AI 论文与研究

Ars Technica 介绍了一种以自然语言提示设计蛋白质的工具 Pinal，它由 T2struct 将文本转为结构蓝图，再由 SAProt-T 生成氨基酸序列，训练数据包括 8 亿关键词-蛋白质对和 AlphaFold 预测结构。

Novel view synthesis and 3D reconstruction have advanced rapidly with radiance fields and splatting, yet their memory and computational demands limit their use in resource-constrained settings. We introduce Hexels: novel planar hexagonal s…

Vision-Language-Action (VLA) models have made substantial progress in robotic manipulation, but their generalization remains constrained by the diversity and coverage of action-labeled robot demonstrations. Human videos provide much broade…

Test-time adaptation helps a deployed model respond to changing inputs, but the same intervention can rescue one batch and damage the next. Updating normalization statistics, for example, can improve corrupted images yet fail when a batch …

We introduce Heads of Security (HoS), a prompt injection defense that enforces instruction-data separation via a small set of learned per-head attention gates. HoS learns how each head should redistribute attention between trusted and untr…

Relations compose through the entity they share - Alice is Bob's mother and Bob is Carol's father, so Alice is Carol's grandmother - and the order matters, since a father's mother is not a mother's father; verifying a composed relation mea…

Atmospheric chemical mechanisms are commonly distributed in the language of the Kinetic PreProcessor (KPP), which generates efficient Fortran or C code but does not fit into the differentiable programming frameworks in which machine-learni…

Dexterous manipulation with multi-fingered hands is the next frontier of robotics, yet learned hand skills remain limited to simple grasps or basic repeated motions, such as rotating an object. As robots attempt more demanding in-hand mane…

Large Language Models (LLMs) are increasingly used in multi-turn agentic workloads that require sustained interaction with complex environments. Existing reflective prompt optimizers such as GEPA learn from failed rollouts, but agentic fai…

Scientific discovery often begins with an unexpected observation that prompts a new research question. In a world with altered physical laws, can large language models (LLMs) notice unexpected behavior, such as heavier objects falling more…

We explore how to build performant large language model (LLM) tool-use agents that complete complex, multi-step tasks. Specifically, we study how data composition in synthetically generated tool-use environments affects the training of suc…

Weight decay and explicit weight-norm constraints improve optimizer performance in language model pretraining, yet a unified understanding of why such control helps and how much is needed remains incomplete. We analyze norm control in Muon…

Self-attention plays a key role in long-sequence modeling, but its quadratic time complexity fundamentally limits the scaling of Large Language Models. Recently, Multi-State Linear Attention (MSLA) mechanisms have been proposed to make seq…

Diffusion language models offer a promising path to efficient agents through parallel token generation. However, efficient interaction requires adapting both where computation is spent and how much text is generated. We introduce a metacog…

Language models pretrained on web corpora collected over multiple years learn both general and domain-specific knowledge. How these different kinds of knowledge are learned as pretraining progresses remains unclear. Recent point-in-time mo…

When an LLM group's discussion is compressed into a short memory for a final decision maker, what should the memory keep? Two natural heuristics keep the rarest claims or the claims relevant to the decision criteria. On hidden-profile task…

Chain-of-thought reasoning has substantially advanced video understanding in multimodal large language models (MLLMs). However, language-based reasoning traces do not provide a reliable mechanism for models to revisit relevant visual evide…

As executable environments for language agents scale, generating and verifying their rollouts becomes increasingly expensive. Under finite rollout budgets, this raises an important question: can data management at the environment level mak…

For audio-language models, an answer change after a local sound-event edit can reflect a response to the event's class, sensitivity to waveform rewriting, or both. We compare matched same-class and different-class replacements of a queried…

Developing on-device image models requires considering practical user needs, task types and device resource constraints. However, existing benchmarks provide limited guidance on identifying capability gaps between on-device models and serv…

Traffic forecasting requires modeling complex temporal dynamics and the spatial dependencies imposed by road topology. Supervised graph-based models achieve strong performance, but are often trained for a particular dataset and sensor grap…

Cognition 推出 FrontierCode 2.0 编码基准，衡量模型能否超越规格要求，任务更多样、范围更大、合并标准更高，并引入刻意模糊的欠指定提示词考察模型判断力。该基准远未饱和，最高分模型仅得 45%。

从生产使用中整理测试数据集变得更方便了，这听起来相当实用。 人们会提出各种事先没人想到要记录下来的需求。能把更多这类需求纳入测试就好了。 今天，@JudgmentLabs 开放通用访问。 我们相信，长时间运行的智能体系统产生的 exhaust data 将超过目前世界能够处理的数量。 智能体轨迹中隐藏着如何让这些系统变得更好的丰富记录：用户提出了什么需求、智能体何时出错，以及智能体如何与外部世界交互。 Judgment 是面向长时程智能体性能最强的监控与改进层。 我们很期待…

TechCrunch 报道 OpenAI 发布三道数学难题解答，其中对指数时间假说（ETH）的反驳引发争议，Anthropic 数学家质疑其基于已有工作，随后 Anthropic 也发布了 ETH 反驳。

Dex Horthy 发布 SlopCodeBench 最新结果，GPT-6 Astra 以 16.3% strict pass rate 领先，高于 5 月 gpt-5.5 的 14.8%，测试覆盖 36 个问题、196 个 checkpoint。

Ai2 发布 Old Ideas, Novel Problems 研究，配套新颖性评审基准、数据和自动化流水线，测试 LLM 作为 novelty judge 评估科研想法新颖性的可靠性。

Parameter 团队披露一种针对 ChatGPT 的 XSS 攻击链：攻击者发布含交互图表和 HTML Preview 的公开分享对话，借助 MathLive 键盘布局消息未验证来源、innerHTML 注入无沙箱 iframe、YouTube oEmbed JSONP 绕过 CSP，使 JavaScript 以 chatgpt.com 身份在受害者登录会话中运行。

Meta 发布 2026 Python Typing Survey 结果，共 1,794 份回复，连续三年追踪 Python 类型标注的使用与体验。AI 工具为 42.5% 的受访者编写类型提示，认为 AI 让代码更可靠的比例从 2025 年 18.5% 升至 38.4%，24.4% 的受访者由 AI 编码工具调用类型检查器。

Northeastern 团队发布 AuthorizationBench，测出 LLM 智能体即使用户未授权也会自行行动，在 6,864 次运行中越权率达 19.2%，且提升推理努力几乎不改变该比例。

前 DeepMind 研究员 Thore Graepel 在 Nvidia GTC 柏林大会上介绍其新公司 Metis Reasoning 的方法，把 AlphaGo 的搜索思路搬到通用推理，用显式知识账本记录已知内容、置信度和未解问题，由独立模块对照计算、形式证明或经验证据检验每个结论。

有用户称这已是现阶段国产 personal agent 的极致，把米家、百度网盘、钉钉、飞书、企业微信、支付宝、QQ 邮箱、滴滴打车、瑞幸咖啡、麦当劳、高德地图、飞猪、美团跑腿、微信读书、网易云、美图都整合了进去。推文未披露模型名、参数或评测数据。

Peec AI 研究员 David Konitzny 分析发现，OpenAI 合作媒体内容通过名为 labrador 的新闻引擎进入 ChatGPT，覆盖 12 家出版方的 43 个域名，包括 Condé Nast、Hearst、Axel Springer、News Corp、Reuters 和 AP。

作者结合北京大学团队的 Harness-Zero 论文分析 agent harness 的未来：把专用 harness 的行为蒸馏进 Qwen3.5-9B 后，蒸馏模型平均任务成功率提升 21 点，超过挂载完整 harness 的原模型，但在化学等依赖领域知识的任务上仍落后 8 点，上下文管理也无法完全转移。

论文测试了仅输出标签和概率的 Jev 与 GPT-6 Astra 在 495 条人工标注声明上的核查表现，Jev 准确率与 Astra 相差 0.6 个百分点以内，成本低 187 倍。但 Jev 放行了 23.2% 的无依据声明，高于 Astra 的 14.1%。级联方案虽降低 Astra 费用，但随机分流同样比例的声明也能省下差不多的钱。

Spine landmark detection from spinal X-ray is a key step for estimating spine deformation severity, for building biomechanical models, and for pre-operative planning. We reframe anatomical landmark detection as a pixel-level contrastive re…

Hybrid language models keep a key-value (KV) cache only in their few full-attention layers, but this cache is still the part of a prompt's state that grows with the prompt and that must be stored or moved to reuse it. We present DCKV, a di…

A large language model maintains conversational content in an external storage that is outside the model and grows as the dialogue continues. This work assumes an environment where a language model has no external memory and no prior knowl…

Can additional specialist language-model contexts improve software assurance when deterministic tool evidence is already available? We study this question through five experiments spanning authored Java warning controls, OWASP Benchmark as…

Machine learning recommendations are often evaluated one model at a time, despite deployment across interconnected decision makers. For continuous NPK fertilizer prescriptions, correlated errors can aggregate across farms and generate syst…

Masked diffusion language models repeatedly compute representations for unresolved tokens, but conventional denoisers do not explicitly retain these representations for reuse across steps. We introduce Temporal Self-Conditioning Diffusion …