13个顶级AI模型参加2026年数学全国一卷测试，GPT 5.5 Thinking、Deepseek-v4 Pro、Gemini 3.1 Pro Thinking和Claude Opus 4.8 Max并列第一，Kimi 2.6 Thinking以微小分差位居第二梯队。测试采用LaTeX格式统一输入，禁用联网。多数模型在长解答题和多选题上容易失分，如第11题（多选题，正确答案ABD）所有模型均未完全答对。Claude Opus 4.8 Max因额度不足未完成所有题目，Claude Sonnet 4.6 Thinking尝试直接搜答案，Deepseek-v4 Pro曾拒绝完成。模型在上下文记忆、题目理解和符号识别方面仍有明显短板。
Rohan Paul@rohanpaul_aiAI 评分6161
SK hynix and NVIDIA just formed a multi-year memory partnership to build the chips behind the next wave of AI factories. So SK hynix is being pulled deeper into AI servers, personal AI, and physical AI. SK hynix will codevelop memory for NVIDIA Vera Rubin AI supercomputers, Vera CPUs, RTX Spark-powered PCs and Jetson Thor platforms, and will use NVIDIA CUDA-X, PhysicsNeMo, Omniverse, OpenUSD and cuOpt to accelerate chip design, semiconductor simulations and autonomous fab digital twins. Memory can no longer be designed after the GPU is finished, because advanced DRAM and high-bandwidth memory need years of co-design, fabrication planning, and capital spending. The partnership also moves AI into chipmaking itself, using CUDA-X, PhysicsNeMo, and simulation tools to speed semiconductor physics, lithography, and internal engineering code. SK hynix also plans fab digital twins, which are software copies of factories where robots, tools, wafers, and logistics can be tested before changes hit the real production line. It reinforces SK Hynix's position as a key supplier in the AI hardware supply chai
本教程展示如何利用GEPA这一反思性提示进化框架，改善小型语言模型在多步算术应用题上的表现。从弱种子提示出发，构建确定性基准并定义能返回可操作反馈的结构化评估器。多组件设置同时进化指令字段与输出格式规则，最后在保留验证集上对比基线提示与优化提示，检验优化效果是否泛化。
富士康在2026年台北电脑展期间展示了一款液冷版 RTX 6000 Blackwell GPU，面向服务器和数据中心场景。该卡采用单槽设计，全面覆盖液冷散热模块，拥有96GB GDDR7显存，可运行AI大模型、复杂仿真及专业图形任务。电源接口改用远离PCIe插槽的独立插槽，未使用标准12V-2x6接口。
Rohan Paul@rohanpaul_aiAI 评分4545"很快，竞赛数学、竞赛编程将不再有趣。 如果到明年我们还没有一个任何人都能用来在国际数学奥林匹克（IMO）中取得满分的模型，我会很失望。"
Rohan Paul@rohanpaul_aiAI 评分6262We've made a breakthrough in self-evolving AI scientists moving from "search" to "principled discovery": Scientific discovery requires that the search space itself changes, and an AI scientist must perceive this shift without intervention. We built an AI that achieves this for the first time with the ability to discover the scientific vocabulary it reasons in. Evidence, tools, artifacts, verifiers, failures & claims become typed provenance. We show three distinct modalities: 1) retrieval, adding known objects; 2) search, exploring a fixed schema; and critically: 3) discovery, a verified regime transition. We solve the open-endedness evaluation problem by lifting agentic workflows into a typed copresheaf and proving, via a Kan obstruction, that true discovery is not unbounded generation but a verifiable schema expansion: old evidence is transported by Left Kan extension, and genuine novelty is mathematically quantified by the pointwise residual beyond the transported image - separating discovery from mere search and making novelty objective and measurable rather than a subjective judgment or benchmark delta. Our AI scientist is built in a way that does not pre-conceive the approach it chooses; instead, we endow the system with formal power to adapt, evolve, and reason from first principles. Case studies include: 1⃣Builder/Breaker model that discovers mode-conditioned compliance in proteins; 2⃣CategoryScienceClaw that finds anisotropic fiber-network stiffness rules. Great work in collaboration with my graduate student @fwang108_ @MITdeptofBE F.Y. Wang & M.J. Buehler, Self-Revising Discovery Systems for Science: A Categorical Framework for Agentic Artificial Intelligence, arXiv:2606.01444, 2026
Rohan Paul@rohanpaul_aiAI 评分6666We've made a breakthrough in self-evolving AI scientists moving from "search" to "principled discovery": Scientific discovery requires that the search space itself changes, and an AI scientist must perceive this shift without intervention. We built an AI that achieves this for the first time with the ability to discover the scientific vocabulary it reasons in. Evidence, tools, artifacts, verifiers, failures & claims become typed provenance. We show three distinct modalities: 1) retrieval, adding known objects; 2) search, exploring a fixed schema; and critically: 3) discovery, a verified regime transition. We solve the open-endedness evaluation problem by lifting agentic workflows into a typed copresheaf and proving, via a Kan obstruction, that true discovery is not unbounded generation but a verifiable schema expansion: old evidence is transported by Left Kan extension, and genuine novelty is mathematically quantified by the pointwise residual beyond the transported image - separating discovery from mere search and making novelty objective and measurable rather than a subjective judgment or benchmark delta. Our AI scientist is built in a way that does not pre-conceive the approach it chooses; instead, we endow the system with formal power to adapt, evolve, and reason from first principles. Case studies include: 1⃣Builder/Breaker model that discovers mode-conditioned compliance in proteins; 2⃣CategoryScienceClaw that finds anisotropic fiber-network stiffness rules. Great work in collaboration with my graduate student @fwang108_ @MITdeptofBE F.Y. Wang & M.J. Buehler, Self-Revising Discovery Systems for Science: A Categorical Framework for Agentic Artificial Intelligence, arXiv:2606.01444, 2026
北卡罗来纳大学教堂山分校和美国东北大学的一项新研究发现，主流AI模型在分析职业体育比赛时表现很差。研究构建了SVI-bench基准，包含35000小时比赛画面等数据。ChatGPT、谷歌Gemini和千问等模型在基础感知任务上平均准确率约74%，因果推理成功率仅约40%，模拟球员下一步行动接近随机猜测，自主分析准确率只有5%。研究认为AI仅擅长描述画面，无法解释战术为何成功或预判后续发展。
OpenCV 5 正式发布，采用基于图的 DNN 引擎，ONNX 算子覆盖率从 4.x 的不到 23% 提升至超 80%，原生支持 Transformer、视觉语言模型（VLM）和大语言模型（LLM）。其他更新包括：更好的 Python 集成与命名参数、更紧凑核心代码、清晰硬件加速层、原生 FP16/BF16、规范化 0D/1D 张量、扩展 3D 视觉及现代化文档。该库 GitHub 拥有超 86,000 stars，每日安装量超一百万次。
推荐理由：OpenCV 5 是一次架构级大更新，DNN 引擎重写、ONNX 覆盖率从 23% 跳到 80%，原生支持大模型推理，做模型部署的该关注了。
美国企业 AI 累计投入突破 1 万亿美元，模型推理成本飙升，降本增效未达预期。Ramp 报告显示 DeepSeek 首次登上其软件趋势榜榜首。案例包括某企业一个月为 Claude 支付 5 亿美元，Uber 四个月内耗尽全年 Token 预算。DeepSeek 宣布 API 价格永久下调 75%，MiniMax 压至新低。Ramp 首席经济学家称这是美国企业寻找 OpenAI、Anthropic 低成本替代方案的最明确信号。部分企业已直接向 DeepSeek 付费。2025 年初 DeepSeek R1 模型曾登顶 App Store，但企业采用率一度达 0.3%，后回落至 0.1% 并维持到 2026 年 4 月。成本压力是再度受关注的主因。
软银CEO孙正义在CNBC采访中透露，OpenAI正使用AI参与设计其后续模型，认为AI正逼近ASI（人工超级智能），并将到来时间从10年缩短至2年。他定义ASI为比人类聪明10000倍的AI。OpenAI曾在2月称GPT-5.3-Codex是其首个“参与创造自身”的模型。孙正义每日使用ChatGPT 2-3小时，预计未来数年内AI会在70%-80%科目上超过人类。
Rohan Paul@rohanpaul_aiAI 评分7979New Anthropic Science Blog: Making Claude a chemist. To manipulate a molecule, chemists first need to understand its structure. Their main tool is NMR spectroscopy. We found Opus 4.7 matches—and on some tasks beats—dedicated NMR software. Read more: https://www.anthropic.com/research/making-claude-a-chemist
腾讯混元提出 Stem 稀疏注意力算法，已被机器学习顶会 ICML-26 收录。该算法从因果信息流重新审视块级稀疏，引入 Token 位置衰减（TPD）和输出感知度量（OAM），仅用 25% 算力即逼近稠密注意力精度。配套的 HPC 开源 Stem+BSA 算子将稀疏收益转化为实际加速，在 128K 上下文下首字延迟降低 3.6 倍。
Rohan Paul@rohanpaul_aiAI 评分9393Anthropic just disclosed that Claude now writes more than 80% of the production code it merges. Before Claude Code reached research preview in 02-25, Claude wrote only low-single-digit merged code, while output per engineer has since risen to 8x the 2024 baseline. The shift comes from agents that edit files, run tests, inspect failures, spawn helper agents, and keep working across longer tasks instead of only suggesting snippets. Anthropic says reliable task length is doubling about every 4 months, with Mythos Preview reaching at least 16 hours and open-ended Claude Code success hitting 76%. i.e. Claude Mythos Preview could stay useful on a task that would take a skilled human roughly 16 hours of work Claude also moved from a 3x training-code speedup to 52x, while a skilled human reached about 4x in 4 to 8 hours on the same setup. The remaining human edge is research judgment: choosing the right problem, trusting the right result, and knowing when an experiment is dead.
Perplexity AI 发布面向个人电脑的混合本地-服务器推理编排器，可自动将 AI 任务在设备端模型与云端模型之间动态路由，实现推理负载的智能分配与优化。
NVIDIA Dynamo Snapshot 使用 CRIU 和 cuda-checkpoint 工具，在 Kubernetes 上对 vLLM 推理工作节点进行检查点和恢复，实现快速启动。
Stem算法通过Token位置衰减（TPD）和输出感知度量（OAM）两项创新，仅用25%算力即逼近稠密注意力精度。配套HPC算子针对Hopper架构优化，支持FP8量化与vLLM的Paged KV Cache，在混元Hy3 preview上实现首字延迟降低3.6倍。HPC-BSA相比MIT-BSA稳定保持约3倍加速，在8K至256K序列长度上表现一致。
推荐理由：把稀疏注意力从一刀切升级成按信息流分配预算，算子在Hopper上把理论加速几乎无损转化，3.6倍首字延迟下降不是灌水——做长上下文应用的值得照着开源代码试。
百度伐谋进入北京工业大学苗扬团队实验室，辅助解决PEM电解槽制氢故障诊断、空间站空气监测和液体表面波可视化等难题。在制氢安全中，模型准确率从92.26%提升至95.04%，探索周期从“周级”缩短至“小时级”；在空间站色谱柱优化中，72小时内生成方案，体积缩小40%、分离效率提升3倍、误差降低8.17%；海外运维场景预计节省70%–80%维修等待时间。伐谋不替科研人员“写答案”，而是围绕目标、变量和评价标准进行大规模搜索试错。
Pinterest 宣布大幅扩展与 AWS 的合作，计划在 2031 年前投入 40 亿美元，扩大对 AWS 定制芯片（包括 CPU 和 AI ASIC）的使用。目前约 1/3 的计算基础设施已运行在 AWS Graviton 系列 Arm CPU 上，未来还将利用 AWS Trainium 托管和运行 LLM / VLM 模型，支持个性化视觉搜索和 AI 辅助发现。首席技术官 Matt Madrigal 表示，深化合作将加速 AI 创新，提升用户体验和广告主效果。
SiliconFlow@SiliconFlowAIAI 评分6464DeepSeek has now topped our token share rankings 4 weeks in a row: https://openrouter.ai/rankings
Tencent Hy@TencentHunyuan精选AI 评分7474推荐理由：腾讯混元联合人大开源的 PlanningBench，补上了 LLM 从「会说」到「会做」之间规划能力评估的缺口，做 Agent 的同学可以直接用来评测和训练，开源即用。
美团技术团队在ACL'26上分享6篇论文，涵盖代码评测、复杂流程推理、数学竞赛、过度思考分析、推理后训练优化及生成式推荐。CoreCodeBench从12个开源库生成1524个结构化任务，有效性达78.55%；SOP-Maze基于真实业务构建397个流程实例，测试显示前沿模型存在流程遵循、对话脆弱性和计算错误三类短板；AMO-Bench含50道高难度数学题，最强模型准确率仅52.4%；The Evolution of Thought提出推理完成点（RCP）检测器减少冗余生成；MASPO通过软高斯门控等优化推理后训练，提升Avg@32和Pass@32；FLR将隐式推理分解为多维偏好因子，平均提升3.2%。
Rohan Paul@rohanpaul_aiAI 评分60606月3日，中国信通院宣布中国电信、中国移动、中国联通的“词元产品”服务正式登陆中国算力平台。词元是大模型最小信息单元，具有可计量、可定价、可交易特征。天翼云Token Plan分开发者/中小企业版（基于GLM-5）和个人/家庭版（基于DeepSeek V3.2）；移动云Coding Plan（基于MiniMax-2.5）支持Claude Code等编程工具；联通云推出Coding Plan（整合DeepSeek V4、GLM-5、MiniMax M2.5）和Token Plan（个人版支持DeepSeek V4-Flash、MiniMax M2.5，团队版采用Credits弹性计费）。服务已在中国算力平台·算力超市上架。
NotebookLM@NotebookLMAI 评分6060SiliconFlow@SiliconFlowAI精选AI 评分7272推荐理由：后训练模型能直追 GPT-5.5 和 Claude 4.7，免费两周，对做 agent 和 deep search 的人来说是难得的低成本试错机会。
Artificial Analysis@ArtificialAnlysAI 评分7474
StepFun@StepFun_ai精选AI 评分7777许多研究实验室只在事后才考虑推理效率。Step 3.7 Flash 是一个 198B 稀疏 MoE VLM，由 @StepFun_ai 从一开始就为推理而设计。196B 语言主干，搭配 1.8B 视觉编码器。 为真实世界的智能体工作负载而构建，运行速度最高可达 400 tok/sec。原生多模态理解与行动、可靠的工具使用，以及增强的网页与视觉搜索。 Apache 2.0。立即试用 → https://fireworks.ai/models/fireworks/step-3p7-flash-nvfp4
推荐理由：198B稀疏MoE加MTP解码把速度推到400 tok/s，还开源Apache 2.0，这规格做agent的大脑正合适，做实时应用的可以试试手。
SGLang 与 Miles 在发布首日即支持 NVIDIA Nemotron 3 Ultra。该模型为开放前沿推理模型，总参数 550B、激活参数 55B，采用混合 Transformer-Mamba 架构的 MoE，支持最长 1M token 上下文。针对长运行自主智能体优化，具备工具调用、编码、深度研究与编排能力，后训练使用多环境强化学习（NeMo RL）。SGLang 提供高性能推理，支持 NVFP4 和 BF16 精度，NVFP4 检查点可在 Blackwell GPU 运行；Miles 支持强化学习工作流。模型权重、数据与配方均开源。在智能体生产力、指令遵循与长上下文任务上准确率领先，成本节省高达 30%。
OpenRouter 在 30 场机器人冲刺对决中测试了 11 款大语言模型，共耗 482 美元推理成本。结果指向一个发现：应该重新审视模型 benchmark 的解读方式。
推荐理由：第一次看到对齐税被游戏化量化，Grok能赢是因为它没被训练成好人，Claude总想组队则拖后腿，怎么选模型得看你想要哪种人。
在 Nemotron-3 Nano 模型的 100B token 续训练实验中，任务种子合成数据生成（Task-Seeded SDG）使 MMLU-Pro 提升 1.8 分，平均代码提升 1.9 分，常识理解提升 1.6 分，GPQA 提升 11.1 分，数学成绩保持稳定。该流程利用 lm-eval-harness 中约 70 个公开任务（约 700 子任务）的训练集作为种子，生成新示例并补充推理和上下文，经过格式校验、去重和答案验证后得到精选合成数据集，用于 Nemotron Ultra 和 Super 训练。
推荐理由：NVIDIA 公开了 Nemotron 训练中造合成数据的详细方法论，用 70 个公开任务训练集做种子生成新题目，100B token 实验把 GPQA 拉高 11 个点，做预训练的人值得看看他们怎么造的数据。
StepFun@StepFun_aiAI 评分7373StepFun's Step 3.7 Flash sits on the Intelligence vs Output Speed Pareto frontier, scoring 43 on the Artificial Analysis Intelligence Index and is served at over 400 output tokens/s Step 3.7 Flash (open weights, Apache 2.0) is a significant upgrade on Step 3.5 Flash and stands out for its speed and gains in agentic performance (particularly GDPval-AA). 400 output tokens/s is more than double other models of a similar size class. Contributing to this speed is that the model has only 11B active parameters and the model ships with trained Multi-Token Prediction heads (3) that predict several tokens in a single forward pass, letting it decode multiple tokens at once using speculative decoding. Key results for Step 3.7 Flash with the high reasoning level: ➤ 4 point Intelligence Index improvement: Step 3.7 Flash scores 42.6 on the Artificial Analysis Intelligence Index, up 4 points from Step 3.5 Flash 2603 (38.5). It is equivalent to Qwen3.5 122B A10B (41.6) and trails MiniMax-M2.7 (49.6) and DeepSeek V4 Flash (Max Effort, 46.5) ➤ Speed-intelligence frontier: Step 3.7 Flash achieves ~400 output tokens/s on StepFun's first-party API, placing the model on the Intelligence vs Output Speed Pareto frontier. StepFun has released the weights for this model and we expect several third-party providers to serve this model ➤ Agentic capability improvements: Step 3.7 Flash improves over Step 3.5 Flash 2603 across our agentic evaluations, in both GDPval-AA (real-world agentic tasks) and TerminalBench Hard (agentic coding and terminal use). It achieves a GDPval-AA Elo of 1298, up from 1070 for Step 3.5 Flash 2603, and it's TerminalBench Hard score increases to 35.6% from 32.6%. AA-LCR (Long Context Reasoning) improves to 63.7% from 54.3%. Scores for other evals remain relatively flat ➤ Weaker on knowledge and hallucination than peers: While Step 3.7 Flash trails competitors overall on AA-Omniscience (-38), it improves from Step 3.5 Flash 2603 (-44). It has an AA-Omniscience accuracy of 25.4% and a hallucination rate of 84.4% ➤ Native multimodal support, new in this generation: Step 3.7 Flash introduces a 1.8B-parameter vision encoder for native image understanding, where Step 3.5 Flash was text-only. On MMMU-Pro (multimodal reasoning) it scores 75.3%, roughly matching Qwen3.5 122B A10B (75.0%). Among its same-size open weights peers, MiniMax-M2.7, DeepSeek V4 Flash, and gpt-oss-120b are text-only Key model details: ➤ Context window: 256K tokens ➤ Parameters: 198B total, 11B active (MoE). At BF16 native precision, Step 3.7 Flash requires ~400GB to store the weights. StepFun has also released FP8 (~200GB) and NVFP4 (~100GB) versions for lower-memory deployment ➤ License: Apache 2.0 ➤ Availability: Currently Step 3.7 Flash is available on @StepFun_ai 's first-party API
Artificial Analysis@ArtificialAnlysAI 评分6767AMD 客户端业务高级副总裁拉胡尔·蒂库在 2026 台北国际电脑展上回应英伟达推出 RTX Spark 入局 AI PC 处理器市场，表示欢迎竞争，认为大容量本地内存对 AI 智能体工作负载至关重要。AMD 的 Strix Halo 及后续 Gorgon Halo 足以与 RTX Spark 竞争，Gorgon Halo 采用 Zen 5 CPU 和 RDNA 3.5 GPU，统一内存最高 192GB。软件生态方面，AMD 主推 ROCm，称 CUDA 的壁垒影响已较三年前下降，开发者迁移难度较低。
