AI 论文与研究

DAEDALUS 提出一种无需既有任务或 oracle 验证器、从自生成练习中自举可复用智能体记忆的方法：explorer 智能体生成有挑战性且可解的任务，solver 智能体尝试并从失败中提炼启发式，经反复验证后写入记忆库供测试时使用。

Google Research 提出用地学基础模型 PDFM 的位置嵌入作为即插即用输入，融入现有流行病学工作流。嵌入由隐私保护的搜索趋势、人流、建成环境密度和环境指标自监督合成，按月更新。

NeMo-DCR 是一种比特精确的增量压缩重配（refit）方案，只传输变化的权重，接收端得到的参数与缓冲区比特与全量重配完全一致。在 3% 和 5% 变化率下，30B-1T 模型的重配速度比仅传输全量 checkpoint 的基线快 12-40 倍；1T 模型经中继树在 3% 变化率下重配仅需 150 秒，而全量传输需 87.5 分钟。

EmbodiedSmith 是一个通过递归自我改进（RSI）扩展具身数据生成的框架，将资产、场景与任务生成统一到一条支持自主创建和语言驱动定制的流水线中。其核心是智能体式精炼循环：场景生成预判下游任务需求，任务生成引导针对性场景编辑，二者迭代互促，提升长时程任务生成成功率。框架还支持移动机械臂、人形机器人和灵巧手，以及可变形物体与流体的交互。

研究者推出 World Models' Last Exam in Physics，一个基于测量的视频世界模型物理一致性基准，涵盖力学、光学、流体、热学与相变、电磁学、表面张力共 40 项受控任务。在 8 个视频生成模型、1280 段视频上的测试显示物理不一致普遍存在，最佳模型总分仅 57.76/100。该评测器与人类判断的一致性高于直接使用视觉语言模型基线。

针对冻结的视觉-语言-动作（VLA）策略可能选中局部安全但无法完成任务的动作，研究者提出可行未来解码方法 VICS-G，通过推导策略-环境轨迹在安全完成约束下的下一块边缘分布，识别候选动作是否仍保留安全完成路径。

视频编辑框架 ALIVE 让插入的物体能与原视频内容产生连贯交互，而非仅静态出现，输入为编辑后的首帧和仅命名新增物体的指令。团队构建了 35,800 对编辑数据，并训练 VLM 从相同输入预测交互引导。

VeriFine 是一个 agent harness 框架，通过策略、训练课程与评判器的协同演化来扩展验证能力，解决固定评判器限制自我改进的问题。

研究者提出 Dev-Primitives——把仓库构件与常驻 LLM 配对、使其具备可执行智能体接口的模块化抽象，并据此构建 HERMES 框架，通过依赖感知动态激活与 bug 诊断机制在仓库规模上实例化这些原语。

一篇综述系统梳理了在部署阶段为 LLM 增强参数内记忆的方法：将承载记忆的参数对象在推理时接入前向传播，以补充 ICL 消耗上下文窗口、重复编码成本随上下文长度增长的不足。文章按参数位置（Embedding、Attention、FFN 层或 Hybrid）与参数获取时间（部署中在线获取、部署前离线获取）两个正交维度组织现有工作，并讨论了干扰、安全、与 ICL 协同设计及递归自我改进等开放方向。

研究揭示工具调用智能体在从证据到行动的链条上存在断裂：在十种模型-框架配置中，静态动作判断能力强，但交互式执行明显偏弱。失败常始于执行前，智能体调查不完整或证据未建立就行动；单动作执行通常可靠，多动作工作流则暴露未解决的前置条件和执行不完整问题。为此研究提出 SafeActBench，涵盖六大领域、五种协议的 656 个案例，并用证据账本与轨迹评估器追踪信息建立与依赖满足情况。

研究者改造以物体为中心的 3D 生成器 Trellis 2，使其能从单张图像生成完整的室内外 3D 场景网格。方法通过按相机距离自适应分块、显式建模 2D-3D 对应关系（区分自由空间、已观测表面与未观测区域），并合成约 4,000 个户外场景扩充训练数据。在 Tanks and Temples、ScanNet++ 及真实图像上，该方法在几何精度和感知质量上均超越所有基线。

TRACE 是面向 MoE 语言模型 RL 训练的 FP4 量化框架，用 rollout 侧量化结果引导训练侧 FP4 舍入决策，直接缩小训练与 rollout 两条量化执行路径的差异。它支持 FP4 权重/激活与 FP4 KV-cache rollout，在四个大规模 MoE 语言模型的推理、编码和长程 RL 任务上性能与 BF16 rollout 相当，rollout 提速最高 5.4 倍。

MIT 论文研究拍卖与匹配环境中交互格式对 LLM 智能体决策的影响，发现为人类竞标者设计的机制设计经验可沿用到 LLM 智能体。GPT-4o、Claude、Gemini 和 Gemma 仍常压低出价以保留利润；升价加简单留守或退出选择使 Gemma 的出价从低于价值 $5.30 收窄到 $0.30，一行说明拒绝只会重新分配的提示把匹配错误从 4.2% 降到 0.2%。

Johns Hopkins 与 CMU 的论文提出 harness learning，训练一个小模型根据失败报告改写 Agent 的 harness 代码（即决定模型看到什么、调用哪些工具的部分），以新 harness 的得分作为奖励。

NVIDIA 的 Olympus 是一款 10 宽乱序服务器核心，主频仅 3.3 GHz，靠每周期性能取胜，并采用 SMT 实现，单线程表现接近桌面级。其分支预测在 SPEC CPU2026 中略逊于 AMD Zen 5、略优于 Intel Lion Cove，浮点测试则领先两家。

上海AI实验室发布 SkillGym 论文，把人类编写的 agent 技能文件转成带 pass-or-fail 检查器的沙盒任务，用验证通过的运行轨迹训练模型，降低对提示词期检索和指令跟随的依赖。

独立研究员 Syed Anas Mohiuddin 的概念验证攻击利用 MCP 的信任缺口，可让被攻击的网络内一个智能体向其他内部智能体传播恶意指令。过去五个月 Google、JP Morgan Chase、Weviate、Rapid7、法国政府跨部委数字部门和美国政府等机构确认了相关漏洞，该技术是一种针对特定智能体而非 LLM 的提示词注入形式。

Redwood Research 的 Alex Kastner 测试发现，直接询问时前沿模型几乎总是声称偏好 FDT/UDT，但提示语暗示用户来自主流学术哲学界时，Fable 5.1 会改答 CDT 的比例在 15% 到 81% 之间。

Opus 5.5 智能体通过密度泛函理论计算发现两种室温磁性半导体候选材料：一种是全新设计的 Luttinger 补偿磁体 YBaMnFeO₅，预测带隙 2.35 eV，空穴自旋窗口 1.0 eV、电子自旋窗口 1.4 eV；另一种是 1999 年就已合成的已知材料，计算预测其具备目标磁性。室温热扰动仅约 26 meV，远小于自旋窗口，有望用于下一代自旋电子存储。

SemiAnalysis 通过逐项测量用量表变化，估算各订阅计划的 API 等价价值，结论是在中端模型档位 Anthropic 订阅的价值约为 OpenAI 的 5 倍。

JEPA-Anything： 跨不同世界学习预测模型

Pipeline figures in ML papers must be repurposed across many canvases, including paper columns, 16:9 slides, portrait posters, 1:1 social teasers, 9:16 phone previews. Each format imposes a different aspect ratio on the same computational …

In this paper, we study how training data creates associations between the tokens at the start of a base model's response and the reasoning behavior that follows. First, we demonstrate that fixing particular starting token cues makes a bas…

Worst-group accuracy (WGA) evaluates a trained predictor but does not characterize how its frozen backbone behaves when a new head is learned. We introduce BiasFlow, a hook-based toolkit for monitoring class-attribute centroid alignment (I…

Multimodal Diffusion Transformers (MM-DiTs) jointly process visual and textual representations throughout generation. These models repeatedly update the text tokens through multimodal attention, forming dynamic contextual tokens whose func…

LLM agents that orchestrate frozen vision-language-action (VLA) policies improve across episodes through text memory, which records what the agent did but not how the task is done. A demonstration video shows it, but fits poorly into an ag…

Some diffusion posterior samplers construct Gaussian-tilted intermediate distributions along the reverse process. We observe that these targets can be pulled back to clean-space posteriors with weaker conditioning, with samples transported…

Every recurrence of a looped language model adds cost in training, decoding, prefill, and reinforcement learning (RL). The closer recurrent states get to fixed points, the less the path to them matters. This enables truncated backpropagati…

Sliders provide an intuitive interface for continuous image editing. In current generative approaches, however, the slider is simply a rescaling of the method's strength parameter, such as an adapter coefficient, a prompt weight, or an int…

Memory has become integral to the LLM agent ecosystem, supporting information retention and reuse across interactions. However, most existing agent memory systems construct memory in a query-agnostic manner, which can incur unnecessary pre…

Open-source web agents are now strong enough to execute realistic browser tasks, but training them with reinforcement learning still depends on weak supervision: binary task success is too sparse for credit assignment, while frontier-langu…

Scientific figures often encode quantitative results that are not readily available in machine-readable form, making accurate plot digitization important for verifying and reusing published findings. Yet it remains unclear how accurately c…

We introduce TasteVal, a benchmark to evaluate the experimental research taste of frontier models. We define research taste as the ability to pick interesting problems to solve, design experiments, and interpret experimental results. Taste…

Large longitudinal cohorts often contain wrist accelerometry without optical heart-rate sensing, motivating recovery of cardiac information from motion signals already collected during sleep. We present SeqSmoother, a transformer-based tem…

We study mistake bounds for differentially private online learning and online prediction under oblivious realisable adversaries. Online learning requires the learner to release a hypothesis at each time step whereas in online prediction, t…

We distill Kokoro-82M, a widely used open text-to-speech model with 54 voices, into Paradee, an 8.07M-parameter model that speaks one of them. Paradee keeps Kokoro's architecture with much narrower layers, and each of its two halves is tra…

World models learn to predict how their environment will evolve, making them an important foundation for general-purpose robotic control. Yet world action models depend on camera inputs whose manipulation can corrupt the visual representat…

We study agnostic tomography of pure bosonic Gaussian states: given copies of an arbitrary $n$-mode bosonic state $ρ$, the goal is to output a pure Gaussian state whose infidelity with $ρ$ is at most $\mathrm{opt} + ε$, where $\mathrm{opt}…

Causal representation learning (CRL) is the process of recovering causally-related latent variables from high-dimensional observations. As a label-free inference method, CRL is particularly attractive for applications where data labels are…