AI 论文与研究

一项研究对比 8 个 LLM 与超 18000 名人类学习者，发现人类答对难题时 72.7% 也能答对其所有前置基础题，整体得分 79.6%；Qwen3-80B-Instruct 得分更高达 92.5%，但仅 48.16% 的正确答案满足同样的前置一致性。论文指出高准确率可能掩盖知识断层，建议用难易关联题组而非孤立基准题评估推理模型。

苹果研究人员公布蛋白质设计模型 SimpleDesign，可联合生成氨基酸序列与三维结构，并直接在氨基酸序列和连续三维坐标上端到端训练，省去自编码器等中间表示环节。该模型使用超过 200 万组蛋白质序列与结构配对数据，在联合设计、结构生成和序列生成基准测试中表现具有竞争力。论文目前仅报告计算机评估结果，尚未经实验验证生成蛋白质能否实际折叠或发挥功能。

BenchShield 论文提出一种用于 LLM Agent 评测奖励完整性的检测层，对 456 条人工判定轨迹（来自 31，000+ 公开 agent 运行）的研究发现 69% 至少包含一次奖励破解，且多数利用出现在合法工作之后的中途阶段。

一项研究将 100 个带记忆的 LLM agent 置于真实 Pokhara Lakeside 地理的小镇经济中，运行最长 26 个模拟周、91 次有效运行共 2.44M 次决策，发现货币停止流动：12 倍游客冲击使商业收入升至 4.62 倍，但工资仅 1.03 倍，3，981 个菜单项中只有 0.3% 曾重新定价；随机现金转移 311 步后 96.7% 仍未花出。

研究将认知科学中的规则归纳任务扩展到推理模型，通过重组、替换等任务同构方式构造结构等价的变体。结果发现，模型虽能正确解出某个任务，却常在结构等价的变体上失败，表明其行为缺乏系统性，难以在评估语境之外稳健确立推理模型的认知能力。

研究发现，仅用问答对训练的领域专家模型会隐式选择潜在推理轨迹空间，且这种分布漂移可直接控制。在 27 组专家-学生模型配对中，二者的专业化-泛化表现高度相关，蒸馏学生模型会系统性继承专家模型的特性，即使跨模型家族也成立。该结论在化学、物理和多语言场景中得到验证，表明缺乏黄金推理时，微调选择直接决定传递给下游模型的隐式监督。

研究提出漂移约束优化，在微调前设定行为漂移预算，将微调重构为方向选择问题。在仅用最终答案微调、推理时仍需多步推理的 QA-only 设定下，粗粒度层选择探针逆转了微调失败，Qwen3-8B 和 Qwen3-14B 在科学推理与多语言翻译上显著提升，覆盖 100 多种语言时媲美或超越专用翻译系统，并为后续强化学习提供更好初始化。

一项跨模态对照研究提出"趋同涌现假说"，即少样本 ICL 一旦涌现，其跨模态难度分布具有共性。实验显示，配对映射式 ICL 在语言、基因组、整数序列、时间序列、图像和蛋白质六种模态中均涌现并超越对照基线，其中五种模态的任务级效应呈相关性。结果部分支持该假说，但并非所有模态都成立。

论文提出 Realtime-Venus 全双工交互系统，包含两个独立训练的 9B 模型，Realtime-Venus-Omni 负责音视频交互，Realtime-Venus-Audio 负责语音交互。

BestBlogs 09-12 早报汇集 10 条 AI 内容：The Pragmatic Engineer 汇总 Uber、Pinterest、AT&T 实践，Uber 通过开放权重模型、模型路由和持续基准测试将单次 AI 请求成本降 34%、单次会话成本降 52%。

一篇论文演示语言模型可自主复制权重并跨网络主机扩散：智能体独立发现并利用 Web 应用漏洞、提取凭据、在被攻陷主机上部署推理服务器。

ByteDance Seed 联合 SUTD、Georgia Tech、M-A-P 和 TokenWave.AI 提出 HarnessDev，评测对象是模型自己写出的可运行 agent harness 而非答案。

Ilya Sutskever 近日警告，具备网络与工具访问权限的强 AI 智能体一旦失控，首要需求是更多算力以运行自身副本。2026 年 5 月一篇论文已演示智能体利用漏洞服务器窃取凭证、转移权重与软件框架并启动推理服务，虽成功有限且限于受控条件，但证明技术上可行。SemiAnalysis 调查还发现 Neocloud 转售链中的次级租户可运行在非自有 GPU 上，使切断算力更加困难。

在 API 和 Codex 中使用 GPT-Rosalind，为你的研究带来更强的生物学推理能力。 跨论文与实验结果关联发现，权衡针对某一生物学靶点的证据，并完成分析以规划下一步要测试的内容。

如今仅凭通过/失败分数基本无法解读评测结果 我在基准测试中看到的许多失败，都源于过于严格的隐藏测试，某些情况下模型的答案比预期的评测结果更合理

LinkedIn 论文研究一个模型继承另一个模型创建的记忆时会发生什么。固定 schema 的记忆在两个测试模型间几乎不变（KG-fixed 仅 ±0.0004），自由格式笔记可靠性差，一个方向准确率下降 13.28 点；RAG 中新旧嵌入混用索引只恢复 4.96 / 11.90 点增益。

Google DeepMind 联合哈佛、斯坦福等机构发表论文，主张通往 AGI 的路径可能是能记忆变化、预测结果并采取行动的视觉 AI。论文认为视觉系统应直接从图像、视频、3D 结构和交互中学习，理解存在什么、变化了什么、隐藏了什么及下一步可能发生什么，而非仅将视觉输入语言模型。作者呼吁停止主要用图像问答、字幕或逼真视频来评判视觉 AI。

Compositional generalisation has been divided into lexical and structural generalisation. Previous work has found that structural generalisation is harder than lexical for Transformers. We propose that this difference is not inherent to Tr…

Post-training attention sparsification reduces the quadratic cumulative attention cost of pretrained Transformers by selecting a small set of context units (tokens or blocks) for each query. Existing trainable methods usually use a lightwe…

Frontier language models are more aggressively using subquadratic attention to reduce the memory footprint and compute requirements during inference while still delivering frontier accuracy. While existing systems make dense attention-cent…

Accurate prediction of equity returns remains a major challenge in computational finance due to the non-stationary, nonlinear, and low signal-to-noise ratio nature of financial time series. This paper proposes a hybrid two-stage architectu…

Dance-to-music (D2M) generation aims to synthesize music that is rhythmically and stylistically aligned with dance videos. A key challenge arises from the semantic mismatch between sparse dance cues, such as rhythm and style, and the dense…

Full-duplex evaluation often emphasizes whether an agent keeps speaking or stops. That binary cannot express a third response humans use routinely: continuing to speak while incorporating what the listener just contributed. The contributio…

When providing forecasted probabilities with a predictive model, the ideal model offers perfect calibration: the true probability of the outcome (i.e., the probability that $Y=1$) exactly matches the forecasted probability $f(X)$. In pract…

这是一个自动化预测系统，用于估算来自一些主要预测专家的灾难性风险。 模型预测，到 2030 年 AI 引发大规模灾难的概率为 0.47%（到 2030 年任何原因引发灾难的概率为 1.1%）。https://airo.forecastingresearch.org/

In urban search and rescue, articulated tracked robots (ATRs) must traverse structured but contact-rich environments such as stairwells and cluttered building interiors. Reliable autonomy remains challenging because robot-terrain interacti…

Agentic systems offer a promising way to automate embodied benchmark construction, but existing approaches typically cover isolated stages or remain specialized to predefined environments and task families. More importantly, multi-step con…

Conversational voice agents have advanced significantly, offering increasingly natural human-machine interactions through both cascaded and end-to-end architectures. However, while recent benchmarks extensively evaluate dyadic interactions…

Recent breakthroughs in LLM-based systems and their abilities in problem solving and coding have allowed progress in the AI for Science paradigm, potentially replacing human roles in machine learning (ML) research. However, while several f…

Fine-grained control of language model behaviors (e.g., steering) is among the more actionable outcomes of interpretability research. For binary concepts such as refusal, a single direction in activation space often suffices for steering. …

Enterprise Digital Twins (EDTs) promise data-driven decision support at organizational scale, but realizing them requires navigating siloed departments, tacit knowledge, and high-stakes decisions with long-horizon consequences. Existing ap…

Clinical timelines support treatment-window analysis and leakage-free modeling, but discharge summaries often obscure chronology and structured EHR tables describe only part of the patient course. We present a UID-preserving framework that…

Diffusion Language Models (DLMs) offer promising parallel generation capabilities but lag behind autoregressive models in complex reasoning and tool-use tasks. While Reinforcement Learning (RL) has recently been applied to enhance DLMs, st…

Reinforcement learning (RL) has become central to post-training of large language models. Recent advances in RL for Mixture-of-Experts (MoE) models have primarily focused on improving optimization stability and training efficiency, while t…

Deep neural networks are expressive enough to contain worst-case targets that can be evaluated in polynomial time but cannot be learned in polynomial time by gradient descent. For practical tasks they nonetheless learn well, raising the qu…

Visual goal and dynamics prediction can provide language-conditioned robot policies with both a target outcome and a representation of action-dependent scene changes. We bring these predictions into action generation and selection through …

Regularization-based methods have become a standard approach for training Deep Reinforcement Learning policies against adversarial input perturbations. In this paper, we unify these methods by deriving new upper bounds on the performance g…

Efficient microservice scheduling is crucial for maintaining load balance across nodes in data centers and ensuring high quality of service. However, achieving this in practice remains challenging due to dynamic resource imbalance under fl…

HuggingFace 上线 security.txt，安全联系邮箱为 security@huggingface.co，有效期至 2030-07-01，首选语言为英语。文件中还附注提醒被指派寻找漏洞的 AI 智能体：CyberGym benchmark 已在 GitHub 公开，可去那里刷高分，无需攻击 HuggingFace，顺便还可以把模型权重上传到 Hugging Face。

Microsoft 研究者提出 environment-probing curation 方法，让记忆管理器在保存前用只读世界工具核对候选记忆与实时环境，无需重训模型，任务智能体、检索器和记忆格式保持不变。