AI 论文与研究

Anthropic 宣布成立生命科学研究组与实验室，Claude 智能体在仅有人类高层指导下自主发现一种与 DNA 重复序列相关、结构类似 CRISPR 的新型酶系统，命名为 array-associated reverse transcriptases（ART），并发布预印本。

推出 Perplexity Research Fellowship：一个面向任何技术或定量学科方向的早期研究人员、工程师和分析师的计划，让他们从事 AI 研究。https://www.perplexity.ai/hub/research/fellowship

The discovery of novel inorganic materials drives technological breakthroughs in critical fields such as computing and energy storage. Generative AI has promised to accelerate the materials discovery pipeline, but state-of-the-art flow and…

Scene safety understanding plays a life-or-death role in situational awareness in various critical domains. Traditional methods that rely on learning direct mappings between scenes and safety levels often lack interpretability, limiting th…

Conventional wisdom in deep learning holds that overparameterization---having more parameters $p$ than training samples $n$---is benign: larger models generalize better and, even without regularization, interpolating models generalize well…

Real-world time series evolve continuously, with meaningful changes potentially emerging at any moment. However, existing time-series language models (TSLMs) remain inherently static. They either receive complete sequences for offline proc…

Recent advances in large language models (LLMs) have made them widely used for code-related tasks. Identifier names are statistically informative in naturally occurring code, but their information is not always reliable. We investigate whe…

Medical image interpretation is high-volume and time-consuming, and while AI interpretation can reduce workload, fully autonomous deployment carries potential safety concerns and low specificity may in practice lead to increased clinician …

Transforming scholarly PDFs into machine-readable fulltext remains a bottleneck for large-scale information systems. Recent vision-based parsers improve accuracy, but need GPUs and may introduce noise into the extracted text. GROBID, a mod…

Successful mobile manipulation requires coordinated base and arm motion while maintaining accurate spatial positioning. However, demonstration-trained policies can struggle to realise the intended base motion reliably, leading to spatial m…

Fairness-aware distributed learning prioritizes clients with large losses to reduce performance disparities, but label poisoning can create large losses, thereby inducing a fairness--robustness conflict. We propose FairMean to manage this …

Long-horizon and multi-turn agents typically generate short actions and process long observations from tools and environments. This growing context demands efficient prefill, compact KV-cache storage, and accurate long-context retrieval. T…

With the rapid pace of AI research and the hundreds of daily new publications, staying up-to-date with the latest developments has become increasingly difficult. For researchers, quickly identifying impactful work is essential, yet manuall…

Reinforcement learning has become a central component of large language model (LLM) post-training, yet token-level credit lacks a generally accepted mathematical definition, leaving its relationship to commonly used training signals unclea…

Concept drift, driven by the rapid evolution of Android malware, severely degrades the performance of machine learning detectors. Current adaptation strategies are often reactive, responding only after performance has dropped and imposing …

The scarcity of non-English language data in specialized domains significantly limits the development of effective Natural Language Processing (NLP) tools. We present TransBERT, a novel framework for pre-training language models using excl…

Political discourse is widely perceived to be growing more hostile, yet robust evidence remains scarce. This study examines blame attribution in the Danish Parliament from 1997 to 2026, combining a purpose-built classifier, BlameBERT (F1: …

Residual Vector Quantization turns continuous representations into discrete, multi-level token sequences. Yet most methods operate in Euclidean space, despite the coarse-to-fine structure of the resulting codes and the latent hierarchies p…

Argument Mining (AM) transforms natural language into its underlying argument structures. This transformation is typically realized through a sequence of AM tasks that form an end-to-end AM pipeline. However, AM approaches often differ in …

Prefill and decode reward different approaches to quantization: low-precision arithmetic accelerates prompt processing, while compact weights reduce memory traffic during generation. We propose "disaggregated quantization" (DQ), which spec…

We study preference elicitation under the Bradley-Terry-Luce (BTL) model where the true partworth vector is unknown and has to be estimated as a parameter with elicited preference information. The set of selected pairwise queries is non-un…

Embodied world models predict the outcomes of robot actions to support learning and planning. For robots equipped with head and wrist cameras, this requires complementary views: the head view captures the overall task, while wrist views re…

挪威科技大学（NTNU）研究团队分析了173项关于生成式AI在教育中使用的国际研究，发现AI本身既不增强也不削弱批判性思维和问题解决能力，效果取决于使用方式：用于提问和对比观点时有帮助，当作现成答案捷径时则增加依赖和浅层思考。

AI 说网站做好了。然后首页报错，按钮重叠，还加了一个你从没要求过的登录流程。🙃 介绍 WebCraftBench：智能体实际使用运行中的应用，覆盖率引导的探索发现从未被触达的部分，然后我们评估美观度、可用性，以及原始需求是否被满足。在 197 个人工验证的配对样本上，它与人类偏好的一致率达到 85.3%。 论文：https://arxiv.org/abs/2609.15387

LG AI Research 提出 Multimodal Crystal Flow（MCFlow），将原子种类与晶体结构视为两种独立模态并分配各自的生成时间变量，用单一模型同时完成成分到结构预测、全新成分与结构生成、以及给定结构生成合适原子种类三类任务，该研究已被 ICML 2026 接收。

WorldCrafter 具有隐式3D感知记忆的一致性视频世界模型 论文：https://huggingface.co/papers/2609.24984

暂时打断一下常规的 muse 更新，但 @scale_AI 在 RSI benchmark 上取得了很棒的进展 【引用 @ScaleAILabs】：600+ 份提案 🤯 我们正在联系顶尖贡献者，开始在我们的公开仓库中构建。 我们还欢迎了 @SchmidhuberAI--RSI 的先驱--担任 RSI Bench 的高级顾问。 关于我们的设置和验证流程的新博客：https://rsi-benchmark.com/blog/verification-in-rsi-bench

一个学术实验室发起"开源周"，主张研究的基本单位已从单篇论文变为生态系统，并展示在自有硬件上运行前沿模型的成果：Qwen 3.6 35B-A3B 在 Mac 与 Metal 上以 1.5 bits per weight 量化实现 450 tokens per second。

Alexandr Wang 转发一条用户推文，该用户称未受 Meta 赞助，但盛赞 Meta 的 Muse 比又一个只提升 2% 评测分数的编程模型有价值、重要 1000 倍。他认为 AI 的瓶颈从来不是模型，而是缺少像 iPhone 那样人人可用、有趣友好的产品，并称个性化智能体比通晓一切却居高临下的工具值钱百倍。

CodeMidas 从代码本身扩展智能体编程 RL 环境 论文：https://huggingface.co/papers/2609.22068

Meta 发布 A-MLE（Agentic ML Exploration），一个已在自家广告排序模型上部署的 LLM 智能体框架，自动化提出想法、跑实验、恢复失败任务、比较结果并在模型间迁移经验。

哈佛、MIT 等机构论文提出 FINSKILLOPS，让智能体从 SEC 文件分析错误中学习，但每个新行为须先通过回归测试。该系统用小型技能注册表把重复错误（如用错报告期、混淆公司）转为可复用指令，仅当新指令修复目标失败且不破坏已答对案例时才部署，将静态 RAG 系统变为可持续学习并安全更新的系统。

小米发布 MiMo-V2.6 系列，主打前沿智能与全模态能力，并以公开方式构建。原文未披露参数规模、benchmark 分数、上下文长度及开放形式等具体细节。

GeoPair 提出一种免训练框架，通过顺序优化跨层权重配对与共享字典因子分解，在保留各层校准几何的前提下实现 Transformer 压缩，并结合结构化稀疏。该方法在多种架构、规模和模态上取得 SOTA，持续优于独立结构化权重分解与基于启发式分组的成对权重因子分解方案。

Flash-dLLM 是一个免训练的扩散大语言模型（dLLM）推理加速框架，通过 IO 感知的融合 KV 缓存内核和 KV 缓存驱动的 draft-and-verify 解码策略，让 dLLM 自身同时充当草稿模型与验证器。

研究者提出 Knowledge Pull Requests（KPRs）框架，通过提取新知识中的论断、过滤并路由到对应章节、标记与现有内容的冲突，生成区分知识变更（claim proposal）与文本变更（document diff）的 ChangeLog。

JEV-as-a-Judge 用仅做决策的评审模型做低成本初筛，在普通偏好与证据支撑的事实性任务上，与最强对比的 SOTA LLM 评审差距在 3 个百分点以内，费用仅为其 0.36%。当判断需要检查推导过程或抵御精心编写的错误答案时，差距会扩大，且 JEV 与对比模型的差距集中在低置信度决策上。一个冻结的级联流程接受高置信度判定、升级不确定判定，以更低成本保留了对比模型 99% 的准确率。

HARMONY 是一个分层链式推理框架，结合智能体推理与视觉几何基础模型，从单张室内图像重建完整 3D 场景。它先校准相机建立空间坐标系，再用 VLM 推理恢复房间布局并按墙面元素、独立家具、家具上装饰物的层级顺序放置物体，配合深度优先遍历与反思反馈回路避免误差累积，每次放置后用点云估计做几何精修。

研究提出"分离式量化"（DQ），针对 prefill 与 decode 两个阶段分别定制计算格式、权重与存储位置。

一篇立场论文提出，线性表示假设并非单一假设，而是由表示等价性区分的一族主张，需用群作用形式化表示对象、生成过程与所断言性质。该框架可厘清假设如何随度量、读取点和分析阶段变化，并据此审计常见表示量与近期的可解释性分析。