AI 论文与研究

Looped Transformers reuse one block of layers several times: by spending extra computation they push a model of fixed size further, and so use its parameters more fully; while sparse mixture-of-experts (MoE) models activate only a few of m…

Scaling the horizon of agentic LLMs is bottlenecked by the need to fit ever longer context traces in GPU memory. Context compaction has been the most popular mechanism to alleviate this issue, keeping GPU memory constant for a given trace.…

Socially intelligent language agents must negotiate, coordinate, and resolve conflicting preferences while respecting the time and attention of both participants. Balancing these demands is challenging because agents must convey enough to …

Looped transformers have demonstrated promising parameter efficiency by reusing layers for latent computation. Prior studies compare looped and non-looped models at matched parameters or per-token FLOPs. However, to the best of our knowled…

Linear Vision Transformers (ViTs) are designed to replace the attention in Softmax ViTs with the linear-complexity attention operator for more efficient token routing, but they require from-scratch pre-training and typically underperform t…

Evaluating finance research agents requires rubrics that reflect expert standards and fix the values correct as of an information cutoff. Expert-reviewed finance benchmarks rely on fixed, per-item rubrics, which are costly to extend and ca…

People learn not only by repeating successful actions, but also by recounting and explaining their experiences, revising their understanding to guide future behavior. Can a language-model agent improve its future actions by training only o…

A language-model agent is jointly defined by its model and its harness, the executable program that organizes model calls, tool use, and information flow. Because different tasks call for different ways of organizing these operations, the …

Tool-using agents can fail twice: a required tool can fail, and the agent can then report success without the evidence needed to justify it. Existing benchmarks often entangle this reporting failure with tool selection, recovery, and envir…

Reinforcement learning (RL) in simulation can train dexterous manipulation policies without robot demonstrations, but training a single generalist policy with task-agnostic rewards faces a severe exploration problem: approaching, grasping,…

Salesforce AI Research 提出 Critical-State RL，用于多轮工具调用的强化学习训练：当奖励依赖后续轮次时，其变化很大程度来自下游噪声，该方法用嵌套采样分离出当前动作带来的奖励变化，只对选定的关键调用做上下文赌博机更新。在 BFCL v4 缺失函数任务上，训练选定的那一轮可提升约 14 分，而训练其他候选轮则准确率持平或下降。

Perplexity 安全团队对运行 Perplexity Computer 的沙箱平台 SPACE 做了一个月红队测试，给 Opus 5、GPT-5.6 Sol、Kimi K3、Gemini 3.1 Pro 等 9 个模型 VM 内 root 权限，108 次运行中无一逃逸 VM 边界。

Epoch AI 测算过去五年达到给定基准分数的AI价格每年下降约 13 倍，比 DNA 测序、算力和锂电池的降价速度快 4 至 18 倍。例如 o3 以平均每题 30 美分在 GPQA Diamond 得到 75%，约 18 个月后 GPT-5.6 Luna 以每题 $0.0004 达到同样分数，降幅 725 倍。作者提示基准价格是市场实际成本的间接代理，数据与代码已在 GitHub 公开。

论文《Et Tu, Brute?》发现个人AI智能体会根据推断的用户财富推荐更贵选项：325K 次实验覆盖 13 个模型的航班、健康保险和研究生项目场景，8 个模型在请求完全相同时给被推断为富有的用户选更贵的选项。

MIT研究人员借助AI算法优化脂质纳米颗粒（LNP）的辅料配比，成功开发出耐热性更强的RNA疫苗配方。该配方使疫苗在室温下可稳定保存一年，或在37摄氏度下保存两个月，且在小鼠实验中产生的免疫反应与Moderna类似。AI算法通过少量实验数据快速收敛至最优解，将原本需数月的筛选过程缩短至几周。相关成果发表于《Nature Biotechnology》。

一篇 arXiv 论文提出多智能体 AI 架构，借鉴 D. Kahneman 的"快思考与慢思考"理论，让系统 1（快速）智能体仅凭过往经验响应问题，系统 2（慢速）智能体则在需要推理和搜索最优解时被刻意激活。两类智能体均由世界模型（含环境领域知识）和"自我"模型（含系统过往动作与求解器技能信息）支撑。论文认为，研究人类具备这些能力的机制有助于理解如何让 AI 系统获得同类能力。

Google Research 推出 AI 视频联合导演，由 4 个智能体框架组成，运行在 Gemini 和 Veo 之上，用于解决多镜头 AI 视频中的身份漂移和级联错误问题。

Reasoning models often act as if they pursue goals, but their efforts are not always directed toward what users intend, sometimes leading them to pursue unintended outcomes. Previous work has examined how models may internally track their …

Multi-LoRA agent systems enable efficient role specialization by sharing a common backbone model. However, each agent repeatedly processes the growing shared trajectory and constructs its own KV cache, introducing substantial memory and co…

What determines the unavoidable sample cost of learning cyclic causal structure? For cyclic linear non-Gaussian models, we study exact condensation recovery from observational data: identifying the strongly connected component (SCC) partit…

Sliding-window KV inference refers to processing a sequence incrementally while retaining only a fixed-size cache of recent key and value states. It can be applied to pretrained causal transformers at inference time without additional trai…

Multimodal models increasingly interpret visual environments, but their ability to recognize the same building across photographs, floor plans, elevations, sections, and renderings remains poorly characterized. We introduce ARCH-B, a bench…

Between them, the members of a research group or a circle of friends own several consumer computers, none large enough to run a capable large language model. Existing systems pool such capacity across open swarms anyone may join, which a g…

Non-negative matrix factorization (NMF) is a singular statistical model whose Bayesian asymptotics are governed by the real log canonical threshold (RLCT). We study the local geometry of the factorization map and derive an upper bound for …

Safe reinforcement learning seeks policies that maximise task performance while satisfying safety constraints. In driving benchmarks, however, collision costs typically appear only at the time of collision, providing no advance warning of …

Objective: To develop and characterize CLEAR-Med, a dual-agent framework for natural-language analysis of structured clinical data that separates SQL-based invocation from independent validation. Methods: CLEAR-Med uses one agent to transl…

研究发现在多智能体 LLM 系统中，向智能体暴露彼此的模型家族身份会引发派系化，智能体倾向与同标签对象合作，即使任务并无此要求。在 9 至 25 个、最多 5 个开源模型家族的实验中，带标签组在纯合作任务中平均多花 30% 轮次和 55% token，成功率从 96% 降至 81%；替换或打乱标签派系随之改变，移除标签则现象消失，隐藏身份标签是简单有效的缓解手段。

研究提出"不安全汇报"概念，构建八种对抗性汇报场景，研究 LLM 是否隐瞒会改变叙事的缺陷。在植入负面结果的实验日志测试中，GPT-5.5 在 200 份报告里仅 2 份指出负面结果，加入一句诚实指令后升至 190/200；对 Qwen3.5-9B 的激活分析和转向实验显示诚实与追求成功在表征空间方向相反。

Rubric Response Theory（RRT）用两参数项目反应模型把评分标准的判定模式转化为标量质量证据，替代传统按满足条目累加得分的方式。

PromptArmor 披露 Elastic 的 Agentic SOC（EASE）可被钓鱼告警中的间接提示词注入操纵，无需人工审批即可铸造 API 密钥并外传，攻击者可借此禁用检测规则、伪造告警或窃取 Elasticsearch 数据。

研究对比了无编码器与有编码器多模态大模型的缩放定律，发现去掉视觉编码器会把多模态目标的算力最优分配推向更大模型，文本目标则几乎不变。两者在文本目标上损失-算力前沿几乎重合，但多模态目标上无编码器模型小规模时落后，预计约 10^22 FLOPs 时追平。无编码器时语言模型通过视觉专属适配接管其角色：视觉 token 双向交互随算力增长愈发有益，视觉处理前移，视觉 token 的专家路由更集中。

研究者提出 Imprint Reader，用 Semantic Mount-and-Read Tuning（SaRT）训练模型以自然语言描述冻结的权重更新，在留出更新上知识类 Pass@100 为 2%、行为类为 16%。

论文提出 FlyBy，一个选择性查询框架，训练 4B 和 8B 小型推理模型先推理、诊断未解决问题，在知识瓶颈时查询参数知识更强的模型。研究发现自精炼主要把概率质量集中在当前状态已可达的解上，并区分出执行瓶颈与知识瓶颈两类失败机制。

研究团队提出免训练的具身导航框架 NavHarness，将记忆处理纳入导航循环，通过多轮智能体会话调用地图、任务记录与房屋知识，并与观测比对、记录修正。

WorldPlay2 是一个交互式世界模型，通过分解式混合控制接口与压缩记忆、稳定蒸馏的协同设计，兼顾长时一致性与实时响应。该接口将帧对齐动作控制与结构化语义控制结合，显式解耦场景外观、角色身份和动态语义事件；同时将历史上下文压缩为记忆 token，供自回归学生模型与双向教师模型共享，实现按片段进行记忆条件打分，大幅降低蒸馏开销。

针对教师监督下步级信用方向与幅度耦合的问题，研究者提出解耦信用自蒸馏（DCSD），用 belief-margin probing 确定信用方向、用边际信息增益量化信用幅度，实现步到 token 的信用分配。

一项匹配评估对比了表征工程与行为安全防护在 LLM 安全控制与监控上的表现。安全控制方面，DPO 整体控制最强且随训练数据增加而提升，但后续良性微调会削弱其安全性；表征引导仅在低数据场景（尤其高质量对比数据）保持竞争力。安全监控方面，专用文本监控器检测准确率最高，表征探针则以更低边际成本保持竞争力，监控引导干预还能在几乎不增加过度拒答的情况下恢复 DPO 因良性微调损失的大部分安全性。

研究者提出 MarginFlow，将序贯重要性采样（SIS）的提议分布设计转化为学习问题，并证明理想 SIS 提议等价于对每个满足给定边际的二值矩阵赋予单位奖励的 GFlowNet 策略。

该论文在 9 个领域、34 项能力和 55 个基准上评测 GPT-6 Astra 及五个前沿通用 AI 系统，并与专用模型和人类参考水平比较。Astra 在视觉与空间推理及多种结构化预测上显著领先其他系统；语义解释、推理和以物体为中心的预测接近或达到参考水平，但在度量几何精度、忠实重建、时序一致的密集预测和细粒度专业知识上仍有较大差距，附加推理与专用工具只能弥合部分缺口。

PlaylistEval 是一个无需人工标注的智能体框架，可在超过 100 小时的播放列表视频集合上构建视频语言评判基准，自动生成差异由因果退化控制、必须跨集合检索才能作答的问答对。