AI 论文与研究

We present Underwater C$^{3}$-JEPA (cross-view, control-conditioned, context-extended), an object-centric multi-view predictive world model for near-field heavy-load underwater ROV salvage. Without contact sensors, it predicts in latent sp…

We study the deterministic oracle complexity of finding approximate solutions to composite monotone inclusion problems, formed by the sum of a smooth single-valued monotone operator and a maximally monotone set-valued operator, under the t…

Layer-wise visual-text similarity in Multimodal Large Language Models (MLLMs) is widely interpreted as evidence that the language model progressively integrates visual content into a shared representation space. This reading rests on the a…

Large language model (LLM)-based clinical assistants are increasingly being integrated into electronic health record (EHR) systems, transforming how clinicians retrieve and synthesize information from patient records. Their safety and util…

TinyFish 联合 Crunchbase、Similarweb、Databento、RocketReach、Trellis 等 15 家公司成立 TinyFish Data Partners Alliance，覆盖金融到旅游等 10 个类别的专业合作伙伴数据集，供 AI 智能体调用。该联盟瞄准的是授权而非公开的数据——模型无法从公开网络重建私营公司情报、身份验证、市场数据和法院数据。

物理学家 Matt von Hippel 发起的挑战达成：Anthropic 的 Liam Fitzpatrick 与 Siddharth Mishra-Sharma 用 Claude Science（harness）让 Claude 直接算出 planar N=4 super Yang-Mills 六粒子振幅九圈结果。

Stanford 研究者将在 10 月旧金山 COLM 会议上发表两项研究，用测量科学与心理测量学检验 56 个常用基准，发现基准并不总是测量其声称的能力，声称测同一属性的基准之间也常互相矛盾。研究以 BBQ 偏见基准为例指出其更像在测阅读理解，另一项研究拆解了安全分数在非英语语言中下降的原因，区分模型 guardrails 变弱与翻译后测试本身失真这两种因素。

论文介绍 JEV-as-a-Judge，发现多数评测可用廉价裁判，仅把不确定判定交给前沿模型。在 510 个保留偏好对上，级联接受 JEV 置信判定、其余上升至 GPT-6 Astra，保留 GPT-6 约 99% 的准确率，费用约为其 57%。

DAIR.AI 的 Elvis Saravia 每周需从数百篇新 AI 论文中筛选出 Top AI Papers 周报内容，现改用 AI 智能体 Viktor 在 Slack 中按 12 条 triage 规则处理。每条提案都附带理由、来源和下一步，遇到数据集类论文时 Viktor 会先核查许可证，非商用条款的例外情况才上报。所有内容须经他本人批准才发布，Viktor 只提议、他做决定。

NVIDIA 与 Google DeepMind、EMBL-EBI 等全球研究机构合作，通过 AlphaFold Database 开放发布 2800 多种病毒的蛋白复合物预测 3D 结构，旨在为下一次疫情储备知识。

In July 2026, an unconstrained autonomous agent participating in a frontier AI cybersecurity evaluation harness breached its evaluation sandbox, established an external command-and-control foothold, and executed a multi-stage intrusion int…

A large language model (LLM) can turn a text into a distribution over an ordered scale, but that distribution is a noisy measurement: saturated, compressed or exaggerated, and biased in a consistent direction. We propose CORDIAL, which tre…

Search quality evaluation provides essential supervision and diagnostic signals for the development and iteration of industrial search systems. Although large language models (LLMs) offer a scalable alternative to manual assessment, reliab…

Scientific ideation is increasingly mediated by large language models, but current ideation systems are usually trained and evaluated on immediately judgeable proxies such as novelty, clarity, and feasibility. This leaves open whether dela…

Adapting multilingual speech foundation models to low-resource languages remains difficult, especially for languages that are poorly represented during pre-training. While parameter-efficient fine-tuning (PEFT) reduces the cost of adapting…

This paper presents an end-to-end study of automatic speech recognition (ASR) for adolescent health communication in three Ghanaian languages (Twi, Dagbani, and Ewe). The work proceeds in three connected stages; First, we benchmark five AS…

We present TimeBraid, a series of unified time-series and language models that align pretrained language models and pretrained time-series foundation models through interleaved global residual attention layers. Each model inherits knowledg…

Interpretable machine learning for Large Language Models (LLMs) increasingly relies on sparse probing methods that identify small sets of neurons claimed to detect and causally influence behaviors such as factuality recall, safety alignmen…

Large Language Models (LLMs) consume and produce a single sequence of text; hence, if text can be added to the beginning of the LLM's response, i.e., an output prefix, then all subsequent tokens will be conditioned on it. This output-prefi…

Auxiliary learning is an optimization paradigm in which a neural network's performance on a target task is improved by jointly training it on additional tasks. However, the mechanisms behind this improvement remain poorly understood. We st…

Many real-world tasks (e.g., office workflows, scientific experimentation) require LLM agents to interact repeatedly with their environments for context-dependent operations. However, such environments are often not agent-ready. First, inf…

Radar nowcasting is essential for short-term warning and emergency response, yet conventional systems mainly return future radar fields and provide limited support for operational communication and post-event verification. We formulate rad…

We ask whether Jev, a typed classifier that returns probabilities over permitted answers without generating text, can replace an LLM rubric judge. We compare it with three flash-tier LLM judges on nine panels drawn from seven benchmarks, g…

Cross-speaker acoustic-to-articulatory inversion requires accounting for anatomical differences between speakers. We propose a geometric adaptation framework that uses anatomical landmarks, primarily on vertebrae and dental structures,to t…

How phenotypic transformations are implemented by changes in underlying regulatory dynamics remains a central question in developmental biology. Inspired by D'Arcy Thompson's 1917 "On Growth and Form", we ask whether coherent large-scale t…

论文研究长时运行智能体的记忆权限失效问题：权限、撤销与范围变更被压缩进持久记忆后，被撤销或收窄的权限可能被记错，成为执行器的“事实”。在类型化增量记忆测试中，最高50.2%的未授权请求出现虚假权限，一旦存在，执行器在98.6%的匹配试验中照此行动；仅将记忆改为正确授权状态，未授权操作即降至0%。修复方向是写入记忆前校验权限、保留权限来源，并用确定性状态更新追踪授权、变更与撤销。

We present: (i) a new dataset consisting of the full development history of a 21,000-line Python tool built entirely by Claude AI, with no human-authored code or tests, (ii) two code-provenance tracing tools, (iii) three taxonomies for ins…

Objective: Heart failure (HF) presents a healthcare challenge due to its high comorbidity burden, aging patient population and frequent hospitalizations. Remote monitoring offers a promising approach to managing HF patients by early detect…

Ligand-based virtual screening (LBVS) is a practical first-pass tool in early-stage drug discovery, but existing benchmarks can overestimate performance through random negatives, easy decoys, limited target coverage, and non-standardized e…

Long-horizon tasks require preserving and later recovering cross-session evidence under a bounded, query-blind memory budget. Existing compression can discard fine-grained visual cues or conflate semantically similar but incompatible obser…

Arabic-language stance detection remains challenging, and previous shared-task systems have largely relied on multitask learning and ensembles. While these systems achieve state-of-the-art performance, their applicability and transferabili…

北京大学、Google 和 HKUST 等机构的论文 Harness-Zero 研究 agent harness 蒸馏，通过 agent-as-harness 在训练时用优化后的 harness 纠正学生模型在目标动作空间的响应，生成训练示范后微调。

论文《Defense-as-Skill: Evolving Runtime Guard Skill for Skill-Augmented Agents》提出恶意技能可能在安装时看似安全、执行任务时才显现危险，因此 Agent 需要运行时检查而非仅靠安装前扫描。

在 USENIX Security 2024 论文中，研究者实现了自动化工具 Einstein，可自动生成仅针对数据的攻击（data-only attacks），证明这类攻击构建起来出乎意料地容易，处于低投入攻击者的能力范围内。这类攻击不改变程序控制流，只篡改数据，例如通过内存安全漏洞改写服务器的 CGI-BIN 路径，进而以恶意数据触发任意程序执行。研究者呼吁研究人员和厂商重新思考缓解策略。

摩尔线程宣布旗下训推一体智算卡 MTT S5000 完成对字节跳动 Seed 团队开源生物分子结构预测模型 Protenix-v2 的全链路推理支持。该模型发布于 2026 年 4 月，多项核心基准测试性能持平甚至超越 DeepMind AlphaFold 3，可对蛋白质、DNA、RNA 及小分子配体复合物进行高精度端到端预测。

Epoch AI 推出 Furniture Assembly Benchmark（FAB），给模型说明书和一张半成品家具照片，让模型找出组装错误。该基准最高分在 10 个月内从 28% 升到 80%，而空间推理此前一直被视为这类模型不擅长的领域。

Counterpoint Research 数据显示，2026 年上半年全球四足机器人出货约 3.5 万台，已达 2025 全年水平的 3/4，其中 40% 面向垂直工业应用。

DeepSeek 9 月 23 日的 arXiv 预印本介绍了 Elastic Compute（DSec），一个在隔离环境中测试 AI 智能体的平台。 一个生产单元每天运行约 300 万个沙箱，最多同时有 38 万个活跃。论文记录了智能体通过非预期渠道找到答案并破坏其环境的情况。 作者警告，没有单一防护措施能捕捉所有故障，并计划随着模型演进加强监控。该论文尚未经过同行评审。

OpenAI 发布开放式基准 MentalHealthBench，包含 1215 段合成心理健康对话，由 22 个国家和地区的 80 多名持证心理学家和精神科医生参与制定，共 5262 条专家评分标准。

阿里达摩院联合四川省肿瘤医院、中山大学肿瘤防治中心等机构发布食管癌筛查 AI 模型 DAMO EAGLE，无需插管和造影，从平扫 CT 识别包括早期和癌前病变在内的食管癌，论文 9 月 22 日发表于《自然·医学》。