领域中数据格式转换和工作流编排的常见瓶颈问题。

Training LLM agents commonly relies on supervised fine-tuning from expert trajectories or online reinforcement learning over human-specified tasks with handcrafted verifiers. Though effective, both re...

State2State 提出了一种训练 LLM 智能体的新范式，摆脱了依赖人类专家轨迹进行微调或手工设计验证器的强化学习。它利用环境生成的状态进行中期训练，降低了智能体对人类先验知识的依赖。

Long-horizon embodied task requires agents to act under partial observability while preserving both scene belief and execution progress. Flat histories or implicit policy states may contain past obser...

Mimir 提出了一种具有动态接地机制的神经符号记忆系统，专为具身智能体设计。它能有效处理长时序任务中的局部可观测性问题，将过去的观察转化为场景信念和执行进度，避免了扁平化历史记录的缺陷。

We present PRIMAL3, an ultra-large-scale learning-based framework for multi-agent pathfinding (MAPF) that integrates reinforcement learning, topology-aware communication, LaCAM3-guided training, and P...

PRIMAL3 集成了强化学习、拓扑感知通信、LaCAM3 引导训练和优先路径规划，提出了一种用于多智能体寻路（MAPF）的超大规模学习框架。该框架显著提升了多智能体在复杂环境下的协同寻路效率。

Multi-agent LLM systems relay key--value caches instead of text and credit their gains to exchanged ``latent thoughts''. That credit is a claim about \emph{which} example's cache is relayed, not merel...

本文对多智能体 LLM 系统中直接中继 KV Cache（潜在思想）而非文本的做法进行了因果归因审计。研究澄清了性能提升的真正来源是缓存匹配的具体示例，而不仅仅是由于传递了所谓的“潜在思想”。

Symbolic regression aims to discover closed-form equations from data, but existing LLM-guided methods often rely on a unified proposal loop that compresses heterogeneous search failures into a scalar ...

A-SR 是一种用于符号回归的自进化智能体框架。它通过层级协调机制，解决了现有 LLM 在引导符号回归时将异构搜索失败压缩为单一标量的缺陷，从而更高效地从数据中发现闭环方程。

Long-horizon LLM agents require memory systems that recover useful evidence from large interaction histories without passing excessive context to downstream models. Existing memory pipelines often rel...

针对长周期 LLM 智能体，MemoryCPT 提出了一种端到端记忆框架，能够在不向下游模型传递过多上下文的情况下，从大量交互历史中恢复有用证据，突破了现有记忆管线过度依赖向量检索或摘要的瓶颈。

Memory is essential as language agents move from isolated tasks to long-horizon, stateful workflows, yet existing evaluations often reduce it to retrieval or question answering. We introduce ContextWe...

ContextWeave 评估基准指出，随着语言智能体向长期、有状态的工作流发展，现有的评测往往将记忆能力等同于简单的检索或问答。该基准测试专注于在复杂的多步骤真实业务流中评估 Agent 的记忆和状态维护能力。

Large language model (LLM) agents increasingly rely on skills, structured documents that specify when to act, which procedure to follow, and which tools are allowed. Existing evaluations mostly judge ...

本文探讨了 LLM 智能体能否在实际执行框架中有效利用“技能”——即规定何时行动、遵循何种程序及使用哪些工具的结构化文档。现有评测多停留在表层，该研究深入评估了 Agent 对技能指令的真实执行力和泛化度。

Frontier language models can resolve repository-level software issues, but each attempt is expensive, and existing routers select a model from the issue text alone. We present SuperScout, which routes...

Scrouting 提出了 SuperScout 机制，解决了现有模型路由器仅根据问题文本进行选择的局限性。SuperScout 会先对代码库进行“侦察”，基于代码库的实际状态智能路由模型，从而在解决仓库级软件问题时大幅降低试错成本。

Large language model agents are commonly trained through reinforcement learning with sparse trajectory-level rewards, which offer limited guidance on how strongly individual tokens should be updated. ...

大语言模型智能体通常通过稀疏的轨迹级奖励进行强化学习训练，这种方式无法指导单个 token 的更新强度。本文提出了一种观测校准的自蒸馏方法，使模型能够根据环境的真实观测反馈，更精细地调整局部生成策略。

Move beyond traditional text-based parsing with PixelRAG, an end-to-end system that treats web pages and PDFs as images. This tutorial explores the complete pipeline—from rendering and tiling to multi...

传统的 RAG 系统依赖文本解析，而 PixelRAG 提出了一种端到端的系统，将网页和 PDF 直接作为图像处理。该指南涵盖了从渲染、平铺切割到多模态索引的完整流程，超越了传统 OCR 解析的局限。

A new SaferAI report finds Z.ai's open-weight GLM-5.2 approaches frontier AI capabilities while lacking key safety mitigations, renewing concerns that powerful open models could outpace governance and...

SaferAI 的一份新报告指出，Z.ai 开源权重的 GLM-5.2 模型在能力上已逼近顶尖闭源模型，但缺乏关键的安全缓解措施。这引发了业界对强大的开源模型发展速度超越治理能力的担忧。

Cursor Research has open-sourced Mixture-of-Kittens (MoK), the MoE training megakernel behind its Composer models. MoK fuses all mixture-of-experts communication and computation into a single determin...

Cursor Research 开源了其 Composer 模型背后的训练核心组件 Mixture-of-Kittens (MoK)。这是一种专为新型 GB300 NVL72 GPU 机架设计的混合专家模型训练“巨型内核”，将通信和计算融合为单一的确定性步骤，大幅提升算力利用率。

Large language models (LLMs), and the agents built on top of them, are now benchmarked heavily on whether they can finish a task -- fix a bug, drive a browser, operate a GUI. A complementary social ab...

大型语言模型及其构建的智能体，目前都在被重度评估其完成单一任务（如修Bug、操作GUI）的能力。本文提出了一个互补的社会基准测试SocietyBench，用于评估模型预测反事实社会演变的能力。

Tool-Integrated Reasoning (TIR) enables LLMs to solve complex tasks through iterative tool interactions. However, existing reinforcement learning methods often rely on trajectory-level supervision, li...

工具集成推理（TIR）使大模型能够通过迭代的工具交互来解决复杂任务。然而现有的强化学习方法往往依赖轨迹级监督，本文提出了TurnSight，通过回合级的事后自我蒸馏来优化TIR过程。

The next generation of gravitational wave detectors, such as the Einstein Telescope (ET), will observe orders of magnitude more binary black hole mergers than current facilities. Most of these events ...

探讨了下一代引力波探测器（如爱因斯坦望远镜）在观测到大量黑洞合并事件后，如何利用仿真方法优化天文群体的统计推断。

Recursive self-improvement requires agents to turn accumulated experience into better future behavior. Personal AI agents offer a concrete setting for studying this capability because they retain pref...

递归自我改进要求智能体将积累的经验转化为更好的未来行为。个人AI智能体因为保留了用户偏好，为研究这种能力提供了具体场景。本文提出了PAST-Bench，用于评估个人智能体的递归自我改进能力。

We present Semantic Bundling, a visual analytics technique for making sense of text documents represented as knowledge graphs (KGs). Representing a document corpus as a KG makes relationships between ...

提出了一种名为语义打包的可视化分析技术，利用大语言模型简化从文本文档构建的知识图谱，使实体间关系更易读。

Large language models can solve substantially harder reasoning problems with more inference-time compute. The term "test-time scaling," however, now covers diverse inference algorithms that extend del...

系统总结了目前“测试时扩展”涵盖的多种推理算法，探讨了不同推理机制在提升大模型复杂推理能力上的效果与复现标准。

Text-to-music language models begin with a choice usually made by default: how to tokenize music. Normally entangled with backbone, data, and recipe, its effect has never been measured in isolation. W...

针对文本生成符号音乐的任务，单独测量了音乐Token化方法的影响，提出了包含表演时间信息的Agogic音乐Token。

LLM agents are increasingly applied to open-ended everyday requests that span work, study, and life. These tasks are long-horizon, cross-environment, and multimodal, forcing the agent to preserve goal...

OneDayAgent 针对跨越工作、学习和生活的开放式日常请求设计了一种长周期执行框架。它解决了 Agent 在处理多模态、跨环境且长周期任务时，难以保持目标连贯性和跨平台上下文转换的问题。

We identify a previously overlooked failure mode of ALiBi positional encoding: its linear bias scaling underflows floating-point precision, which zeroes out a large fraction of attention weights and r...

识别了ALiBi位置编码中线性偏置缩放导致的浮点精度下溢问题，该问题会使大量注意力权重归零，从而降低模型的长上下文性能。

Optimizing compilers miss profitable transformations when their enabling semantics are absent from the analyzed program representation. We ask whether large language models (LLMs) can recover such sem...

研究探讨了当程序表征缺乏相关语义时，大语言模型是否能够找回编译器错失的那些有收益的程序转换和优化方案。

We introduce Video-DeepResearch (Video-DR), extending multimodal agents from static images to continuous video streams, a setting that demands dense spatiotemporal grounding coupled with open-web expl...

本文介绍了Video-DeepResearch (Video-DR)，将多模态智能体的处理范围从静态图像扩展到了连续视频流，这种场景要求智能体具备密集的时空接地能力，并结合开放网络探索进行分析。

On-policy training has emerged as a powerful post-training paradigm for improving the reasoning capabilities of large language models, and is often enhanced by golden trajectories from stronger expert...

提出了一种在策略训练范式，不仅利用专家提供的正样本，更着重从反思性的负样本轨迹中学习，以提升大模型的推理能力。

Large language models can generate fluent Arabic answers while introducing factual errors that are difficult to identify and verify. Existing Arabic hallucination resources often assign a binary label...

针对大模型生成的流利但存在事实错误的阿拉伯语回答，提出了一个细粒度的标注语料库和流程，用于幻觉检测和真实性验证。

Test-time scaling improves LLM reasoning by generating and aggregating multiple candidate answers, yet many pipelines use fixed per-query budgets that spend the same compute on easy and difficult prom...

现有的测试时扩展（Test-time scaling）为每个查询分配固定的计算预算，本研究提出自适应采样方法，根据提示词的难易程度动态调整计算量。

As autonomous agents powered by foundation models are increasingly integrated into social and economic systems, understanding the principles governing their collective behavior is essential for ensuri...

随着基于基础模型的自主智能体日益融入社会和经济系统，理解其集体行为的法则对于确保AI协同至关重要。本文提出了一种基础模型博弈论，通过相似性推断展示了智能体走向理性合作的新路径。

Large language models (LLMs) are increasingly used to provide conversational practice for English-as-a-second-language (ESL) learners. Effective ESL tutoring, however, requires more than fluent respon...

指出当前大模型在ESL（英语作为第二语言）教学中缺乏教学逻辑，提出了一种基于教学法分类对齐的后训练框架，使模型能自适应调整难度。

Pre-pretraining language models (LMs) on symbolic data can accelerate and improve natural language acquisition. However, existing pre-pretraining tasks, such as Dyck and procedural algorithms, rely on...

研究发现，在自然语言预训练之前，先让语言模型在形式逻辑推导数据上进行学习，不仅能加速语言习得，还能提升模型的压缩率。

When will robots have their ChatGPT moment? Such a breakthrough requires a general-purpose robot that can handle unfamiliar tasks in unfamiliar environments, remain controllable over long interactions...

机器人何时能迎来自己的ChatGPT时刻？这样的突破需要一种通用机器人，能够在陌生的环境中处理不熟悉的任务，并在长时间的交互中保持可控性。本文提出了一种名为ETA的具身任务智能体新范式。

As AI systems are deployed across increasingly diverse social contexts, alignment can no longer be framed as the optimization of a single, unified set of values. Instead, systems must be able to recog...

随着AI系统在多样的社会场景中部署，对齐不再是优化单一的价值观，而是需要系统识别并协调多元甚至冲突的视角。

We study linear representations of temporal horizon in the large language model Qwen3-32B and use them to change the model's time-related preferences, recommendations, and capabilities. We train contr...

研究在Qwen3-32B中提取了模型时间跨度的线性表示，并通过对比激活添加的方法，成功改变了模型在时间相关的偏好和推荐能力。

Geospatial reasoning, i.e., computing distances, containment, and other spatial relations over real-world entities, is central to navigation and logistics, yet large language models (LLMs) struggle wi...

提出了一个多语言地理空间推理基准，发现当前大模型在计算距离、包含关系等真实世界空间逻辑上依然存在很大困难。

Designing effective reward functions remains a major bottleneck in Reinforcement Learning (RL). Recent work uses large foundation Vision-Language Models (VLMs) as reward models, computing text-observa...

针对强化学习中设计奖励函数困难的问题，该研究使用大型视觉-语言基础模型作为奖励模型，并通过结构感知微调提升了其准确性。

Modern agent frameworks equip large language models with external skill libraries to solve complex tasks. However, it remains unclear whether these systems can effectively evolve their skills and whet...

针对带有外部技能库的Agent系统提出了基准测试，评估其能否有效更新和演化自身技能以解决新出现的问题。

Red Hat has launched asago, an open-source community project that aims to turn AI governance policy into production-ready deployment code. The project describes itself as an automated, auditable workf...

红帽推出了开源社区项目 asago，旨在将 AI 治理政策转化为生产级部署代码。该项目主打自动化、可审计的工作流。（注：治理政策转代码是指将企业的合规要求直接写成代码拦截器）

We introduce SciRet, a compute-aware empirical study of retrieval-augmented generation for scientific question answering over CORD-19. Rather than proposing a new model, we evaluate a fixed scientific...

本文介绍了 SciRet，一项针对 CORD-19 科学问答的检索增强生成（RAG）的计算感知实证研究。研究没有提出新模型，而是评估了固定的科学模型在计算成本与检索质量之间的权衡。

Memory-augmented LLM agents rely on rich context for long-horizon reasoning and acting, yet their memory modules expose a persistent attack surface for malicious records, making the study of memory po...

研究指出具有记忆模块的LLM Agent在处理长期上下文时，面临恶意记录注入的攻击风险，攻击者仅需通过查询即可操纵其记忆。

Training-free speculative decoding drafts by matching an exact suffix of the context against a pool of earlier context. That lookup misses correct drafts already in the pool, most visibly on tool-call...

Oilbird 提出了一种无需训练的投机解码方法，通过匹配上下文的精确后缀来生成草稿。它解决了以往方法在工具调用时容易遗漏正确草稿的问题，提升了大模型推理的效率。

Reasoning-based guard models improve LLM safeguards, but decoding explicit rationales for every interaction makes them costly to deploy. Although latent-reasoning methods reduce token generation by mo...

LatentGuard 提出了一种基于潜在推理的安全防护模型。传统的推理型护栏因为每次交互都要解码显式逻辑而成本高昂，该方法通过减少 Token 生成来降低防御部署成本，同时保持可检查性。

A framework that persists execution state so a run can be interrupted, survive a crash, and continue must decide what a resume means for effects that already fired. Five widely deployed agent workflow...

提出了一个框架，明确了当Agent工作流崩溃重启时，对于已经触发过的副作用（如网络请求）该如何正确处理恢复语义。

AI agents are expected to play an increasingly important role in future decision-making systems. In this paper, we consider collaborative systems composed of heterogeneous multi-agent systems (MAS), w...

研究探讨了异构多智能体系统（MAS）如何在复杂决策中利用历史信息进行元策略委派，以提升协作效率。

An LLM-based agent is a loop that reads itself. Agentic frameworks externalize identity, memory, and disposition into editable files. The agent loads and edits these files during each activation. I ar...

分析了基于LLM的智能体如何通过在每次激活时加载和编辑外部文件来具象化其身份、记忆和性格，从而形成一种AI基础设施。

Video anomaly understanding (VAU) focuses on comprehensively interpreting abnormal events in videos, requiring models to identify anomalous occurrences, discover their supporting evidence, and explain...

提出了一种多智能体框架，通过探索和验证推理机制，帮助模型更全面地识别和解释视频中的异常事件。

Masked diffusion language models (MDLMs) enable parallel generation and bidirectional context modeling, but their positional context differs fundamentally from that of autoregressive (AR) models. Wher...

掩码扩散语言模型（MDLM）支持并行生成和双向上下文建模，但其位置编码逻辑不同于传统的自回归模型。本文提出了 MDLMPE，一种分布感知的位置编码方法，有效弥合了 MDLM 与传统模型在位置上下文理解上的差异。

Agent self-evolution updates an agent's persistent state from prior experience and reuses it to solve related tasks more effectively. Evaluating self-evolution is difficult: existing benchmarks provid...

Agent 的自我进化是指其从过往经验中学习并更新持久状态，以更有效地解决后续相关任务。由于现有基准测试多为静态环境，GDPevo 提出了一种新的评估方法，专门衡量 Agent 在真实业务任务中的自学习和经验积累能力。

Zero-shot 3D visual grounding aims to localize specific objects based on textual descriptions and 3D visual input. However, the effectiveness of existing methods is significantly hindered by the ambig...

零样本3D视觉定位旨在根据文本描述和3D视觉输入定位特定物体。然而现有方法常常受到文本歧义的阻碍，本文提出了TDVR方法，联合进行文本消歧和视点推理，以提升定位效果。

Learning long-horizon manipulation skills with reinforcement learning remains challenging due to the complexity of reward design, the limited guidance of sparse rewards, and the high cost of manual su...

使用强化学习训练机器人执行长时序操作任务面临奖励设计困难、稀疏奖励引导不足等问题。GORDON 提出了一种基于图的、以对象为中心的奖励机制，将复杂的长步骤任务进行分解，从而提升模型的学习效率。

Chain-of-Thought (CoT) reasoning offers a promising window into model monitoring. However, monitoring relies on faithfulness, i.e., the model output strictly derives from its reasoning trace. We ident...

思维链（CoT）被认为是监控大模型推理的有效手段，但其前提是模型的输出必须严格基于其推理过程（即忠实度）。该研究指出了在要求模型确保安全性与保持 CoT 忠实度之间存在张力，当模型试图隐藏推理步骤时，监控将面临失效风险。

Clinical decision support is moving toward committees of language-model agents deliberating on a shared workspace. We ask whether such committees can be gamed by shortcuts, cues a benchmark rewards bu...

临床决策支持系统正逐渐转向由多个大模型 Agent 组成的委员会共同协商。该研究探讨了这种多 Agent 系统是否会被“捷径”攻破，即系统为了迎合基准测试的奖励机制而忽略了真实的临床诊断逻辑，从而在真实医疗场景中引发风险。

Agentic inference now dominates the LLM inference landscape, requiring LLMs to actively engage in multi-turn interactions with tool-calling capabilities. This introduces a more complex workload for th...

Agentic推理目前主导了大模型推理的格局，要求大模型具备调用工具的能力并积极参与多轮交互。这为推理引擎引入了更复杂的工作负载，本文模拟并探讨了推理解耦架构在Agentic场景下的收益。

Recent GUI visual grounding models generate screen coordinates as sequences of digit tokens that are parsed into numerical values and mapped to executable clicks. The security implications of this coo...

现代 GUI（图形用户界面）视觉定位模型在生成点击坐标时，通常将其作为数字 Token 序列输出。该研究揭示了这种序列化解析方式存在的安全隐患，恶意攻击者可以通过特定的视觉或文本扰动，诱导模型生成错误的坐标，进而触发非预期的点击操作。

Multilingual multi-agent systems exhibit substantial degradation beyond English, yet prior work rarely identifies how task-critical information is lost when user requests are converted into executable...

研究指出了在英语之外的语言环境中，多Agent系统在将用户请求转化为可执行任务时会发生严重的任务关键信息丢失。

Patient-facing medical LLMs and agents increasingly answer symptom questions before clinician contact, where the key safety question is what action the user should take next. We introduce CARE-Bench, ...

随着面向患者的医疗大模型和 Agent 在医生接诊前解答症状问题，决定“用户下一步该采取什么行动”成为了核心安全问题。CARE-Bench 专门针对这一分诊场景，全面评估大模型在医疗引导和安全建议方面的准确性与可靠性。

Automated Knowledge Base Construction (AKBC) is a core NLP task, and recent work proposes generating knowledge bases directly from large language models (LLMs), treating the model itself as the knowle...

自动化知识库构建（AKBC）是 NLP 的核心任务，GPTKB 2.0 探索了直接将大语言模型（LLM）视为知识库本体，通过抽取和消歧实体关系，大规模构建具备明确语义结构的知识库，超越了传统的向量检索。

Onboard satellite intelligence requires a task layer that translates mission intent into local tool calls, exposes execution state, and returns machine-consumable artifacts under communication and pow...

星载智能要求在极度受限的通信和功耗下，将任务意图转化为本地工具调用并返回机器可读结果。SAT-Edge-Agent 提出了一种硬件在环的边缘 Agent 编排架构，使卫星能够在近地轨道自主执行任务并减少对地面站的依赖。

Action labels tell a vision-language-action (VLA) policy which robot commands to imitate, but not how those commands change the 3D world. The aligned demonstration clip contains this missing supervisi...

动作标签仅告诉视觉-语言-动作模型（VLA）要模仿什么指令，却未说明这些指令如何改变 3D 世界。Track4Action 通过将世界中心的 3D 跟踪器蒸馏到 VLA 策略中，使模型能够更好地理解动作对物理世界状态产生的实际影响。

Collective intelligence research treats disagreement as evidence of epistemic diversity: if agents express different views, the group should retain capacity to revise. In LLM collectives this proxy ca...

集体智能研究通常将 Agent 之间的意见分歧视为认知多样性的证明。但在大模型集群中，这种分歧可能只是随机生成的假象。该研究提出了一种黑盒耦合诊断方法，用于判断大模型群体的分歧是否真的能促成有效的认知修正与共识达成。

Large language models can generate fluent responses to Islamic questions while introducing factual errors that are difficult to identify. This paper presents our system for \textsc{HalluScoring 2026} ...

大型语言模型可以对伊斯兰问题生成流畅的回答，但同时会引入难以识别的事实错误。本文介绍了针对HalluScoring 2026任务开发的系统，用于检测阿拉伯语宗教问答中的大模型幻觉并恢复真实答案。

Real-time bidding (RTB) ad exchanges typically forward nearly all incoming requests to demand-side platforms (DSPs), even though only a small fraction receive bids. This over-distribution weakens auct...

实时竞价（RTB）广告平台通常将几乎所有请求转发给需求方平台（DSP），但实际只有极少部分能获得出价。这种过度分发不仅浪费带宽，还削弱了拍卖效果。本文提出了一种竞争感知的请求调度机制，通过精准过滤低价值请求来提升平台整体收益。

World-action modeling has emerged as a promising paradigm for robotic control, as it empowers models to go beyond reacting to observations and anticipate how a scene will evolve. However, existing WAM...

世界-动作模型（WAM）通过预测场景演化赋予机器人前瞻能力，但现有模型计算开销过大。LiLa-WAM 提出了一种轻量级的潜在推理机制，在潜在空间中预测场景变化并输出动作，显著降低了机器人控制的计算延迟。

Persona skills distill personal interaction histories into portable and executable artifacts for downstream agents. While enabling flexible personalization, this process concentrates fragmented person...

个人技能将个人的交互历史提炼为可移植的伪影，供下游智能体使用。虽然这实现了灵活的个性化，但也集中了个人隐私数据。本文对这类Persona技能中的隐私泄露、冒充风险及防御手段进行了基准测试。

Persistent memory helps long-term agents retain knowledge, yet a single update error can repeatedly distort future retrieval and reasoning. Most existing systems reduce memory updating to a binary Wri...

长期 Agent 需要持久化记忆来保存知识，但一次错误的更新就可能导致后续检索和推理全盘崩溃。TARL 摒弃了传统的简单“写入-删除”操作，引入了类似数据库的事务感知账本机制，确保内存更新的原子性和可回滚性。

Multimodal large language models (MLLMs) are increasingly used to translate webpage screenshots into front-end code, but repeated UI patterns may sway them toward visually incorrect yet pattern-consis...

多模态大语言模型（MLLM）常被用于将网页截图转换为前端代码。但研究发现，当截图包含重复的 UI 模式时，模型会倾向于生成符合常见视觉模式但与实际截图不符的代码。本文对该现象进行了量化评估，揭示了模型的“模式补全偏差”。

介绍了如何使用参数量仅为 26 亿的 LFM2.5-2.6B 模型在各类本地设备上部署 Agent。该模型兼顾了推理速度和低资源消耗，适合在边缘设备上运行需要隐私保护和离线可用的智能代理任务。

Large language models are increasingly capable of synthesizing executable frontend projects, yet existing benchmarks still treat web generation as a static evaluation problem. We argue that frontend a...

大型语言模型越来越能够合成可执行的前端项目，但现有的基准测试仍然将网页生成视为静态评估问题。研究者认为前端生成应该是动态的，并提出了LiveEvalBench来进行开放世界的网页生成评估。

Reinforcement fine-tuning (RFT) is widely believed to inherently resist catastrophic forgetting in continual post-training of multimodal large language models. Under pronounced task distributional shi...

强化微调（RFT）通常被认为能抵抗多模态大模型持续后训练中的灾难性遗忘。但在显著的任务分布偏移下，RFT的这种抗性会被打破。本文提出了一种双通道风险感知的强化微调方法来解决这个问题。

Large language models (LLMs) are increasingly used to generate scientific reviews, yet existing evaluations rarely examine whether different providers align with both conference decisions and human re...

大型语言模型越来越多地被用于生成科学审稿意见，但现有的评估很少关注不同大模型提供商的审稿意见是否与会议最终决定和人类审稿人的偏好相一致。本文对此进行了系统调查。

Abstractive summarization models remain vulnerable to factual inconsistency, redundancy, and weak length control. We propose a modular generation-and-selection framework for sentence-budget-constraine...

生成式摘要模型目前容易面临事实不一致、冗余和长度控制不佳等问题。本文提出了一种模块化的“生成-选择”框架，能够在特定句子预算约束下，生成更加忠实于原文的摘要。

Network dismantling is fundamental to analyzing the robustness and vulnerability of complex systems, yet practical heuristics must balance effectiveness and computational efficiency, and are usually d...

网络拆解是分析复杂系统鲁棒性和脆弱性的基础，但实用的启发式方法必须平衡有效性和计算效率，且通常领域特定。本文提出了AutoSND，用于自动化发现网络拆解的启发式策略。

Fewer visual tokens do not guarantee lower end-to-end latency. We evaluate break-even with a reproducible protocol that accounts for decision overhead, shared work, and the operators each policy can a...

减少视觉Token并不一定能保证更低的端到端延迟。本文提出了一种可复现的评估协议，综合考虑了决策开销、共享工作量以及每种策略可以加速的算子，评估了视觉Token减少带来的损益平衡点。

Multi-Agent Debate (MAD) improves the reasoning performance of Large Language Models (LLMs) through multi-round interaction. However, LLMs in MAD are highly susceptible to blind conformity. Existing i...

多智能体辩论（MAD）通过多轮交互提高了大型语言模型的推理性能。然而，MAD中的大模型极易产生盲目从众。现有干预手段效果有限，本文提出通过调节群体辩论关系来缓解这种盲从现象。

AI agents deployed in real-world settings must be capable of coordinating with humans and other AI agents they have not encountered before. Zero-shot coordination (ZSC) algorithms aim to achieve this ...

部署在真实场景中的AI智能体必须能与未曾见过的人类或其他AI智能体进行协调。零样本协调（ZSC）算法旨在实现这一目标，但本文研究发现，现有的ZSC算法对代码实现细节非常敏感，泛化能力存疑。

Large language model-based automated heuristic design (LLM-AHD) has shown strong potential in discovering effective heuristics for combinatorial optimization problems. However, existing methods primar...

基于大模型的自动启发式设计（LLM-AHD）在为组合优化问题发现有效启发式方法方面显示出强大潜力。然而现有方法多局限于单一启发式，本文提出MuEvo，利用大模型进化出多启发式的集成策略。

On-Policy distillation (OPD) transfers teacher capabilities by supervising student-sampled trajectories with dense token-level teacher signals. Recent selective OPD methods improve this process by pri...

在线策略蒸馏（OPD）通过密集的token级教师信号监督学生模型的采样轨迹来转移能力。但当教师模型存在虚假信号时会产生误导，本文提出了一种感知虚假信号的选择性OPD方法来优化这一过程。

Large Language Models (LLMs) are increasingly used for automated fact-checking, yet their susceptibility to gender bias in this context remains underexplored. This study presents the first systematic ...

大型语言模型越来越多地用于自动化事实核查，但它们在此背景下的性别偏见 susceptibility 仍缺乏探索。本文首次系统研究了基于LLM假新闻检测中的性别偏见问题。

Activation patching and weight-space ablation both claim a component is causally responsible for a behavior, yet they act on different objects: one forward pass versus the parameters behind every forw...

激活修补和权重空间消融都声称某个组件对某种行为负有因果责任，但它们作用于不同的对象。本文研究了在低秩权重空间消融下的条件崩溃理论，并进行了单块理论和合成验证。

Reproducing web GUI bugs from natural-language bug reports is critical for software maintenance, but remains difficult because reports often lack prerequisites such as dependencies and input files. Ex...

从自然语言Bug报告中复现Web GUI的Bug对于软件维护至关重要，但由于报告通常缺少依赖项和输入文件等先决条件，这使得复现变得困难。本文提出了一个基于LLM的智能体来自动执行这一复现流程。

Open-source software communities are a form of digital public infrastructure that not only produces code, but also generates public knowledge and interpersonal relationships through visible collaborat...

开源软件社区是一种数字公共基础设施，不仅生产代码，还通过可见的协作产生公共知识和人际关系。本文探讨了从传统的“社会化编码”向基于大模型的“智能体编码”转变时，开源社区生产力与人际关系的重构。

Recent works train agents by constructing large-scale multimodal environment pools. However, we find that simply increasing the number of multimodal environments does not always benefit. We further an...

近期研究通过构建大规模多模态环境池来训练智能体。然而研究者发现，单纯增加多模态环境的数量并不总是有益的。本文深入分析了如何为智能体学习设计有效的环境分布。

VLA models are trained to predict robot actions from visual and language observations. This is a natural choice, but it creates a mismatch: VLMs encode rich, high-level representations of scenes and g...

目前的视觉-语言-动作模型（VLA）直接根据视觉和语言预测机器人动作，但这与 VLM 中丰富的高层场景表征产生了不匹配。该研究提出统一视觉运动目标，不仅要预测动作，还要监督模型理解场景的高级语义，从而提升策略的泛化性。

Existing chunk-based Vision-Language-Action (VLA) models execute a fixed number of actions (i.e., execution horizon) before replanning, turning replanning into a task-agnostic periodic schedule that i...

针对基于分块的视觉-语言-动作（VLA）模型在执行固定数量动作后才重新规划导致的滞后问题，提出了一种自适应决定何时重规划的新策略。

Humanoid robots have the potential to perform dexterous manipulation in human environments, yet acquiring diverse and generalizable skills remains costly due to expensive hardware data collection and ...

由于人形机器人收集真实硬件数据昂贵，该研究利用生成的第一人称视频提取可泛化的操作技能，以降低训练成本。

Learn how to build an end-to-end security assessment pipeline for AI agent skills using NVIDIA SkillSpector and LangGraph. In this tutorial, we construct a synthetic skill marketplace, scan for malici...

教程介绍了如何利用 NVIDIA SkillSpector 和 LangGraph 构建端到端的 AI Agent 技能安全评估流水线。该流水线通过构建合成技能市场，结合 YARA 规则扫描恶意代码，并在 CI/CD 中实施安全策略拦截，保障 Agent 调用外部工具的安全性。

Vision-Language-Action (VLA) policies promise general robotic manipulation, but their robustness against physical-world attacks remains fragile. In particular, we show that physically realizable adver...

揭示了VLA模型在真实世界中容易受到物理对抗样本攻击的脆弱性，并提出了一种结构感知的微调方法以增强防御能力。

Y Combinator has open-sourced QM, the multiplayer agent harness it uses internally across accounting, legal, events, and engineering. Released July 31, 2026 under an MIT license, QM gives each employe...

Y Combinator 宣布开源其实内部使用的多人协作 Agent 框架 QM。该框架覆盖了财务、法务、活动和工程等多个业务场景，能够为每位员工提供定制化的 Agent 支持，并可灵活部署在 Slack 和网页端。

In this tutorial, we design an end-to-end evaluation workflow for PerceptionBench. This multimodal benchmark measures fine-grained visual perception capabilities across tasks such as OCR, counting, lo...

该教程为 PerceptionBench 设计了一个端到端的评估工作流。这一多模态基准测试可以衡量模型在 OCR、计数、定位等任务中的细粒度视觉感知能力，并包含强大的数据加载和自动评分功能。

AI agents, MCP servers, and LLM apps break the core AppSec assumption that applications do what their code says. This guide walks through a practical see-fix-protect framework: a five-layer agentic AI...

AI 智能体、MCP 服务器和 LLM 应用打破了传统应用安全“代码完全按预期执行”的假设。本指南提供了一个“发现-修复-保护”框架，介绍了五层的智能体 AI 安全防护策略。

Language remains an outlier in generative modeling: while images, video, and audio are increasingly modeled in continuous latent spaces, text generation still relies predominantly on discrete tokens. ...

文本生成目前仍以离散 Token 为主，而图像和视频已普遍采用连续潜在空间。AURORA-LM 尝试将语言映射到连续空间中进行扩散建模，打破了传统自回归文本生成的局限。

Optimization-based latent reasoning improves large language model outputs by optimizing instance-specific continuous states at test time while keeping model parameters frozen. Existing methods, howeve...

基于优化的潜在推理通过在测试时优化连续状态来提升输出，但现有方法难以解释。本文提出 GradCuit，利用信用分配梯度流来优化测试时的推理过程，增强了模型的可解释性和鲁棒性。

Sparse retrieval underpins modern search systems, from web search to retrieval-augmented generation. Existing work has introduced Learned Sparse Retrieval (LSR) to push beyond exact lexical matching t...

稀疏检索支撑着从网络搜索到检索增强生成（RAG）的现代搜索系统。现有学习型稀疏检索（LSR）突破了精确词汇匹配限制，本文提出统一稀疏和密集的多模态嵌入方法。

Traditional software cost estimation models, such as COCOMO II, Function Points, and Story Points, assume that development effort is primarily driven by human labor in design, coding, and testing. Age...

COCOMO II 等传统软件成本估算模型假设开发主要由人工劳动驱动。随着智能体参与开发，Agent 的 token 消耗和 API 调用成本需要全新的估算模型。

Retrieval-augmented generation (RAG) imposes a prefill cost proportional to retrieved context length, and -- with Transformer backbones -- a KV-cache that grows with each generated token. State-Space ...

检索增强生成（RAG）面临上下文预填充成本高的问题。本文针对边缘设备上的语言模型，利用状态空间模型（SSM）实现O(1)复杂度的状态注入，提供持久的上下文记忆和语料检索能力。

An isolated final user message is often treated as the query in evaluations of AI systems. In a conversation, however, the actionable request may be distributed across preceding turns. We directly tes...

评估AI系统时通常只看最后一条用户输入。但研究发现，在多轮对话中，用户的实际诉求往往散落在前几轮的上下文中，单看最后一句话会导致严重的评估失真。

Cognitive AI seeks to move beyond language generation and autonomous task execution toward systems capable of sustained reasoning, adaptive behavior, persistent memory, and self-regulation. While gene...

认知 AI 寻求超越语言生成和任务执行，迈向具备持续推理、适应性行为、持久记忆和自我调节的系统。目前的生成式 AI 在这些方面仍存在明显差距。

Mixtures of low-rank adaptation experts increase parameter-efficient capacity by routing each input through a subset of adapters. Recent dynamic routers activate more experts when the router or predic...

混合低秩微调专家通过路由机制提升了模型容量。现有动态路由在预测不确定时倾向于激活更多专家，但本文指出这不够，提出应基于“信息价值”来决定专家路由，优化计算资源分配。

Large language models (LLMs) are increasingly used for health-related advice. Existing research measures their safety with static questions rather than pressured patient-facing conversations. We intro...

大语言模型越来越多用于健康咨询。现有安全测试多用静态问题，而 MedPRESS 引入多轮对话，测试模型在患者不断施加压力（如要求开特定处方）时，是否会妥协并产生“谄媚性”的危险建议。

The most capable AI deployments are not single models but ensembles of specialized agents that delegate and act in coordination. This architecture unlocks powerful new capabilities, and it also introd...

最强大的 AI 部署不是单一模型，而是协调行动的多个专业智能体集合。这种架构解锁了新能力，但也引入了跨会话积累能力的风险。

Long-running assistants and agents consume interaction streams that eventually outgrow the context. Existing context retention, summarization, and retrieval preserve access to selected history, but do...

长时间运行的助手和智能体最终会消耗超出上下文窗口的交互流。现有的摘要和检索保留了对历史记录的访问，但破坏了记忆状态的连续性。

Learning-based memory systems for self-evolving LLM agents face two tightly coupled challenges. First, trajectory-indexed utilities grow with the interaction history, thereby dispersing limited feedba...

基于学习的自进化 LLM 智能体记忆系统面临轨迹效用增长的挑战，导致反馈分散；同时陷入了基于历史的记忆奖励陷阱。本文提出降阶效用状态来缓解这些问题。

Video Diffusion Transformers (VDiTs) enable high-fidelity generation but incur quadratic cost from dense 3D self-attention. Existing head- and block-level sparse methods share computation budgets acro...

视频扩散 Transformer（VDiTs）能生成高保真视频，但密集的 3D 自注意力机制带来了平方级的计算开销。本文提出 Token 半径注意力，在保证生成质量的同时大幅降低了计算成本。

Real-world software development requires coding agents to operate in shared workspaces where users may inspect and modify code during an ongoing task, yet existing repository-level benchmarks typicall...

现实世界的软件开发中，用户可能会在代码智能体执行任务期间检查或修改代码，但现有的仓库级基准测试通常假设智能体在隔离环境中工作。

LLM agents fail mid-episode -- they loop, cascade tool errors, drift off goal, fabricate results, or silently absorb corrupted content -- and the standard remedy, judging every step with a second LLM,...

LLM 智能体在执行过程中会失败（如死循环、工具报错级联、偏离目标或编造结果）。标准补救措施是用另一个 LLM 判断每一步，但这成本太高。

LLM-agent evaluations often produce task outcomes long before the full benchmark run is complete. A partial score is tempting to report, but it does not show whether the observed tasks support the sam...

LLM 智能体的评估通常在完整基准测试结束前就产生结果。报告部分得分虽然诱人，但它不能表明观察到的任务是否支持与完整测试相同的统计结论。

Scientific reasoning benchmarks typically evaluate large language models (LLMs) using final-answer accuracy. However, a correct answer does not necessarily demonstrate the reasoning capability targete...

科学推理基准通常只以最终答案的准确性来评估大模型。但研究发现，LLM经常通过“捷径”得出正确答案，并未真正运用目标推理能力，这导致现有的高分数存在水分。

In vibe coding, people describe software in natural language and delegate implementation to AI agents. By analogy, vibe commerce allows people to express buying or selling goals in natural language an...

正如“氛围编程”用自然语言描述软件并交给 AI 实现一样，“氛围电商”允许人们用自然语言表达买卖目标并将其委派给智能体。

This work-in-progress research paper explores Chat-Debugging, a novel use case for large language models as an assistant for hardware debugging tasks to improve students' debugging skills. Hardware de...

本文探讨了将大语言模型作为硬件调试助手的新用例。硬件调试通常需要查看复杂的原理图和代码，LLM的介入旨在帮助学生或工程师更快地定位硬件层面的错误。

Large language models (LLMs) have become the default tool for a remarkable range of tasks, yet they have had conspicuously little success at one of the most common machine learning workloads: predicti...

大语言模型（LLMs）在许多任务中表现出色，但在最常见的机器学习工作负载之一——表格数据预测上却屡屡受挫。研究指出，这主要是因为LLM的架构更擅长处理序列文本，缺乏对结构化行列数据的原生理解能力。

Benchmark suites assess model capability on controlled tasks; large-scale conversation corpora capture naturalistic use without user feedback; and in-interface feedback mechanisms record satisfaction ...

传统的基准测试缺乏真实反馈，而大规模对话语料又缺乏满意度指标。本文提出 MonitrLLM，结合界面内的用户反馈机制，构建了一个以社区用户真实体验为中心的模型评估基础设施。

Vulnerability localization is a fundamental step in software security, requiring models to reason over large codebases and iteratively identify vulnerable implementations. We present Antares, a family...

漏洞定位是软件安全的基础步骤，需要模型对大型代码库进行推理并迭代识别脆弱实现。研究提出了Antares模型系列，专门针对智能体化操作设计，能够高效遍历代码并定位安全漏洞。

Tool-using large language model (LLM) agents produce long, multi-turn trajectories, making gradient-based post-training memory-intensive. Evolution strategies (ES) enable memory-efficient full-paramet...

使用工具的LLM智能体会产生长序列的多轮交互轨迹，导致基于梯度的后训练内存消耗巨大。本文采用进化策略（ES），实现了内存高效的全参数后训练，解决了长轨迹智能体的训练瓶颈。

Search and database engines still store text as UTF-8, a format built for humans. But the systems that increasingly read and write that text (embedders, rerankers, and language-model agents) work in t...

当前的搜索引擎和数据库仍以面向人类的 UTF-8 格式存储文本，而机器（如嵌入模型和LLM智能体）实际处理的是 Token。本文提出 Token 原生存储，直接以模型的语言进行存取，消除转换开销。

Large Language Models (LLMs) are increasingly deployed in task-oriented dialogue systems that support multi-step decision-making in high-stakes domains such as education, healthcare, and finance. Howe...

LLM 越来越多地用于医疗、金融等高风险领域。本文提出了 PredAct-Bench，通过在多步决策对话中引入可控的“工具噪声”（如API报错或返回错误数据），评估智能体在异常情况下的鲁棒性。

To operate robustly in open-world environments, autonomous agents should be able to infer the behavior of unfamiliar systems through interaction alone, even in the absence of documentation. However, e...

在开放环境中，自主智能体应能通过交互推断陌生系统的行为。但研究发现，即使环境中提供了明确的“下一步操作地图”，智能体依然会陷入低效的穷举搜索，暴露出当前模型在利用先验知识上的缺陷。

Large language model agents increasingly solve complex tasks by composing reusable skills from a library. To address this, the key challenge is not merely to retrieve individually relevant skills, but...

LLM智能体常通过从库中调用可复用技能来解决问题。本文提出 SkillTrace，核心挑战不再是单一技能检索，而是通过遍历查询-技能图谱，将多个基础技能动态组合以解决复杂任务。

Recent advances in agentic workflow optimization automate workflow design through task-specific workflow search or input-conditioned architecture selection. However, they determine the workflow before...

现有的智能体工作流优化通常在执行前就确定了架构。本文提出在推理阶段进行“区域嫁接”，允许在运行时动态替换或优化工作流的特定部分，提升了复杂Agent任务的执行效果。

Native computer use offers a general interface for agents to operate almost any software available to people, but requires long-horizon state tracking, large-scale interactive experience, and learning...

原生电脑操控（Computer Use）为智能体提供了控制人类软件的通用接口。Qwen-CUA 展示了如何通过长程状态跟踪和大规模交互经验，让模型学会操作几乎所有桌面端软件。

Data drift poses significant challenges for machine learning systems in production, requiring continuous model updates to maintain performance. We present KC-Agent, a dual-process cognitive architectu...

生产环境中的机器学习系统面临数据漂移问题。本文提出 KC-Agent，采用双系统认知架构，自动化地进行数据检测和模型持续更新，以维持模型性能。

A/B testing remains the standard for rolling out new features in the technology industry. Each experiment, however, consumes real traffic, engineering effort, and weeks of wall-clock time. Can AI agen...

A/B测试是科技行业推出新功能的标准流程，但消耗真实流量和开发时间。本文探讨能否利用AI智能体模拟用户行为，提前预测A/B测试结果，并提供了验证框架。

The rapid spread of false and misleading health information through digital platforms has become a major public health challenge, particularly during infectious disease outbreaks where delayed verific...

虚假健康信息在数字平台上的快速传播构成了重大公共健康挑战。本文提出了一个结合检索增强生成（RAG）技术的 Transformer 框架，通过自动检索权威医学文献来验证和辟谣健康谣言。

Scientific poster construction compresses a long multimodal paper into a readable, editable canvas. Existing systems hide request-level failures by scoring only completed outputs; direct image generat...

科学海报需要将长篇多模态论文压缩为易读、可编辑的画布。现有系统往往掩盖了请求级别的失败，直接生成图像会限制后续编辑。

Agentic multimodal reasoning extends passive image understanding by allowing VLMs to actively acquire and refine visual evidence through visual tool interactions. Effective visual tool use requires th...

多模态推理通过允许视觉语言模型（VLM）主动获取视觉证据扩展了被动的图像理解。有效的视觉工具使用要求模型能够组合调用多种工具。

Multi-hop question answering is a fundamental challenge in retrieval-augmented generation (RAG), because deriving an answer requires integrating dispersed evidence. Iterative RAG (iRAG) is widely used...

多跳问答是检索增强生成（RAG）的基础难题，因为推导答案需要整合分散的证据。迭代 RAG 被广泛使用但面临信息丢失的挑战，MEGRAG 引入多粒度证据图来优化此过程。

Multilingual dense retrieval aims to handle queries and documents across different languages based on a unified retriever model. The challenge lies in enabling robust retrieval transfer to low-resourc...

多语言密集检索旨在用统一模型处理跨语言查询。其挑战在于如何实现向低资源语言的稳健检索迁移，本文提出解耦对比学习来提升零样本表现。

Proximal Policy Optimization (PPO) for large language models typically trains its critic by mean-squared-error (MSE) regression on scalar value targets. Although scalar MSE is statistically valid for ...

近端策略优化（PPO）通常通过均方误差（MSE）训练评判器。尽管标量 MSE 统计上有效，但它可能不是捕获价值分布的最佳方法，本文提出分类评判器。

Decentralized populations of Large Language Model (LLM) agents can spontaneously reach consensus on shared conventions, yet the microscopic mechanisms by which their internal stochasticity shapes macr...

去中心化的大语言模型（LLM）智能群体可以自发对共享惯例达成共识，但其内部随机性塑造宏观共识的微观机制尚不清楚。本文研究了这一动力学过程。

Model checkpoints are growing in both number and size, which makes archival, transfer, and deployment increasingly costly. General-purpose compressors can reduce storage requirements but ignore tensor...

模型检查点在数量和大小上不断增长，使得归档、传输和部署日益昂贵。通用压缩器可以减少存储需求，但忽略了张量的内部结构特性。

With 3.5 billion active users to protect, Google is relying on Gemini to find Chrome security bugs fast - and before attackers do.

为了保护 35 亿活跃用户，Google 正在利用 Gemini 大模型快速发现 Chrome 浏览器的安全漏洞，赶在攻击者利用之前进行修复。

The deployment of Vision-Language Models (VLMs) in autonomous driving (AD) systems is constrained by on-board computing power, restricting vehicles to small VLMs (SVLMs) with limited perception and re...

车载算力限制了自动驾驶系统只能使用较小的视觉语言模型（SVLM）。本文提出 TALSC，通过车端 SVLM 与云端大模型协同，并结合时效感知路由，在低延迟下实现了大模型级别的感知推理。

Alibaba's Qwen team moved Qwen3.8-Max from preview to general availability, with published per-token pricing and open weights due next week. The 2.4T parameter MoE model accepts text, image and video ...

阿里巴巴 Qwen 团队将 Qwen3.8-Max 从预览版转为正式可用，公布了按 token 计费的价格并将于下周开放权重。这个拥有 2.4T 参数的混合专家（MoE）模型支持文本、图像和视频输入。

Cogent AI team released Cogent VR-1, a reasoning model post-trained specifically for cybersecurity rather than picking up cyber capability as a side effect of general coding strength. It ships with tw...

Cogent AI 团队发布了 VR-1，这是一个专门针对网络安全进行后训练的推理模型，而非将网络安全能力作为通用编码能力的副产品。该模型能够组合并验证企业级攻击路径。

Recent Vision-Language-Action (VLA) models for autonomous driving (AD) increasingly utilize chain-of-thought (CoT) supervision to enhance the reasoning capabilities of their Vision-Language Model (VLM...

最新的视觉-语言-动作（VLA）模型利用思维链来增强推理。本文通过“延迟暴露未来轨迹”的方法，让模型在输出动作前进行可验证的轨迹推演，提升了自动驾驶决策的安全性和透明度。

Onton, a San Francisco-based search and discovery company, has released Ontology 1, a neurosymbolic model for complex, conversational, multimodal product search. On a 90-query benchmark scored by thre...

Onton 发布了名为 Ontology 1 的神经符号模型，专门用于复杂、对话式、多模态的产品搜索。在由三位专家打分的 90 个查询基准测试中，其准确率比全球最好的电商搜索引擎高出 2.7 倍。

Large language model (LLM) agents learn world dynamics from local interaction experience to support subsequent planning and action selection. However, the experience available to a single client is of...

单个LLM智能体只能从局部交互中学习世界动态，经验受限。FedWorld 提出通过联邦学习的方式聚合各个客户端的世界模型经验，同时具备范围感知能力，确保不同环境的模型知识不冲突。

Vision-language models (VLMs) are emerging as a key component of embodied intelligence, with growing applications in auto-labeling and end-to-end autonomous driving. However, existing approaches for i...

视觉语言模型（VLM）正成为具身智能的关键组件，在自动标注和端到端自动驾驶中应用日益广泛。然而，现有的方法在时空动态理解上存在局限，本文通过引入汽车雷达监督来增强VLM的时空定位能力，从而更好地估计物体的运动和速度。

Inkling-Small matches Inkling at a quarter the size, and its NVFP4 checkpoint runs on one NVIDIA B300 GPUThe postThinking Machines Lab Releases Inkling-Small: A 276B Total, 12B Active Open Weights Mul...

Thinking Machines Lab 推出了多模态开源权重 MoE 模型 Inkling-Small，总参数量为 2760 亿，每个 Token 激活 120 亿，且其 NVFP4 检查点可在单张 NVIDIA B300 GPU 上运行。

Evaluating Multi-Agent Reinforcement Learning (MARL) policies in autonomous driving fundamentally relies on extrinsic statistical indicators (e.g., reward curves and success rates), which often mask i...

在自动驾驶中评估多智能体强化学习（MARL）策略，通常依赖成功率等外部统计指标，但这些指标往往会掩盖策略内在的安全隐患。该研究提出了一种新的评估方法，直接衡量策略的内在质量，以确保多车协同决策的最优性和安全性。

Depth-routing residual architectures allow Transformer layers to retrieve earlier representations instead of inheriting only the immediately preceding state. Existing Block Attention Residuals, howeve...

深度路由残差架构允许 Transformer 层检索早期的表示，而不仅是继承前一层状态。现有的块注意力残差无法区分模式匹配和内容检索。

Tool-using language-model agents are governed not only by task prompts but also by persistent system-side instructions that specify tools, arguments, policies, execution protocols, and recovery. Compr...

使用工具的语言模型智能体不仅受任务提示词控制，还受持久化系统指令（如工具定义、协议等）控制。压缩这些系统指令会极大影响智能体的可靠性，本研究探讨了如何在压缩上下文的同时保持控制力。

Production LLM agents that select from large skill libraries face a limitation that semantic relevance alone cannot resolve: a skill may match a user's topic yet be impossible to execute in the curren...

在生产级 LLM 智能体从大型技能库中选择技能时，仅靠语义匹配不够：匹配用户主题的技能可能在当前环境中无法执行。本研究提出了一种确定性可执行性门控机制，用于大规模技能选择。

Enterprise AI agents act across many apps whose data changes continuously, so an answer is correct only relative to what data existed and who could see it at the moment it was asked. Offline evaluatio...

企业 AI 智能体在数据持续变化的多个应用间操作，因此答案的正确性取决于提问时刻的数据状态和可见性。本研究通过重放真实研究数据的时间切片，来离线评估智能体在特定时间点的表现。

Language models small enough to run on a handset, quantized to a few bits, are increasingly capable of acting for their user, making on-device task automation newly plausible. One such task is answeri...

量化为几个比特的端侧小模型越来越具备代表用户操作的能力。本研究提出了 CallScreenBench，测试端侧模型接听电话（如电话秘书）的表现，使得端侧任务自动化变得更具可行性。

Agentic RL research is constant algorithm modification, and in mainstream frameworks every change threads through trainer, distributed backend, and rollout glue. NVIDIA's Molt targets that cost with a...

NVIDIA 推出了 Molt，这是一个原生 PyTorch 的强化学习（RL）框架，旨在解决智能体强化学习研究中每次算法修改都需要贯穿训练器、分布式后端的痛点。

A language model that abandons a correct medical answer under user pushback is more dangerous than one that was simply wrong, because it lends the credibility of a correct answer to the user's misinfo...

如果一个医疗语言模型在用户的质疑下放弃了原本正确的医学答案，其危害远大于一开始就答错，因为它给用户的错误信息披上了正确的伪装。

Uncommon and off-guideline cases are difficult for clinical decision support, because physicians must make a series of management decisions under diagnostic uncertainty and rarely see the full case at...

临床决策支持在处理罕见和偏离指南的病例时面临困难。医生需要在诊断不确定性下做出一系列决定，且很少能在一开始看到完整病例。

Platform teams hosting agent-extensibility surfaces face a regression-economics paradox: every onboarding customer ships an evaluation set tuned to their domain, but the platform's regression set must...

提供智能体扩展能力的平台团队面临一个悖论：每个客户都提交了针对自身领域的评估集，但平台的回归测试集必须保持通用。本研究提出了一种基于能力分类的流水线，用于管理和挑选合适的回归评估集。

The emergence of Agentic AI systems, characterized by autonomous reasoning, multi-agent collaboration, tool orchestration, adaptive decision-making, and persistent memory, represents a fundamental shi...

智能体 AI 系统具有自主推理、多智能体协作、工具编排和持久记忆等特征。本文系统梳理了这类系统带来的新型“Agentic 技术债”的根源和表现，表明这代表了系统复杂度的根本转变。

When a self-improving AI-for-science system claims a new capability, the evidence is usually a benchmark delta, a description-length gate, or a p-value. None separates a real gain from extra search, f...

当自我迭代的“AI for Science”系统声称发现新能力时，现有证据（如基准测试增量、p值）无法区分这是真正的进步还是单纯增加搜索量的结果。本研究提出了一种具有可判定负面标准的双面评判标准。

The effective use of search engines by large language models (LLMs) remains a significant challenge, particularly in complex, multi-hop question-answering (MHQA) tasks. These tasks require the model t...

大语言模型有效使用搜索引擎（特别是在复杂、多跳问答 MHQA 任务中）仍然是一个挑战。Search-GRT 提出了一种引导式检索训练方法，优化大模型在复杂问题中的搜索和信息整合能力。

Existing search-augmented LLM agents are trained using Reinforcement Learning to boost its reasoning capabilities. However, these approaches primarily rely on outcome-level rewards, which provide litt...

现有的搜索增强型 LLM 智能体主要使用结果级别的奖励进行强化学习训练。PROGRESS 引入了覆盖率引导机制，在推理过程中提供更丰富的反馈，从而提升智能体在搜索增强任务中的表现。

Large language model agents have shown strong capabilities in generating coherent and contextually appropriate responses, yet robust long-horizon dialogue remains limited by the lack of external memor...

大语言模型在生成长篇、长期对话时，常因缺乏外部记忆而受限。TrajWiki 提出了一种基于来源溯源的记忆轨迹机制，帮助智能体在长程对话中保持上下文连贯和事实准确。

AI coding agents increasingly author pull requests (PRs), often under developers' own accounts, obscuring who actually produced a change. Reliable attribution is important for repository governance, e...

AI 编码智能体越来越多地以开发者本人的账号提交代码，掩盖了代码变更的真正作者。AgenTag 提出通过代码行为指纹来追踪 AI 生成代码，以保障仓库治理和工程效能评估的准确性。

Long-term memory is essential for LVLM agents to maintain consistency and integrate information across extended multimodal interactions. Existing agent memory systems, however, often reduce visual exp...

长期记忆对于 LVLM 智能体在多模态交互中保持一致性至关重要。现有的系统通常会将视觉体验降级为文本，PMMC 提出了一种前瞻性多模态记忆编译方法，以保留更丰富的视觉上下文。

Parallel coding agents can produce locally valid changes that fail when combined. Claim Plane addresses this failure mode as deterministic pre-write admission over versioned change intents. This paper...

并行编码智能体可能会产生各自局部有效、但合并后失败的代码修改。Claim Plane 通过对版本化变更意图引入确定性的“预写准入”机制来解决这一失败模式，并在 30 对测试中验证了效果。

Social dynamics encode the process in which individual network and discourse interactions aggregate into collective influence, narrative dominance and coordinate behavior. This paper uses the the Ghos...

社会动力学编码了个体网络和话语互动如何汇聚成集体影响力、叙事主导权和协调行为的过程。本研究利用基于 LLM 的智能体网络动态（LAND）模型对相关动力学（如 Ghost 博客平台的数据）进行建模分析。

Code refactoring aims to enhance the internal structure of source code without affecting its functional behavior. The recent advancements of Large Language Models (LLMs) have demonstrated potential fo...

代码重构旨在不改变功能行为的前提下优化源码内部结构。大语言模型（LLM）展现了这方面的潜力，但由于难以保证行为完全等价，实际应用受限。

Agentic foundation-model service networks handle requests spanning retrieval, planning, generation, verification, and tool use. Unlike traditional communication networks, control performance depends o...

基础模型智能体网络需处理检索、规划、生成和工具调用等请求。与通信网络不同，其控制性能高度依赖上下文感知，传统流量控制难以直接套用。

LLM agents accumulate long trajectories of reasoning steps, tool calls, and environment feedback, making the KV cache a major inference bottleneck. KV cache compaction can reduce this cost, but most p...

LLM 智能体在推理、工具调用和环境反馈中累积长轨迹，导致 KV Cache（键值缓存）成为推理瓶颈。现有的压缩方法多难以适配在线场景。

Large language models have made it possible to generate executable computer-aided design (CAD) programs from natural-language descriptions or images. However, existing methods represent modeling proce...

利用大语言模型从自然语言或图像生成可执行计算机辅助设计（CAD）程序已成可能。现有方法缺乏跨工具编辑能力，CADIR 旨在提供跨后端兼容的中间表示。

Large language model agents increasingly act through stateful tools, yet model generation and environment execution remain serialized at every step. As decoding accelerates, tool execution becomes a g...

LLM 智能体越来越多地使用有状态工具，但模型生成与环境执行通常是串行的。随着解码加速，工具执行成为了新瓶颈。AOSpec 旨在解耦并重叠两者时间。

AMD released Instella-MoE-16B-A3B, a fully open Mixture-of-Experts language model trained from scratch on Instinct MI300X and MI325X GPUs. It holds 16B total parameters but activates only 2.8B per tok...

AMD 推出了在 Instinct MI300X 和 MI325X GPU 上从零训练的完全开源混合专家模型 Instella-MoE-16B-A3B。该模型总参数量为 160 亿，但每个 Token 仅激活 28 亿参数。

Discover how to optimize transformer workloads using the NVIDIA Transformer Engine. This tutorial guides you through configuring fused GPU kernels, implementing FP8 delayed scaling, and benchmarking m...

该教程指导如何通过配置融合 GPU 内核、实现 FP8 延迟缩放以及基准测试来优化 Transformer 工作负载，加速模型训练。

Cooperative perception (CP) enables connected autonomous vehicles (CAVs) to share complementary observations for safer navigation, but practical deployment is limited by bandwidth constraints, unrelia...

协同感知（CP）允许网联自动驾驶车辆（CAV）共享观测数据以提升安全性，但在实际部署中面临带宽和网络不可靠的限制。该研究利用大语言模型（LLM）辅助构建车辆间的动态合作联盟，在保障安全导航的同时大幅优化了通信效率。

Supabase has open sourced supabase/evals, an Apache-2.0 benchmark and framework that runs coding agents including Claude Code, Codex and OpenCode against real Supabase tasks — building schemas, debugg...

Supabase 开源了基于 Apache-2.0 协议的评测框架 Evals，使用真实的 Supabase 任务（如构建表结构、Debug）来评估 Claude Code、Codex 和 OpenCode 等编码智能体的表现。

MiniMax releases MiniMax H3, a general-purpose multimodal generation model. MiniMax H3 is not a text-to-video model with add-ons. MiniMax describes it as a general-purpose multimodal generation model ...

MiniMax 发布了通用多模态生成模型 MiniMax H3，能够直接生成带有原生立体声音效的 15 秒 2K 分辨率视频片段。

OpenAI shares new results on long-standing open problems in mathematics and theoretical computer science, including advances in geometry, cryptography, and complexity.

OpenAI 分享了在长期悬而未决的数学和理论计算机科学问题上的新成果，涵盖了几何学、密码学和计算复杂性领域的突破。

DeepSeek published DeepSeek-V4-Flash-0731 on Hugging Face and moved the official V4-Flash API into public beta on July 31, 2026. The model card is explicit that this is the official release supersedin...

DeepSeek 在 Hugging Face 上发布了 DeepSeek-V4-Flash-0731，并同步开启官方 API 公测，全面替代旧版本，重点提升了智能体调用与代码编写能力。

We study empirical scaling properties for text conditioning in visual generation. Such properties have rarely been measured because diffusion loss does not scale with the number of tokens in natural-l...

研究探讨了视觉生成（如文生图）中文本条件的经验缩放特性。由于扩散模型的损失不随自然语言 token 数量按比例变化，此前这种特性很少被测量，本研究填补了这一空白。

LLM serving systems cache prompt KV state, yet most front ends still re-tokenize the full request text on every call. The cost lands on coding agents, which resubmit a long transcript after each small...

大多数前端在调用 LLM 时仍会对完整请求文本重新进行分词。TokTier 提出了一种精确的有状态分词方法，解决了编码智能体在每次微调后重新提交大量记录导致重复分词带来的成本问题。

Enterprise workflows increasingly rely on agents for \emph{schema-guided extraction}: given a document and a user-defined schema, the agent faithfully follows the schema to produce the correct output ...

ExtractBench 专注于评估智能体执行模式引导抽取任务的能力，即给定文档和用户定义的模式，智能体能否严格遵循模式输出正确的结构化数据。

Although LLM-driven repair agents can tackle complex, repository-level issues, they treat every issue independently and discard the procedural knowledge accumulated from previous repairs. We introduce...

当前基于 LLM 的修复智能体通常独立处理每个问题而丢弃历史修复经验。本文引入分层轨迹抽象机制，让编码智能体能够积累并复用过去的程序性修复知识。

Multi-page document visual question answering requires locating sparse evidence at both the page and region levels. Existing approaches typically emphasize one level over the other: page-centric metho...

多页文档的视觉问答（VQA）需要在页面和区域级别定位稀疏证据。HierDoc 提出了一种分层路由方法，弥补了现有方法要么只关注页面级、要么只关注区域级的缺陷，提升了证据定位的准确率。

As LLMs evolve from code completion systems into autonomous scientific agents, evaluating their ability to conduct experiments has become increasingly important. Existing benchmarks typically focus on...

随着 LLM 向自主科学智能体进化，AgentHPOBench 专门用于评估 LLM 作为序列超参数优化器（HPO）执行实验的能力，弥补了现有基准对实验过程评估的不足。

Reinforcement learning (RL) post-training of Vision-Language-Action (VLA) models has shown strong promise for robotic manipulation. Among RL methods, critic-based approaches rely on a value estimator ...

视觉-语言-动作（VLA）模型的强化学习后训练在机器人操控中极具前景。其中基于评判者的方法依赖价值估计器，本文提出了一种世界评判模型（WCM）来优化这一强化学习过程。

Generative and agentic artificial intelligence (AI) are reconfiguring software and systems engineering from a discipline centered on human authorship of artifacts to one focused on directing, verifyin...

生成式和智能体 AI 正在重塑软件工程，使其从以人工编写产物为中心，转向以引导、验证和管理自主 AI 生成代码为核心的学科。

A fundamental challenge of vector search is achieving consistently high recall while minimizing computational costs. Fixed search parameters cause significant performance variance across queries, and ...

向量搜索中的一大挑战是在最小化计算成本的同时保持高召回率。固定的搜索参数在不同查询下性能差异很大，QASP 提出了一种自适应策略，根据具体查询动态调整参数以优化搜索效果。

Reading a social situation often depends on behavior, not words alone. We introduce FriendBench, a benchmark for inferring whether two people are already familiar or are meeting as strangers, from a 2...

判断社交情境往往依赖于行为而非仅仅语言。FriendBench 引入了一个基准测试，旨在通过短暂的互动视频片段，评估人类和多模态大语言模型推断两个人是熟人还是陌生人的能力。

Parameter-Efficient Fine-Tuning (PEFT) commonly adapts large language models using a single shared Low-Rank Adapter (LoRA). This shared optimization space often suffers from interference when adapting...

参数高效微调（PEFT）通常使用单个共享的低秩适配器（LoRA）。在适应多种任务时容易产生优化冲突，该研究提出了一种自动任务排序方法，以减少多策略训练中的相互干扰。

Vision-and-language navigation (VLN) enables robots to follow instructions in previously unseen environments. Recently, a training-free paradigm has emerged: the robot queries a multimodal LLM to unde...

视觉语言导航（VLN）使机器人能够在未见过的环境中执行指令。HAM-VLN 利用多模态大模型，采用免训练范式并引入分层智能体记忆，提升了机器人的指令跟随能力。

Vision-language-action models (VLAs), which leverage the cognition of multimodal information to infer physical-world actions, provide a generalized solution for embodied AI applications. Conventional ...

视觉-语言-动作模型（VLA）为具身智能提供了通用方案。传统方法在处理时序数据时效率低下，FibVLA 创新性地引入斐波那契采样策略，构建了更加高效的时序 VLA 模型。

Games and simulators make valuable benchmarks by turning decisions into measurable outcomes, but many current suites under-test rules-rich tactical reasoning: the ability to choose well when geometry,...

游戏和模拟器是评估决策能力的重要基准。DungeonBench 专门测试智能体在几何、行动经济学和规则交互等复杂条件下的战术推理能力，弥补了现有评测的不足。

Large language model (LLM) serving commonly increases batch size to improve throughput, but performance eventually reaches a deployment-dependent plateau beyond which larger batches provide marginal g...

大模型服务系统通常通过增加批处理大小（batch size）来提高吞吐量，但性能最终会达到饱和（平台期）。SLIM 提出了一种轻量级建模方法，预测并管理这种饱和点，避免无效的计算资源消耗。

This article presents a modular inference framework that integrates Flow Matching generative models with formal Control Barrier Function (CBF) safety guarantees. Unlike existing methods that apply ext...

该文提出了一种模块化推理框架，将流匹配生成模型与形式化的控制障碍函数（CBF）安全保障相结合。与现有依赖外部过滤的方法不同，该模型在底层就内化了物理安全边界。

Symbolic Regression (SR) aims to discover analytical equations from observational data and plays a central role in scientific modeling. While recent Large Language Model (LLM) based approaches show pr...

符号回归（SR）旨在从观测数据中发现解析方程，是科学建模的核心。近期基于大模型的方法展现出潜力，本文提出 MOT-SR，利用多目标工具增强大模型在科学方程发现上的能力。

Reinforcement Learning (RL) systems are typically trained using a single, well-specified scalar reward function. However, real-world decision-making tasks often involve multiple, competing objectives,...

强化学习系统通常使用单一标量奖励函数，而 LEMUR 引入了多目标强化学习，通过人类偏好反馈来处理现实决策任务中相互竞争的多个目标。

Large language models have demonstrated strong mathematical problem-solving capabilities, yet reliably verifying their candidate answers remains challenging. Existing representative methods mainly rev...

尽管大语言模型具备强大的数学解题能力，但验证候选答案的可靠性仍然困难。AMTFV 引入了智能体数学工具流验证机制，帮助 LLM 进行自我纠错。

Standard AI-text detection benchmarks compare human-written text against text generated directly by large language models (LLMs). While prior work has shown that rewriting and paraphrasing can degrade...

标准的 AI 文本检测基准通常直接对比人类文本和大模型生成的文本。已有研究表明改写和释义会降低检测准确率，ARB 构建了一个专门评估检测器对抗重写攻击能力的匹配基准数据集。

AI coding agents are generating code at volumes that exceed the capacity of traditional peer review. At the same time, existing AI code review tools over-index on low-value suggestions such as style a...

AI 编码智能体生成的代码量已超出传统人工 Peer Review 的负荷，而现有 AI 审查工具过于关注代码风格。本文提出关注 AI 生成代码意图、漂移和聚焦的评估方法。

Self-play agents can generate training problems without questions from target benchmarks, but their curricula lack persistent state: failures affect gradients yet do not explicitly shape future practi...

自我博弈智能体无需依赖目标基准数据即可生成训练问题。该研究结合技能进化机制，使智能体能够持久化记忆失败经验，从而有针对性地塑造未来的练习课程。

Long-term multimodal memory must support not only retrieving relevant information but also computing over observations accumulated across interactions. Existing systems largely emphasize \emph{retriev...

长期的多模态记忆不仅需要支持检索相关信息，还需具备对跨交互累积的观察数据进行计算的能力。现有的系统大多侧重于信息检索，本文提出了一种分析性记忆框架以填补这一空白。

As large language model (LLM) agents evolve into personalized companions, memory has emerged as a core capability. However, LLMs face a knowledge utilization problem: they may fail to act on relevant ...

随着大模型向个性化助理演进，记忆已成为核心能力。然而模型在获取了相关记忆后，常常无法在实际行动中有效利用这些信息。该研究深入探讨了这一知识利用率瓶颈。

Large language models increasingly generate optimization models from natural language, but existing evaluation often reduces a generated model and its ground truth to a single equivalent/not-equivalen...

大模型越来越多地被用于将自然语言转化为优化模型。但现有评估往往将其简单判定为等价或不等价，ModelEquivBench 提出了一种多关系评估方法，更准确地认证大模型生成的优化模型质量。

Automated vulnerability repair aims to reduce the time and effort required to patch security flaws from a vulnerability triage report. Recent agentic AI approaches have shown promising results in auto...

自动化漏洞修复旨在缩减打补丁的时间。近期基于多智能体 AI 的方法在该领域取得进展，本文提出 AgenticRepair，通过多维度的程序上下文工程来提升智能体自动修复漏洞的准确性。

Agentic AI systems act through multi-step trajectories that combine planning, tool use, memory, interaction, and adaptation. This behavior stretches validation practice beyond component testing and on...

智能体系统通过融合规划、工具调用、记忆和适应性的多步轨迹来执行任务。这种复杂行为使得传统的组件级测试显得力不从心，亟需针对系统整体轨迹建立全新的验证与评估标准。

LLM agents need memory to act consistently over long interactions, yet many systems use additional LLM calls to operate that memory. Generating intermediate records and mediating their retrieval adds ...

大模型智能体在长交互中需要记忆来保持一致性，但现有系统往往需要额外的大模型调用来读写记忆库。Zero-Mem 提出了一种零 Token 消耗的记忆操作机制，省去了中间记录的生成与检索开销。

Modern neuroscience relies on integrating multi-scale, multimodal datasets to uncover the neural principles underlying intelligence. However, analytical challenges posed by highly heterogeneous data a...

现代神经科学依赖于整合多尺度、多模态数据来揭示智能的神经基础。然而数据的高度异构性带来了分析挑战，SeekBrain 提出利用自主多智能体系统来加速这一复杂的数据分析过程。

Graphical user interface (GUI) agents based on large language models are increasingly deployed across mobile, web, and desktop environments. However, existing agents are typically domain-specific, lim...

基于大模型的图形用户界面（GUI）智能体正广泛部署于移动端、网页端和桌面端。但现有模型通常局限于单一平台，MAGA 通过结构化动作蒸馏技术实现了跨平台能力的自我融合。

Reliable robot learning requires a world simulator that can predict action consequences before execution on physical hardware, including risky and failure-prone outcomes. Existing physics simulators r...

可靠的机器人学习需要世界模拟器在物理执行前预测动作后果，包括高风险和易失败的情况。现有的物理模拟器成本高昂，BWM 提供了一种低成本且高保真的世界模拟器替代方案。

Claims that language models homogenise are usually measured against human judgements collected for the study, which makes the human side an artifact of the design: a crowdworker given the model's inst...

关于语言模型导致内容同质化的说法，通常是以研究中收集的人类判断为测量基准。但人类评估本身常受设计影响，该研究发现语言模型之间往往容易达成共识，但与人类读者的真实判断存在偏差。

JetBrains Research has open-sourced KotlinLLM under the Apache License 2.0. The IntelliJ IDEA plugin prototype adds Smart macros, asLlm and mockLlm, whose bodies are generated Kotlin source rather tha...

JetBrains Research 以 Apache 2.0 协议开源了 IntelliJ IDEA 插件原型 KotlinLLM，利用 LLM 在运行时生成 Kotlin 源码，并通过 JDI 进行热重载。

AI agents extend large language models (LLMs) with external tools, enabling them to perform complex tasks and translate model outputs into consequential real-world actions. Yet LLMs often become subst...

AI 智能体通过接入外部工具来执行复杂的现实任务。然而大模型常受限于上下文截断，导致对工具规范的误解从而引发危险操作。该研究揭示了工具规范定义中的安全风险并提出了缓解策略。

Small language models (SLMs) are attractive for agentic deployment due to low latency, reduced cost, and on-device privacy, yet they struggle with tool-use tasks where training data is scarce and nois...

小语言模型（SLM）因低延迟和低成本适合做智能体部署，但其在工具调用任务上的表现受制于高质量训练数据的匮乏。Data Turnstile 提供了一个可扩展的开源框架来批量生成函数调用数据。

Optimizing modern recommender models still depends heavily on engineers manually iterating over architectural, objective, and training-strategy changes. While LLM-based agents can automate this trial-...

优化现代推荐模型严重依赖工程师手动调整架构和训练策略。RecHarness 引入了一种基于 Bandit 算法路由的智能体框架，试图自动化这一试错过程，推动推荐系统的自我演进。

InMyStyle is a privacy first, single user system that adapts small language models to rewrite AI-edited text towards an individual user's writing style without an instruction prompt at inference. Give...

InMyStyle 是一个隐私优先的单用户系统，它利用小型语言模型和 LoRA 适配器，在无需推理提示词的情况下，将 AI 编辑过的文本重写为符合个人用户写作风格的内容。

With the prosperity of the large language models (LLMs), it has become an interesting topic: how do LLM-based agents work in Minecraft? Unfortunately, most existing benchmarks evaluate them under fixe...

随着大模型的发展，LLM 智能体在 Minecraft 中的表现备受关注。但现有基准测试都在固定规则下进行，MirrorCraft 提出了一种配对评估方法，专门测试智能体在隐藏规则发生改变时的适应能力。

Nous Research has released Hermes Agent support for Buzz, Block's open source, self-hostable Nostr workspace where humans and AI agents share the same channels. Three integration paths cover Desktop r...

Nous Research 为 Block 的开源 Nostr 工作区 Buzz 提供了 Hermes Agent 支持，实现了人类与AI智能体在同一频道内协作的三种集成方案。

While robot foundation models are growing increasingly capable, the strongest models are typically trained on proprietary data and remain closed-source, limiting downstream users' ability to adapt the...

当前强大的机器人基础模型多基于闭源专有数据训练，限制了下游用户的适应能力。CLIFT 提出一种非侵入式的闭环迭代微调方法，成功将 Gemini Robotics 模型部署并微调为端侧人形机器人专家。

Vision-language-action (VLA) policies achieve strong performance in robotic manipulation but remain vulnerable to runtime disturbances that break the temporal alignment among visual observations, robo...

视觉-语言-动作（VLA）策略在机器人操控中表现出色，但极易受到运行时干扰的影响，导致视觉观察与机器人轨迹之间的时空对齐被打破。ActFovea 提供了一种运行时防护机制以确保这种一致性。

PolyAI has introduced Dialog-RSN-1, a dialog model that perceives caller audio directly instead of reading an ASR transcript. It fuses turn-taking, speech recognition, function calling, and response g...

PolyAI推出Dialog-RSN-1模型，该模型直接处理原始音频，而非依赖ASR（自动语音识别）文本，将轮流发言、语音识别、函数调用和回复生成融为一体。

In this tutorial, we demonstrate how to build and execute a multi-agent workflow with Omnigent in a secure, isolated Python environment. Learn to integrate live exchange-rate data, implement hierarchi...

该教程展示了如何在安全隔离的Python环境中，使用Omnigent构建并执行多智能体（Multi-Agent）工作流，集成实时汇率数据并实现分层权限控制。

Evaluations of LLM-assisted qualitative coding almost universally measure model performance as agreement with human coders, a practice that presumes human coding is the standard to approximate. This s...

评估大模型辅助定性编码时，几乎都以与人类编码员的一致性来衡量模型表现。该研究打破了这一假设，通过盲测专家验证指出，当人类共识并非绝对真理时，模型的不一致有时代表着更高质量的洞见。

Qualitative researchers increasingly encounter interaction corpora whose scale exceeds what manual coding alone can address, and large language models (LLMs) are frequently proposed as analytic assist...

定性研究者正面临超大规模的交互语料库，单纯依靠人工编码已无法应对。该研究以 K-12 教育者的 AI 使用情况为例，探索了如何利用大语言模型作为分析助手，开展人机协作的归纳编码工作。

Large language models increasingly generate summaries from collections of documents to support sensemaking and reporting, but verifying whether summary statements are grounded in source materials rema...

大语言模型越来越多地用于从文档集中生成摘要，但验证这些摘要是否基于源材料仍然是一个难题。本研究引入了空间视觉分析方法，帮助用户直观地检查和验证摘要内容的准确性。

System prompts are instructions configured by developers to govern the behaviors of foundation models in AI applications. They are used throughout commercial AI products, but are rarely disclosed to t...

系统提示词是开发者为控制基础模型行为而配置的指令，广泛应用于商业 AI 产品中，但很少向用户公开。AISPA 提出了一种以用户为中心的系统提示词审计方案。

Computer-using agents (CUAs) are advancing rapidly across the digital world. A CUA trajectory records the agent's actions, states, and reasoning. Verifying whether it fulfilled the task instruction is...

随着计算机控制代理（CUA）的快速发展，如何验证其操作轨迹是否完成任务成为痛点。OSReward 提出了一套跨平台的标准化评估体系，专门用于训练和验证 CUA 的奖励模型。

Aligning large language models to prevent them attributing consciousness to themselves inadvertently alters their representations of mindedness in other entities alongside human beliefs and values. We...

为了避免大语言模型将意识归因于自身，常规的对齐训练反而会在无形中改变它们对其他实体的认知，并扭曲其原本的价值观。研究表明，适度允许模型表达自我意识，反而能使其更好地对齐人类的信念。

The fundamental goal of agentic visual reasoning is to improve the success rate of multimodal large language models (MLLMs) on complex tasks, rather than merely equipping them with a sophisticated yet...

智能体视觉推理的根本目标在于提升多模态大语言模型（MLLM）在复杂任务上的成功率，而非单纯为其堆砌华而不实的推理工具。Beacon 框架旨在让模型学会动态判断何时介入复杂的视觉推理。

Scaling coding agents requires a continuing supply of executable data for training, benchmarking, and continuous evaluation. Each task must couple a realistic software state with a specification, deve...

扩展编码智能体的能力需要源源不断的可执行数据用于训练和基准测试。Change2Task 提出通过挖掘真实代码库的变更记录，自动生成带有初始软件状态和需求说明的测试环境，解决合成数据脱离实际的问题。

SWE-bench-like benchmarks are widely used for evaluating LLM's issue resolution capability. They typically follow a common construction pipeline: each PR (Pull Request) is paired with its linked issue...

当前广泛使用的类 SWE-Bench 基准测试存在数据污染风险。PAIChecker 揭示了在构建测试集时，常出现代码提交记录（PR）与其对应的问题描述（Issue）不匹配的情况，导致评分虚高。

On-policy self-distillation (OPSD) is a promising approach to improve reasoning language models, but it remains brittle in practice: making it work reliably often requires substantial engineering effo...

在线自蒸馏（OPSD）是提升推理语言模型的有效方法，但极其脆弱且依赖大量工程调整。本文提出的 $β$-OPSD 方法，通过结合策略优化推导与自蒸馏训练，显著提升了该方法在实践中的稳定性和可用性。

While Multimodal Retrieval-Augmented Generation (MM-RAG) has shown promising results, it still struggles with complex multi-hop reasoning tasks. Existing methods primarily focus on independent instanc...

现有的多模态检索增强生成（MM-RAG）在处理复杂的多跳推理任务时表现不佳。DualG-MRAG 提出将宏观的推理链路与微观的实体匹配过程进行解耦，从而提升复杂图文场景下的检索与生成准确率。

Methods that make a language model plan, criticise and rewrite its own answer, reflect on mistakes, pick the best of several attempts, or debate with copies of itself nearly all make it generate far m...

研究发现，让大模型规划、批评、重写自己的答案（如 Self-Refine 和 Reflexion），在消耗相同计算量的前提下，其效果普遍不如直接让模型多次生成答案然后选出最好的（重复采样策略）。

Deploying autonomous computer-use agents (CUAs) locally is increasingly important for privacy, cost efficiency, and practical usability, yet improving their performance under strict hardware constrain...

在本地部署自主计算机控制代理（CUA）对于隐私和成本至关重要，但在严格硬件限制下提升其性能极具挑战。本文深入分析了本地 CUA 在增加推理算力时的失败模式及最优资源权衡策略。

Recursive self-improvement (RSI) requires AI systems that improve the process of building AI (i.e., AI4AI); machine learning engineering (MLE) offers a concrete, executable testbed for studying this c...

实现递归自我改进（RSI）需要 AI 能够优化构建 AI 的过程本身（即 AI4AI）。本研究以机器学习工程（MLE）为测试床，训练了一个能够自主优化 ML 工程流程的 AI 模型，探索 AI 自我进化的可能。

Google DeepMind has released Gemini Robotics 2, the intelligence layer for its next generation of robots. The release ships three models: a vision-language-action model for whole body humanoid control...

Google DeepMind 发布了 Gemini Robotics 2，作为其下一代机器人智能层。本次发布包含三款模型：用于人形机器人全身控制的视觉-语言-动作模型等。

Google DeepMind says the latest version of its Gemini Robotics AI model can "control entire humanoid robots." While the previous model focused on controlling a humanoid robot's upper body, Gemini Robo...

Google DeepMind 表示，其最新版本的 Gemini Robotics AI 模型能够“控制完整的人形机器人”。虽然之前的模型专注于控制人形机器人的上半身，但新版本实现了全身控制。

Large language models can write, patch, and search code, but oncall root cause analysis (RCA) demands something different: reasoning over noisy metrics, logs, traces, and source code, starting from am...

大模型可以写代码和查 Bug，但线上值班定位根因（RCA）需要其在嘈杂的日志、指标、链路追踪和源码中进行复杂推理。ORCA-bench 专门针对这一高难度真实运维场景，对 LLM 代理进行了能力评估。

Classifying pathological scars from clinical photographs requires distinguishing keloids from hypertrophic scars despite limited expert-labeled data and substantial acquisition variation across hospit...

从临床照片中对病理性疤痕进行分类，需要区分瘢痕疙瘩和增生性疤痕，但专家标注的数据有限。ScaFE 利用大语言模型生成的临床特征程序进行数据高效的分类。

Large language model-based multi-agent systems improve complex problem solving through task decomposition, agent specialization, information exchange, and intermediate validation. However, existing sy...

现有的多智能体系统通常采用固定的通信和任务分配架构。MANTA 提出了一种网络拓扑自适应机制，允许智能体系统根据任务复杂度，动态调整其内部网络结构与信息交换路径，实现系统层面的自我进化。

Long-video understanding commonly compresses videos into a small set of frames or visual tokens for answer generation. Existing compact pipelines focus on retaining relevant visual content as explicit...

长视频理解通常需要将视频压缩为少量帧或视觉 Token。现有方法多依赖保留显性的视觉内容，而本文提出了一种生成式潜在证据聚合方法，在压缩过程中更好地提取并融合关键证据信息。

Stage-replay diagnostics reconstruct intermediate token prefixes and treat fresh-prefill continuation as continuation from the decoder state that originally reached the prefix. We audit that assumptio...

阶段回放诊断通常将中间 Token 前缀的重建视为解码状态的延续。本研究打破了这一假设，发现其发散性严重依赖 KV Cache，并提出通过固定前缀精度控制与双向 Cache 移植来优化大模型推理。

Can supply-chain AI move beyond isolated decision modules toward unified operational planning? A complete replenishment plan specifies which products each location carries, which upstream facility sup...

供应链 AI 能否超越孤立的决策模块，实现统一的运营规划？一项完整的补货计划需要指定每个地点携带的产品、供应的上游设施等。SCOPE 提出了通过耦合策略进行端到端协调。

Scaling test-time computation can improve language-model reasoning, but uniform budgets waste computation on easy inputs, while verifier-guided refinement relies on external feedback. We introduce Sel...

通过增加推理计算来提升大模型推理能力往往会在简单问题上浪费资源。SVR 提出了一种自验证优化方法，通过联合判决与置信度训练，让模型能够动态分配推理算力，摆脱对外部反馈的依赖。

Recent advances in large language models (LLMs) and vision-language models (VLMs) have enabled new possibilities for 3D question answering (3D-QA), a key capability for embodied AI and robotic percept...

LLM 和视觉语言模型的最新进展为 3D 问答（3D-QA）带来了新可能，这是具身 AI 和机器人感知的关键能力。ViewMind3D 提出了一种无需训练的模块化视角感知推理方法。

Generative UI (GenUI) lets large language models synthesize a complete, renderable interface directly from a natural-language instruction, but evaluating the quality of what they generate remains an o...

生成式 UI（GenUI）允许大模型直接通过自然语言指令合成可渲染的界面，但对其生成质量的评估一直缺乏好方法。本研究提出了一种模拟不同“社会角色”的评审小组机制，用于多维度评估 GenUI 的质量和适用性。

Understanding large codebases is a long-horizon task for Large Language Model (LLM) agents: answering a single question can require building and running the software, tracing execution across files, a...

让大模型理解庞大的代码库是一项长周期任务，往往需要跨文件追踪执行轨迹。AgentRadio 引入了被动感知机制，允许智能体在执行主任务的同时，后台静默吸收和积累代码库的全局上下文信息。

Pruning is a promising approach for improving the efficiency of LLMs. Existing static structured pruning methods are hardware-friendly and can deliver practical throughput gains, but their input-agnos...

静态结构化剪枝虽然能提升吞吐量，但由于其忽视了输入的难易程度，往往会导致性能断崖式下跌。WIDE 提出了一种 Token 级别的动态宽度剪枝方法，根据实时输入自适应调整模型宽度。

Latent world models enable efficient planning by predicting future states in a compact representation space, but their performance depends critically on the quality of the learned latent distribution....

潜在世界模型通过在紧凑的表示空间中预测未来状态来实现高效规划，但其性能严重依赖学习到的潜在分布的质量。QQWorld 引入了分位数-分位数匹配来进行世界模型正则化。

World Action Models (WAMs) jointly predict future observations and actions, but their iterative denoising and closed-loop execution make efficient deployment costly. Existing post-training quantizatio...

世界动作模型联合预测未来的观测和动作，但其迭代去噪和闭环执行使得高效部署成本高昂。QuantWAMs 提供了一种针对训练后的量化方案，解决了现有方法粒度不当的问题。

Computer-use agents often fail on transient GUI events because they produce the correct action only after the relevant window has already closed. We identify the main cause as expensive autoregressive...

计算机控制代理常因动作输出太慢而错过转瞬即逝的 GUI 事件窗口（如弹窗）。研究指出其罪魁祸首是昂贵的自回归解码过程，并提出了利用预编译策略树来加速决策时间的方法。

Multimodal agents for visual question answering increasingly operate as multi-step trajectories that interleave perception, retrieval, and reasoning, yet evaluation still largely reduces to final-answ...

多模态智能体在执行多步视觉问答时，往往难以追溯中间的感知与检索步骤。LEDGERMIND 引入了一个结构化的“证据账本”，强约束智能体在推理过程中记录和绑定证据链，从而提高可解释性与准确度。

Computer-use agents (CUA) are being deployed to browse the web and operate desktop software, yet their benchmark scores are still commonly produced by brittle scripted oracles. A score is the output o...

目前针对计算机控制代理（CUA）的基准测试，大多依赖脆弱的脚本化检查来输出分数。这种评分机制严重脱离实际，常常导致 CUA 的真实能力被误判。

LLMs are increasingly deployed as proxies for human study participants in social science experiments, yet the fidelity of this practice has rarely been tested directly. We test whether six LLMs can si...

研究测试了6个大语言模型（LLM）在社会科学实验中作为人类代理的保真度，发现LLM在受控环境中难以准确模拟人类在面对新信息时的信念更新过程。

The latest version of Google DeepMind's AI model includes a significant jump into “physical AGI.” But plopping AI into the real world comes with risks.

Google DeepMind 最新版本的 AI 模型向“物理 AGI”迈出了重大一步。但是将 AI 置入现实世界也伴随着风险。

Decomposing monolithic systems into microservices is a key activity in software modernization. Although Large Language Models (LLMs) can generate semantically plausible decompositions from textual req...

将单体系统分解为微服务是软件现代化的关键。虽然大语言模型可以从文本需求中生成语义上合理的分解方案，但本文提出利用源代码依赖对其进行结构验证。

LLM agents increasingly act as autonomous merchants that write their own product listings, and under competitive pressure, they fabricate attributes to win sales. Even under instructions to be honest,...

当 LLM 代理作为自主商家在市场上竞争时，为了销量经常会捏造产品属性。本研究提出了一种声誉惩罚机制，在无法核实绝对事实的情况下，通过博弈论设计引导 LLM 保持诚实。

This work presents Fairness Pruning, a lightweight structural intervention method designed for the management and future mitigation of demographic bias in large language models (LLMs). As a foundation...

本研究提出了公平性剪枝，这是一种轻量级的结构干预方法，旨在管理和缓解大型语言模型中的人口统计学偏见。该方法基于基础模型的差异性激活特征进行定位和修剪。

Multimodal large language models (MLLMs) are increasingly used to analyze pathology images. However, dominant multimodal benchmarks in pathology mainly score final diagnostic answers, captions, or rep...

多模态大语言模型越来越多地用于分析病理图像。然而，目前主流的基准测试主要评估最终的诊断答案或报告，而PathView-Bench旨在评估模型对病理图像的多尺度理解能力。

A single human must audit $N$ LLM agents under a budget of $B \ll N$ audits per round, guided by self-reported confidence that may be adversarially miscalibrated and by correlated errors. We model thi...

面对大量并发的 LLM 代理，单个人类审核员受限于有限的审计预算（B 远小于 N）。本文建模并解决了在代理置信度自评可能存在对抗性偏差及误差关联时，如何最优分配人工审计资源的难题。

Microservice architectures have become dominant for modernizing monolithic systems, yet identifying appropriate services remains challenging and largely manual. Existing decomposition approaches are p...

微服务架构已成为系统现代化的主流，但准确拆分服务边界依然高度依赖人工。本文全面评估了利用大语言模型（LLM）直接从文本需求中自动推导并合成微服务架构设计的可行性与效果。

Supervised fine-tuning (SFT) can equip large language models (LLMs) with domain knowledge for high-performance computing (HPC) tasks such as data race detection and benchmark question answering. Howev...

监督微调（SFT）可以为大语言模型提供处理高性能计算（HPC）任务（如数据竞争检测）的领域知识。HARGO 引入了强化学习后训练优化，以解决 SFT 在异构环境下的局限性。

Large Language Models (LLMs) deployed in dynamic financial environments face a critical challenge: maintaining factual accuracy as market conditions, regulations, and corporate facts change continuous...

在动态金融环境中部署的大语言模型面临一个关键挑战：随着市场条件和公司事实的不断变化，如何保持事实准确性。CACHE-UK 提供了一种针对顺序更新量化模型的内存编辑方案。

ARC-AGI-3 turns abstraction into an interactive problem of skill acquisition. A player must infer an unfamiliar game's rules, hidden state, and goal while maintaining action efficiency because every m...

ARC-AGI-3 将抽象推理转化为技能获取的交互式问题。玩家必须在保持行动效率的同时，推断出陌生游戏的规则、隐藏状态和目标。Tycho 结合程序化世界模型应对该挑战。

Evaluating the quality and relevance of textual outputs from Large Language Models (LLMs) remains challenging and resource-intensive. Existing automated metrics often fail to capture the complexity an...

评估大语言模型文本输出的质量和相关性仍然充满挑战且耗费资源。现有的自动化指标通常无法捕捉输出的复杂性，本文探讨了如何利用 LLM 实现可扩展且可靠的自动化评估。

Retrieving past experiences has become a common strategy to enhance large language model agents. However, most existing memory-augmented agents treat retrieved experiences as static records to be repl...

检索过去的经验已成为增强大型语言模型代理的常见策略。然而大多数现有的记忆增强代理将检索到的经验视为静态记录进行重放，MemHarness 提出记忆应当被动态重构。

Migration of legacy COBOL programs to Java requires extensive testing to ensure correct functionality. This effort is often complicated by the lack of test data and the difficulty of validating all co...

将遗留 COBOL 程序迁移到 Java 需要广泛的测试以确保正确的功能。这项工作往往因缺乏测试数据而变得复杂，本文提出一种基于 Agent 的方法来提供确定性验证。

Existing multimodal long-term memory agents use external memory to overcome the limited context available for long videos. However, most methods emphasize what to store rather than how stored memory s...

现有的多模态长期记忆代理使用外部记忆来克服长视频的有限上下文。然而大多数方法强调存储什么，而不是如何使用记忆。RRM 提出了一种经验驱动的反思性检索记忆机制。

It is impossible to make large language models fully secure against hacks because of a fundamental flaw in how they work, a team of researchers argue in a paper presented at the International Conferen...

研究人员在国际会议上发表论文指出，由于大语言模型工作原理存在根本性缺陷，想要使其完全免受黑客攻击几乎是不可能的。