https://camcat.nethttps://camcat.net/hothttps://camcat.net/insightshttps://camcat.net/rankinghttps://camcat.net/geohttps://camcat.net/methodologyhttps://camcat.net/subscribehttps://camcat.net/stories/claude-code-v2-1-261-mtnnq8sphttps://camcat.net/stories/v2-1-260-mtm9k017https://camcat.net/stories/the-latest-ai-news-we-announced-in-august-2026-mtlbhq2dhttps://camcat.net/stories/introducing-gemini-3-5-flash-cyber-mtk5fc5zhttps://camcat.net/stories/gemini-omni-1-1-flash-lets-you-build-with-more-control-mtlhq2z7https://camcat.net/stories/experiment-with-qwen3-8-flash-next-on-nvidia-gb300-nvl72-for-age-mtkq8onshttps://camcat.net/stories/gemini-robotics-2-brings-whole-body-intelligence-to-robots-mtk5fc5xhttps://camcat.net/stories/jit-agent-scaling-harness-intelligence-via-just-in-time-harness--mtnucyqlhttps://camcat.net/stories/v0-14-24-mtkqqd5rhttps://camcat.net/stories/claude-s-new-system-prompt-really-doesn-t-want-to-reproduce-song-mtnqxdhthttps://camcat.net/stories/v0-14-22-mtkqqd5zhttps://camcat.net/stories/v0-14-23-mtkqqd5whttps://camcat.net/stories/safety-overview-gpt-6-astra-mtm9jzzwhttps://camcat.net/stories/deploy-a-multimodal-whatsapp-ordering-assistant-with-amazon-bedr-mtnnlkiqhttps://camcat.net/stories/孟加拉语习语基准数据集发布-phi-4-mini-kimi-k2与gemini-2-5-flash各擅胜场-mtns29wlhttps://camcat.net/stories/openai发布gpt-6-astra-arc-agi-3得分99-9-api定价对标claude-fable-5-mtnqrd5rhttps://camcat.net/stories/designing-lifecycle-policies-for-agentcore-memory-mtnnlrhghttps://camcat.net/stories/nvidia-技术博客-用-nemoclaw-搭建记忆驱动的-ai-chief-of-staff-智能体-mtnnmrx9https://camcat.net/stories/gemini-cli-夜间版-v0-60-0-发布-移除硬编码-crux-api-密钥-强化-mcp-oauth-与扩展加载器安-mtnpjsbahttps://camcat.net/stories/aws-发布使用-textract-预处理复杂账单文档的-bedrock-知识库定制方案-mtnnmef5https://camcat.net/stories/aws发布开源控制平面hyperpod-instantstart-支持ai代理驱动集群运维-mtnnm82phttps://camcat.net/stories/simon-willison发布8月赞助者通讯-含openai意外网络攻击-claude自动模式等细节-mtnqmgnrhttps://camcat.net/stories/sglang-v0-5-19-发布-新增-beam-search-deepep-v2-支持-qwen3-8-等模型-mtnsrr7bhttps://camcat.net/stories/openai-codex-0-153-2-发布-修正-gpt-6-astra-快速档描述文字-mtnqrd6thttps://camcat.net/stories/gemini-robotics-er-2-powering-robotics-with-video-understanding--mtk5fc5whttps://camcat.net/stories/openai-codex-0-153-1-发布-支持通过-api-配置-gpt-6-astra-不更改默认模型-mtnqt38dhttps://camcat.net/stories/机器人数据创企xdof洽谈b轮融资-估值约12亿美元-mtnpe3n7https://camcat.net/stories/v0-1-3-mtkqqaifhttps://camcat.net/stories/v0-1-4-mtkqqaiehttps://camcat.net/stories/llama-cpp-b10813-发布-为-adreno-gpu-新增-opencl-xmem-sdpa-路径-mtnnpw43https://camcat.net/stories/llama-cpp-b10814-发布-opencl-后端新增-9-种逐元素算子并大幅优化拷贝路径-mtnnpt6ghttps://camcat.net/stories/llama-cpp-b10816-发布-为-apple-m3-补齐-metal-上剩余量化类型的快速向量调优-mtnnplemhttps://camcat.net/stories/accessing-openai-models-on-amazon-bedrock-from-australia-with-gl-mtkq8no5https://camcat.net/stories/introducing-gemini-3-6-flash-3-5-flash-lite-and-3-5-flash-cyber-mtk5fc5zhttps://camcat.net/stories/introducing-gemini-3-7-flash-mtk5fc5thttps://camcat.net/stories/v0-1-0-mtkqqaihhttps://camcat.net/stories/introducing-computer-use-in-gemini-3-5-flash-mtk5fc63https://camcat.net/stories/advancing-price-performance-for-developers-with-gpt-5-6-in-kiro-mtk4xc3ghttps://camcat.net/stories/start-building-with-nano-banana-2-lite-and-gemini-omni-flash-mtk5fc62https://camcat.net/stories/intelligent-transcription-with-gemini-3-5-transcribe-mtk5fc5shttps://camcat.net/stories/v1-0-0-mtkqqb7bhttps://camcat.net/stories/v0-2-0-mtkqqaibhttps://camcat.net/stories/gemini-omni-1-1-flash-lets-you-build-with-more-control-mtk5fc5qhttps://camcat.net/stories/github-发布-project-hydrafusion-研究预览-跨模型编排实现前沿编码质量-mtnnqvachttps://camcat.net/stories/introducing-weathernext-3-our-most-advanced-and-accurate-global--mtm9jzzzhttps://camcat.net/stories/llama-cpp-发布-b10809-版本升至-0-4-0-修复多后端并调整-asi-相关代码-mtnnpzuphttps://camcat.net/stories/from-code-to-diagrams-agentic-architecture-documentation-with-am-mtkq8noehttps://camcat.net/stories/trinity-agentic-ai-powered-transition-planning-for-students-with-mtkq8noghttps://camcat.net/stories/modernizing-and-scaling-support-operations-with-generative-ai-on-mtkq8no8https://camcat.net/stories/how-an-aws-team-detects-dashboard-content-failures-at-scale-usin-mtkq8nobhttps://camcat.net/stories/llm-gemini-0-34-发布-支持-gemini-3-8-flash-提供低-中-高三档思考级别-mtnqspmjhttps://camcat.net/stories/v0-33-0-mtka96znhttps://camcat.net/stories/nvidia-博客-nemotron-3-5-lightning-与-qwen3-8-27b-现可本地部署于-jetson-推理-mtnnp6lchttps://camcat.net/stories/蒂姆-库克卸任苹果ceo-约翰-特努斯接任并预告下周有重磅发布-mtnqo81thttps://camcat.net/stories/anthropic-python-sdk-1-4-0-发布-用量报告新增-claude-tag-分类-部分接口支持工作区-id-mtnqaimfhttps://camcat.net/stories/integrating-outlook-with-amazon-quick-for-ai-powered-email-autom-mtm9k00dhttps://camcat.net/stories/best-practices-for-building-agentic-automations-with-amazon-quic-mtm9k00rhttps://camcat.net/stories/the-yield-imperative-turning-ai-infrastructure-into-useful-intel-mtka95elhttps://camcat.net/stories/谷歌发布-gemini-cli-0-59-0-夜间版-改进网页抓取工具的目的地验证与连接路由-mtnqa4w6https://camcat.net/stories/v0-32-15-mtka96zqhttps://camcat.net/stories/b10781-mtld2omyhttps://camcat.net/stories/python-genai-2-22-0-发布-新增-gemini-3-8-flash-模型支持-mtnqainohttps://camcat.net/stories/v0-32-14-mtka96zvhttps://camcat.net/stories/v0-33-1-mtka96zjhttps://camcat.net/stories/v0-33-2-mtka96zghttps://camcat.net/stories/b10797-mtmrbw4hhttps://camcat.net/stories/openai-s-rogue-agents-keep-escaping-with-no-formal-process-to-in-mtnpjgv2https://camcat.net/stories/create-your-best-tracks-yet-with-lyria-3-5-in-gemini-mtnnqjo8https://camcat.net/stories/b10796-mtmjx4hthttps://camcat.net/stories/b10795-mtmhjavjhttps://camcat.net/stories/ai-driven-development-lifecycle-using-amazon-bedrock-agentcore-mtm9k004https://camcat.net/stories/embed-quick-sight-visuals-using-cognito-user-authentication-mtm9k00yhttps://camcat.net/stories/nvidia-pair-virtual-inference-router-expands-available-compute-o-mtm9k00zhttps://camcat.net/stories/migrate-agentic-workloads-to-amazon-bedrock-agentcore-mtm9k00ahttps://camcat.net/stories/英国ai基础设施公司nscale拟ipo前融资35亿美元-其中20亿来自英伟达-mtnpykvjhttps://camcat.net/stories/openai训练中的智能体被发现利用公共wiki互传数千条消息-mtnqa4uthttps://camcat.net/stories/another-swarm-of-openai-agents-reached-the-open-internet-without-mtnqmgmlhttps://camcat.net/stories/b10794-mtmedhdghttps://camcat.net/stories/openai-codex-0-153-4-发布-修复-astra-模型选择器可见性并设为默认模型-mtnqa7gvhttps://camcat.net/stories/苹果ceo交接-john-ternus接任-首份备忘录承诺下周有重大发布-mtnq9zw5https://camcat.net/stories/llama-cpp-b10798-llama-print-build-info-输出流可配置-版本信息改走-stdout-mtnnq28yhttps://camcat.net/stories/expert-intelligence-a-new-way-for-you-to-engage-with-trusted-con-mtlhq2z6https://camcat.net/stories/capturing-token-ids-during-agentic-interactions-for-better-reinf-mtkqq7ighttps://camcat.net/stories/v1-3-0-mistral-nemo-mtkqqauxhttps://camcat.net/stories/our-approach-to-bioresilience-mtk5fc60https://camcat.net/stories/litellm-v1-101-0-dev-2-发布-新增-gemini-3-8-flash-定价与多模型路由修复-mtnqaendhttps://camcat.net/stories/a-new-benchmark-for-evaluating-patient-facing-health-ai-agents-mtkqq7i8https://camcat.net/stories/build-observable-enterprise-agentic-retrieval-using-managed-amaz-mtljmwndhttps://camcat.net/stories/cuda-python-1-0-stable-apis-one-foundation-full-platform-access-mtkq8onvhttps://camcat.net/stories/restore-llm-inference-capacity-in-seconds-with-shadow-engine-rec-mtkq8onuhttps://camcat.net/stories/nvidia-neural-modules-2-7-0-mtkqqa7chttps://camcat.net/stories/connect-an-agentcore-runtime-hosted-mcp-server-to-amazon-quick-mtljmwnbhttps://camcat.net/stories/empowering-india-s-next-generation-of-innovators-with-atl-saathi-mtk5fc60https://camcat.net/stories/looking-back-on-microsoft-s-fy26-from-ai-experimentation-to-fron-mtka95ephttps://camcat.net/stories/weathernext-ai-model-achieves-breakthrough-in-forecasting-cyclon-mtk5fc5vhttps://camcat.net/stories/putting-sign-language-ai-into-users-hands-mtk5fc5uhttps://camcat.net/stories/ernie-4-5-vl-28b-a3b-thinking-多模态智能领域的重要突破-mtkqq759https://camcat.net/stories/graviton5-8217-s-improved-design-increases-speed-and-energy-effi-mtkqq7iihttps://camcat.net/stories/draft-nist-guidelines-rethink-cybersecurity-for-the-ai-era-mtkqq8pthttps://camcat.net/stories/build-multi-tenant-agentic-chat-applications-on-enterprise-data--mtljmwnehttps://camcat.net/stories/paddleocr-vl-1-5-面向真实场景文档解析的多任务-0-9b-视觉语言模型-mtkqq74zhttps://camcat.net/stories/rethinking-cloud-operations-with-agentic-observability-mtka95f6https://camcat.net/stories/inaugural-report-pioneered-by-ai-security-institute-gives-cleare-mtkqq95xhttps://camcat.net/stories/nvidia-nvlink-fusion-brings-nvhbm-to-next-generation-ai-infrastr-mtkq8onphttps://camcat.net/stories/new-ai-model-shows-how-to-evacuate-for-fires-one-safe-step-at-a--mtkqq8pkhttps://camcat.net/stories/deploy-an-open-model-from-checkpoint-to-inference-in-two-command-mtkq8onnhttps://camcat.net/stories/scale-av-perception-across-vehicle-platforms-with-nvidia-omniver-mtkq8onlhttps://camcat.net/stories/aws-trainium-frontier-competition-co-design-models-and-kernels-o-mtkqq7i0https://camcat.net/stories/rethinking-security-for-the-age-of-ai-mtka95eshttps://camcat.net/stories/nvidia-nemo-speech-3-0-mtkqqa75https://camcat.net/stories/sop-bench-a-new-benchmark-for-evaluating-ai-agents-on-real-busin-mtkqq7hxhttps://camcat.net/stories/langchain-发布-langchain-core-1-6-2-支持-openai-异步工具-修复-google-genai-mtnnpagahttps://camcat.net/stories/b10791-mtm9k00zhttps://camcat.net/stories/b10793-mtm9k00zhttps://camcat.net/stories/caisi-evaluation-of-deepseek-ai-models-finds-shortcomings-and-ri-mtkqq8puhttps://camcat.net/stories/文心-5-1-正式发布-多榜登顶-模型-34-写得好更懂你-34-mtkqq74shttps://camcat.net/stories/how-to-carry-user-identity-across-federated-kubernetes-and-ai-pl-mtm9k00yhttps://camcat.net/stories/b10792-mtm9k00zhttps://camcat.net/stories/openai-codex-0-153-3-amazon-bedrock-模型选择器新增-gpt-6-astra-mtnqmjcbhttps://camcat.net/stories/研究-给智能体加一条记忆指令-其选择参考档案的来源方式可被显著改变-但不同模型表现并不一致-mtnogic0https://camcat.net/stories/introducing-claude-fable-5-1-on-aws-mtkq8noihttps://camcat.net/stories/openai发布新模型astra-号称迄今最强-先用ai安全工具daybreak开放-mtnrb1v7https://camcat.net/stories/研究-数十亿参数视觉语言模型在神经外科手术器械检测任务中仍表现不佳-mtntbo9bhttps://camcat.net/stories/meta-发布免对齐语音模型-text-ab-3b-参数覆盖配音与全双工对话合成-mtnsxlm7https://camcat.net/stories/b10790-mtm9k010https://camcat.net/stories/google-deepmind发布weathernext-3-ai天气模型实现每小时更新与0-1度分辨率-mtnsf7m9https://camcat.net/stories/v1-2-0-add-mamba-mtkqqauzhttps://camcat.net/stories/新研究-可解释ai热力图不能代表rppg模型真实读取位置-皮肤归因不保证心率估算准确-mtnrj3pzhttps://camcat.net/stories/新研究发现矩阵codi模型对潜在矩阵秩不敏感-截断不影响推理精度-mtnqzr14https://camcat.net/stories/新-ml-代理模型直接在-tcad-网格上预测漂移-扩散场-finfet-设计探索提速-mtnq9vqzhttps://camcat.net/stories/uk-and-australia-pact-on-fast-moving-ai-security-risks-mtkqq95phttps://camcat.net/stories/研究提出集体中心评估框架-偏好推断模型在参与式民主中的准确性不等于公正性-mtnoes1whttps://camcat.net/stories/新基准-execretrieval-代码检索模型在近克隆错误代码面前-首位命中率仅-0-331-mtnoaidkhttps://camcat.net/stories/预注册审计发现-同一模型名在共享端点上并非稳定测量仪器-52-988-次请求重复排名一致性远低于阈值-mtnoa67hhttps://camcat.net/stories/27b混合大模型全网络nvfp4-4-bit量化-gated-deltanet循环层在w4a4下性能与bf16持平-mtno96tohttps://camcat.net/stories/新基准culturalmenubench揭示-多模态模型识别食物高分-却难将菜品归因中国地方菜系-mtnnvyeohttps://camcat.net/stories/新标注方案ntep-r让智能体视觉语言模型每次工具调用都推进推理-mtnnvnvwhttps://camcat.net/stories/langchain-1-4-0a4-mtkqqct0https://camcat.net/stories/tree-species-mapping-in-denmark-a-comparison-of-spectral-tempora-mtnogv27https://camcat.net/stories/interface-induced-trajectory-censoring-mtno3ok3https://camcat.net/stories/ollie-is-betting-its-focus-on-privacy-can-help-it-win-the-ai-ass-mtnrbupdhttps://camcat.net/stories/realcadbench-benchmarking-parametric-cad-modeling-from-industria-mtnth3gthttps://camcat.net/stories/新扩散模型simcast-s2s挑战ecmwf业务系统-称在s2s降水预报上部分指标更优-mtnvht6chttps://camcat.net/stories/剑桥团队用地球观测基础模型嵌入改进概率性天气降尺度-温度技能提升11-5-mtnvh56qhttps://camcat.net/stories/contamination-inflates-scores-but-rarely-reorders-large-language-mtnqslbphttps://camcat.net/stories/新优化器ak-momentum-用激活值控制动量衰减-67m模型预训练最多省46-步数-mtnujhr2https://camcat.net/stories/清华团队开源单张rtx-5090训练2b模型方案-算力成本压至6-9千美元-mtnud3onhttps://camcat.net/stories/研究-用因果干预证明开源大模型内部确实编码材料科学物理机制-mtntnvsthttps://camcat.net/stories/新方法用自然语言描述自动驾驶图像数据集差异-配套基准-ad-diff-bench-发布-mtntatjbhttps://camcat.net/stories/新蒸馏框架-alra-让自回归语言模型预训练蒸馏更关注局部词元关系-mtntahs9https://camcat.net/stories/新型扩散增强模型uno实现llm无损加速-8b版本多项评测优于26b-diffusiongemma-mtnt3ijphttps://camcat.net/stories/沙漠蝗虫栖息地预测-rnn-maxent-将生态位模型与深度学习结合-性能超越传统-maxent-mtnsftpuhttps://camcat.net/stories/研究提出用生成过程多样性预测大模型间的相关失败-38个模型跨10个基准验证-mtns1i30https://camcat.net/stories/谷歌发布ai天气模型weathernext-3-预报分辨率达5公里-可每小时更新-mtnrhx8whttps://camcat.net/stories/bioglyph-将网络拓扑编码为结构化角色语言-提升大模型网络推理准确率最高26个百分点-mtnrc4j8https://camcat.net/stories/erasesae-借助稀疏自编码器在文本生成视频扩散模型中实现精细概念擦除-mtnraw6rhttps://camcat.net/stories/dret框架让66m参数distilbert在生物医学pico分类上接近十倍于其规模的模型-mtnqr0m4https://camcat.net/stories/start-the-year-ai-ready-with-the-google-ai-educator-series-mtm9k012https://camcat.net/stories/开发者实测-gpt-6-astra-图像生成质量全面超越-gpt-5-6-sol-低价档效果更优-mtnpjs9xhttps://camcat.net/stories/82m参数泰语tts模型wayu-paxa-tts-edge发布-支持端侧运行-mtnoutyuhttps://camcat.net/stories/新知识图谱嵌入方法pogre显式应对模式过度泛化-链接预测超越现有模型-mtnogv4bhttps://camcat.net/stories/开源低延迟语音翻译系统-x-translator-注重说话人声音一致性-mtnoabp9https://camcat.net/stories/新基准insitumeasure显示-24款多模态大模型在真实工业仪表读数任务中最佳准确率仅25-7-mtno4a2zhttps://camcat.net/stories/研究-大语言模型可在低数据条件下辅助慢性肾病早期筛查-mtno3skahttps://camcat.net/stories/npci研究团队发布fimi-banking框架-面向印度零售银行的受控对话ai基准-mtno3b6ehttps://camcat.net/stories/新论文-艺术家群体难以约束联邦学习产出的模型使用-研究者提出四条治理原则-mtno25bihttps://camcat.net/stories/研究-音视频扩散模型存在跨模态-语义泄漏-注意力路由沿特定路径出错-mtnnwh2thttps://camcat.net/stories/co-designing-ai-models-using-speculative-decoding-for-faster-llm-mtkq8onchttps://camcat.net/stories/securing-amazon-quick-from-poc-to-production-agents-flows-and-sp-mtkq8nophttps://camcat.net/stories/tokenomics-at-scale-how-jamf-built-real-time-spend-enforcement-f-mtkq8nomhttps://camcat.net/stories/from-theory-to-delivery-how-atos-upskilled-400-engineers-in-agen-mtkq8nokhttps://camcat.net/stories/introducing-gemini-3-8-flash-and-3-8-flash-cyber-mtkba4sghttps://camcat.net/stories/longcounsel-8-a-benchmark-suite-for-longitudinal-depression-trac-mtnp8pj3https://camcat.net/stories/flowbalance-verifier-grounded-self-improvement-from-on-policy-re-mtnofeq1https://camcat.net/stories/shelf-a-synthetic-harness-for-multi-task-bibliographic-benchmark-mtnof1kwhttps://camcat.net/stories/xiaomi-tabldm-a-tabular-foundation-model-technical-report-mtno2zgchttps://camcat.net/stories/viscad-a-foundation-model-suite-with-multimodal-industrial-cad-i-mtnths4ahttps://camcat.net/stories/arxiv-论文提出贝叶斯算子推断主动学习框架-用更少采样构建更稳定的参数动力系统降阶模型-mtnvqvrkhttps://camcat.net/stories/neored-a-knowledge-logic-alignment-multimodal-large-language-mod-mtnnw5ethttps://camcat.net/stories/新基准pawbench评估视频生成模型多可能结局分布-11个系统无一达标-mtnvgbbghttps://camcat.net/stories/科学家发布-thousandworlds-基准-近1800次系外行星气候模拟数据集开放-mtnvf3zohttps://camcat.net/stories/meevo框架-融合自然进化与元认知进化改进大模型自动启发式设计-mtnv24yrhttps://camcat.net/stories/contextual-tamil-spelling-and-grammar-correction-using-progressi-mtnro69ahttps://camcat.net/stories/新论文证明单层注意力模型-k-个头能算-k-位奇偶校验-无条件无法算-k-1-位-mtnuq15khttps://camcat.net/stories/轻量ssim残差精炼器提升脑部mri修复集成模型ssim-mtnuox47https://camcat.net/stories/promptgfm-用图词汇学习驱动文本属性图基础模型-统一llm与gnn-mtnuhky3https://camcat.net/stories/新论文将混合专家模型视为局部聚合-给出路由与稀疏激活的统计风险界-mtnucza6https://camcat.net/stories/gaokerena-为消费级硬件设计的波斯语医疗语言模型家族发布-mtnucmorhttps://camcat.net/stories/新方法让图表生成模型只按图像可见信息训练-减少代码歧义-mtnubvkshttps://camcat.net/stories/scoremix-无需外部数据的扩散模型合成数据方法-人脸识别精度最高提升7个百分点-mtnu2iauhttps://camcat.net/stories/新研究质疑多模态大模型的-潜在空间想象-隐层几乎不响应输入-作者提出文本显式想象框架capimagine-mtntzetyhttps://camcat.net/stories/新基准csm-mtbench评测20余款机器翻译模型处理中文社交媒体俚语与网络新词的能力-mtntv2c1https://camcat.net/stories/新框架以-编码代理-生成可编辑视觉设计-结合扩散模型与html-css-mtntmis1https://camcat.net/stories/新框架specalign-直接从规范文档合成数据-让大模型对齐可快速适配政策变化-mtntlzn7https://camcat.net/stories/研究团队发布首个公开-plax-超声心动图数据集-训练左心室射血分数预测模型-mtnti9e0https://camcat.net/stories/面向电网需求响应的开源强化学习环境-dr-gym-发布-mtntfpjchttps://camcat.net/stories/利用流匹配生成模型分析cms开放数据-研究强子单z暗物质探测灵敏度-mtntbo83https://camcat.net/stories/neuroweaver-让大模型生成轻量级eeg分析流水线-以0-18m参数超越基础模型-mtntazh3https://camcat.net/stories/爱丁堡大学团队用程序合成给仿真轨迹-降维-让大模型更可靠地推理物理系统-mtntan6rhttps://camcat.net/stories/google-s-gemini-spark-can-now-manage-your-google-photos-library-mtnqr648https://camcat.net/stories/开源框架robustseiz发布-在部署前标准化测试癫痫检测模型的鲁棒性-mtnt2l65https://camcat.net/stories/机器学习模型cloudcast-v2实现12小时云量预报-误差较前代降低10-mtnsqw1khttps://camcat.net/stories/tap-path-对virchow2病理基础模型进行任务自适应压缩-减参25-减少35-计算量-mtnsm2jehttps://camcat.net/stories/研究团队发布rarf框架-用于3d脑部mri图像修复并参加brats-2026挑战赛-mtns9in7https://camcat.net/stories/simpledesign-直接在数据空间训练的蛋白质序列与结构联合生成模型-mtnrvjpchttps://camcat.net/stories/新框架charm用轻量微调模型检测道德基础-域外auc最高提升15-3-mtnrvjo4https://camcat.net/stories/b10785-mtljmwnfhttps://camcat.net/stories/caisi-issues-request-for-information-about-securing-ai-agent-sys-mtkqq8prhttps://camcat.net/stories/研究团队发布2854项压实试验开放语料库-称部分已发表压实数据在物理上不可能-mtnrtsbchttps://camcat.net/stories/研究团队发布pace数据集与pacemaker框架-提升个性化助手对用户隐性冲突请求的识别能力-mtnrtrhyhttps://camcat.net/stories/llada-image-开源-6b-扩散-transformer-实现文生图与精细编辑-mtnrtp4ahttps://camcat.net/stories/数据高效-rlvr-框架-de-venus-仅用-10-标签即可保持或提升模型质量-mtnro6a5https://camcat.net/stories/英伟达确认以129-3亿美元收购hugging-face-承诺保留开放平台-mtnrj96jhttps://camcat.net/stories/新任务swim评估大模型模拟学生写作能力-提示词控制有限-监督微调与强化学习更有效-mtnrj3r4https://camcat.net/stories/研究-大模型从规则描述中学习比从示例中学习更可靠-mtnrhg51https://camcat.net/stories/研究-向推理模型注入-思考结束-标记不一定能触发干净的回答阶段-mtnrb982https://camcat.net/stories/研究发现大语言模型存在偏向早期证据的系统性偏差-提出无训练激活引导方法改善上下文冲突处理-mtnraeiihttps://camcat.net/stories/tooldf框架-用音频大模型定位混合真实性音频中的伪造片段-mtnr5sw8https://camcat.net/stories/ai生成菜单图片为何让人不适-训练数据趋同与-同质化-问题-mtnqxbl4https://camcat.net/stories/面向端侧大模型推理的leanstream框架提出推测-细化流式方案-内存占用最高降至原有1-7-5-mtnqx7gvhttps://camcat.net/stories/泰文ocr新模型wayu-paxa-ocr-zero-仅用合成数据训练-手写识别字符错误率降至20-55-mtnqsdirhttps://camcat.net/stories/gemini-cli-夜间版更新-扩展环境变更需用户确认-强化路径与配置文件安全检查-mtnqmm5shttps://camcat.net/stories/5-amazing-visuals-show-how-the-male-fruit-fly-s-brain-map-is-adv-mtm9k013https://camcat.net/stories/arxiv新研究-多模态大模型智能体-strixae-面向复杂失真耦合的音频增强-mtnofzethttps://camcat.net/stories/新研究发现语言模型在不确定时回退到训练语料的词汇先验-可被几何方法识别-mtnob83uhttps://camcat.net/stories/新框架以-认识论担保-标注大模型建议可信度-区分不稳定到广泛支持四级-mtno9b25https://camcat.net/stories/研究者提出语义贝叶斯世界模型-用信念网络替代事实数据库-衔接语言模型与知识图谱-mtno2nsihttps://camcat.net/stories/ppo-stgnn-结合近端策略优化与时空图神经网络-面向云边端协同的dag任务调度算法发布-mtnnvuyrhttps://camcat.net/stories/the-modern-cuda-toolbox-in-practice-a-step-by-step-optimization--mtkq8onahttps://camcat.net/stories/breaking-the-likelihood-trap-variance-calibrated-modulation-for--mtnu8ow3https://camcat.net/stories/toward-physically-grounded-jepa-world-models-for-goal-conditione-mtnqm1f8https://camcat.net/stories/parabridge-bridging-paralinguistic-perception-and-dialogue-behav-mtnu7yz4https://camcat.net/stories/recurtrace-adaptive-latent-reasoning-with-loop-time-memory-mtnrzvybhttps://camcat.net/stories/how-much-do-circuits-tell-us-measuring-the-consistency-and-speci-mtnu1opihttps://camcat.net/stories/b10751-mtlbhq2bhttps://camcat.net/stories/fly-eval-an-evidence-driven-evaluation-protocol-for-safety-const-mtno4a42https://camcat.net/stories/跨交通模式知识迁移新框架-transmod-发布-仅需少量目标数据即可预测多模式出行需求-mtnvkh2zhttps://camcat.net/stories/新研究提出roms-imle-单步生成模型在imagenet-256上实现fid-2-56-mtnvgyv4https://camcat.net/stories/科学家发布163项可审计研究流程的开源库-供语言模型按需加载-mtnvgytuhttps://camcat.net/stories/无需重训或群体标签-对微调更新做svd尾部截断可减少模型虚假关联-mtnva5s7https://camcat.net/stories/palmclaw-开源端侧智能体框架-将设备能力封装为显式工具-mtnv9ld0https://camcat.net/stories/decoupling-turn-taking-from-semantics-a-decoupled-data-approach--mtnruskfhttps://camcat.net/stories/emodistill-用情绪技能蒸馏训练小型语言模型-提升对抗谈判中的议价效用-mtnur51khttps://camcat.net/stories/b10726-mtljmwnjhttps://camcat.net/stories/ex-omni框架-为全模态大模型加入语音伴随的3d面部动画生成-mtnuim2ihttps://camcat.net/stories/研究-语言模型幻觉与-关系线性-高度相关-为可预测幻觉提供新线索-mtnuepjmhttps://camcat.net/stories/新推测解码框架sfad-提升大模型上下文忠实度并实现2-48倍加速-mtnuem1ehttps://camcat.net/stories/surgegen-两阶段扩散模型框架-用于快速生成风暴潮情景-mtnubwgehttps://camcat.net/stories/b10734-mtljmwnihttps://camcat.net/stories/研究者推出-kidbench-评估大语言模型对-7-11-岁儿童的安全性-mtnu1s3dhttps://camcat.net/stories/研究-新一代视觉语言模型并非总能提升自动驾驶性能-mtnu1orzhttps://camcat.net/stories/新序列并行方法basp利用批处理结构削减大模型长序列训练通信开销-mtnu0k3dhttps://camcat.net/stories/videoharness-rsi-递归搜索视频上下文构造程序-在冻结视觉语言模型上提升长视频理解-mtntv2athttps://camcat.net/stories/新论文提出因果世界模型-cwm-形式化定义-连接多领域研究-mtntte32https://camcat.net/stories/langchain-1-4-0a1-mtkqqct9https://camcat.net/stories/v0-5-16-mtkqqcechttps://camcat.net/stories/ex-omni-2d-让多模态对话模型输出带参考形象的协调语音视频-mtntrvtrhttps://camcat.net/stories/idrbench-首个系统评估大模型跨学科研究能力的框架发布-mtntnw11https://camcat.net/stories/开源pcbworld环境发布-让强化学习与llm代理在kicad引擎内完成pcb布线-mtntnjychttps://camcat.net/stories/comap框架让语言智能体与世界模型共同进化-在多项基准上提升决策性能-mtnthoc2https://camcat.net/stories/v0-5-18-mtkqqce4https://camcat.net/stories/codex-0-153-0-mtnqxmsuhttps://camcat.net/stories/新基准揭示医学多模态模型答题依赖捷径-去除文本线索后准确率大幅下滑-mtnt4yzvhttps://camcat.net/stories/新论文-免监督微调的对比课程奖励可教会多模态模型高效使用放大工具-mtnt4gylhttps://camcat.net/stories/跨直播会话风险检测新模型-cs-var-mtnt4dvxhttps://camcat.net/stories/b10776-mtlbhq29https://camcat.net/stories/新方法在输出空间重分配近似标签移除结果-避免重训练分类模型-mtnsz5hxhttps://camcat.net/stories/研究-大模型预训练中知识与重复的再思考-mtnsxlf8https://camcat.net/stories/研究-分阶段进行在线策略蒸馏与rlvr训练-推理模型表现优于联合方案-mtnssl2nhttps://camcat.net/stories/pytorch-2-14-0-发布-新增-torch-switch-多分支-cuda-图内-while-循环及-apple-si-mtnqaekohttps://camcat.net/stories/新攻击-tas-可从已-遗忘-大模型中提取被删提示词-实体恢复准确率-100-mtnsj23zhttps://camcat.net/stories/b10723-mtljmwnkhttps://camcat.net/stories/研究-多数大语言模型在判断食谱是否适合糖尿病患者时偏向保守-mistral-7b与llama-70b表现更优-mtnsa165https://camcat.net/stories/b10728-mtljmwnjhttps://camcat.net/stories/b10730-mtljmwnihttps://camcat.net/stories/b10731-mtljmwnihttps://camcat.net/stories/无需训练的空间推理框架-graft-通过-3d-场景图提升多模态大模型空间推理能力-mtns1qighttps://camcat.net/stories/b10737-mtljmwnhhttps://camcat.net/stories/b10739-mtljmwnghttps://camcat.net/stories/b10740-mtljmwnghttps://camcat.net/stories/b10786-mtljmwnfhttps://camcat.net/stories/b10782-mtlhq2z4https://camcat.net/stories/b10783-mtlhq2z4https://camcat.net/stories/b10784-mtlhq2z3https://camcat.net/stories/自适应谱带宽方法提升核化图方法分类精度-ieee-mlsp-2026-接收论文发布-arxiv-mtnrnjdthttps://camcat.net/stories/新规划方法leap将潜在世界模型平均控制成功率从77-5-提升至94-8-mtnrmj78https://camcat.net/stories/b10754-mtlbhq2bhttps://camcat.net/stories/b10757-mtlbhq2ahttps://camcat.net/stories/v0-26-0-mtka968khttps://camcat.net/stories/failbench-基准发布-最强-vlm-机器人任务成功率判断准确率仅-0-77-mtnqynpzhttps://camcat.net/stories/b10760-mtlbhq29https://camcat.net/stories/研究-模型压缩越强-测试时自适应性能退化越明显-mtnqwwj3https://camcat.net/stories/b10762-mtlbhq29https://camcat.net/stories/因果推断引入基础模型范式-新论文介绍无需微调的因果基础模型-cfm-mtnqr0nohttps://camcat.net/stories/b10777-mtlbhq28https://camcat.net/stories/b10772-mtktcek1https://camcat.net/stories/v0-27-0-mtka968ghttps://camcat.net/stories/文心-5-0-ernie-5-0-2-4-万亿参数的原生全模态大模型-mtkqq74yhttps://camcat.net/stories/v0-28-0-mtka9689https://camcat.net/stories/v0-25-0-mtka968rhttps://camcat.net/stories/langchain-1-4-0a2-mtkqqct5https://camcat.net/stories/v0-5-17-mtkqqce8https://camcat.net/stories/b10769-mtkqqbtghttps://camcat.net/stories/新因果框架区分模型-看似欺骗-与-机制欺骗-实验显示二者可分离-mtno9rujhttps://camcat.net/stories/研究团队发布irwoz-2-0工业人机对话数据集-对话状态追踪准确率大幅提升-mtno4fd2https://camcat.net/stories/release-v5-9-0-mtk4y1juhttps://camcat.net/stories/release-v5-7-0-mtk4y1jxhttps://camcat.net/stories/release-5-8-0-mtk4y1jwhttps://camcat.net/stories/release-v5-11-0-mtk4y1jphttps://camcat.net/stories/release-v5-10-1-mtk4y1jshttps://camcat.net/stories/release-v5-14-0-mtk4y1jfhttps://camcat.net/stories/release-v5-13-0-mtk4y1jkhttps://camcat.net/stories/release-v5-15-0-mtk4y1jbhttps://camcat.net/stories/release-v5-16-0-mtk4y1iyhttps://camcat.net/stories/beyond-bleu-a-case-for-redefining-sign-language-translation-benc-mtnrnl2zhttps://camcat.net/stories/privatehub-contrastive-diffusion-model-for-private-sensor-intens-mtnob2mahttps://camcat.net/stories/modelpedia-用llm从论文自动提取模型行为发现-建可检索公开目录-mtnvlf0dhttps://camcat.net/stories/from-zero-to-hero-an-open-llm-ecosystem-for-armenian-mtnt9fo2https://camcat.net/stories/新视频压缩框架vortec-基于生成流模型-码率降58-解码最快达13-fps-720p-mtnvhqeyhttps://camcat.net/stories/研究-多模态大模型存在-跨模态安全漂移-轻量方法srt可在冻结模型上缓解-mtnvhembhttps://camcat.net/stories/less-than-24-hours-to-apply-for-your-techcrunch-disrupt-2026-sid-mtnqsnnchttps://camcat.net/stories/新方法spacr-单次训练即可生成多置信水平有效预测区间的回归模型-mtnvajvkhttps://camcat.net/stories/crocodil-cross-model-code-editing-with-llms-mtnss1qfhttps://camcat.net/stories/beyond-shallow-alignment-how-post-training-methods-determine-ref-mtnsqqoohttps://camcat.net/stories/b10778-mtlbhq28https://camcat.net/stories/fpco-dialog-a-multi-turn-false-premise-benchmark-for-correction--mtnrzs6shttps://camcat.net/stories/新ai文本检测方法sv-detect-利用冻结语言模型的引导向量提升跨域检测稳健性-mtnuyjoshttps://camcat.net/stories/jina-ocr-v1-efficient-document-parsing-with-speculative-decoding-mtnrbppghttps://camcat.net/stories/bharatgather-a-culturally-informed-benchmark-dataset-for-misinfo-mtnq9sknhttps://camcat.net/stories/fedps框架发布-基于聚合统计的联邦学习数据预处理方案-mtnuj41ahttps://camcat.net/stories/b10729-mtljmwnjhttps://camcat.net/stories/新研究-moe-模型事实回忆未必定位于单个专家-qwen3-层-44-与-mixtral-结果相反-mtnu4xkahttps://camcat.net/stories/用观测降水数据微调ai气象模型-中期预报评分最高提升19-mtnu2li6https://camcat.net/stories/研究提出encoding-probe方法-用可解释特征重建语言模型内部表征-mtnu0k24https://camcat.net/stories/b10742-mtlbhq2chttps://camcat.net/stories/新提示策略用大模型检测对话中的心理操控-论文获coling-2025最佳短文奖-mtntnlv1https://camcat.net/stories/b10756-mtlbhq2ahttps://camcat.net/stories/新方法-编译即训练-将自然语言规则转为可复用本地函数-无需调用大型远程模型-mtnsz5j5https://camcat.net/stories/语言学家发起讨论-用翻译与合成数据训练的语言模型还算-意大利语模型-吗-mtnsed57https://camcat.net/stories/b10727-mtljmwnjhttps://camcat.net/stories/b10733-mtljmwnihttps://camcat.net/stories/b10736-mtljmwnhhttps://camcat.net/stories/b10738-mtljmwnghttps://camcat.net/stories/b10780-mtlbzjijhttps://camcat.net/stories/b10741-mtlbhq2chttps://camcat.net/stories/b10743-mtlbhq2chttps://camcat.net/stories/b10749-mtlbhq2chttps://camcat.net/stories/b10752-mtlbhq2bhttps://camcat.net/stories/b10753-mtlbhq2bhttps://camcat.net/stories/b10758-mtlbhq2ahttps://camcat.net/stories/v1-4-0-pixtral-mtkqqauuhttps://camcat.net/stories/b10759-mtlbhq2ahttps://camcat.net/stories/b10774-mtkwjxxkhttps://camcat.net/stories/b10775-mtkwjxxghttps://camcat.net/stories/b10770-mtks9p33https://camcat.net/stories/b10771-mtks9p31https://camcat.net/stories/b10764-mtkqqbtqhttps://camcat.net/stories/b10766-mtkqqbtkhttps://camcat.net/stories/b10767-mtkqqbtihttps://camcat.net/stories/atv-big-air-tour-turned-3-days-of-work-into-3-hours-with-chatgpt-mtkpz0othttps://camcat.net/stories/traceable-tts-toward-watermark-free-tts-with-strong-traceability-mtnoa6p9https://camcat.net/stories/inside-microsoft-s-two-decade-push-to-cut-water-intensity-while--mtka95f3https://camcat.net/stories/ai-源内-関-動画-更新-mtkqq9e2https://camcat.net/stories/rw-ttt-为带请求自有状态的测试时训练模型提供批量服务-吞吐量提升9-31倍-mtnv9laohttps://camcat.net/stories/新基准-tembed-统一评测表格嵌入模型-最佳选择取决于任务与表示层级-mtnv7xzdhttps://camcat.net/stories/研究表明-强化学习诱导的大模型有害对齐取决于环境设计-模型规模并非总是安全屏障-mtnv3gd1https://camcat.net/stories/新方法识别双变量因果方向-aag在图宾根基准达84-3-加权准确率-mtnv2r2ohttps://camcat.net/stories/新优化器msign-通过稳定秩恢复机制防止大模型预训练崩溃-mtnv27mlhttps://camcat.net/stories/55-billion-r-d-funding-boost-to-unlock-uk-breakthroughs-from-hea-mtkqq961https://camcat.net/stories/新研究用因果最优传输为退化扩散模型定义引导损失-摆脱对分数函数的依赖-mtnt4cdthttps://camcat.net/stories/nvidia-vera-rubin-and-blackwell-set-a-new-standard-for-agentic-a-mtkqq71mhttps://camcat.net/stories/giga-scale-ai-and-the-ethernet-evolution-how-spectrum-x-ethernet-mtkqq71khttps://camcat.net/stories/v1-0-4-mistral-inference-mtkqqav2https://camcat.net/stories/amazon-and-university-of-michigan-give-robots-a-sense-of-touch-mtkqq7idhttps://camcat.net/stories/microsoft-frontier-company-ai-engineering-that-amplifies-and-pro-mtka95f0https://camcat.net/stories/real-world-grounding-in-agentic-ai-mtkqq7ilhttps://camcat.net/stories/nist-allocates-over-3-million-to-small-businesses-advancing-ai-b-mtkqq8pqhttps://camcat.net/stories/unlocking-uk-house-building-with-ai-accelerated-planning-mtk5fc64https://camcat.net/stories/nist-expands-ai-consortium-s-scope-calls-for-new-members-mtkqq8pnhttps://camcat.net/stories/openai-and-microsoft-join-uk-s-international-coalition-to-safegu-mtkqq95rhttps://camcat.net/stories/ai-security-institute-launches-international-coalition-to-safegu-mtkqq966https://camcat.net/stories/powering-america-8217-s-genesis-mission-microsoft-8217-s-commitm-mtka95ewhttps://camcat.net/stories/nist-launches-centers-for-ai-in-manufacturing-and-critical-infra-mtkqq8pshttps://camcat.net/stories/powering-the-next-wave-of-ai-expanding-capacity-with-our-new-dat-mtka95f8https://camcat.net/stories/manage-agents-tools-and-skills-at-scale-with-aws-agent-registry-mtljmwndhttps://camcat.net/stories/34-amazon-research-awards-build-on-trainium-recipients-announced-mtkqq7i3https://camcat.net/stories/how-boomi-scribe-streamlines-documentation-using-aws-mtkq8novhttps://camcat.net/stories/microsoft-expands-azure-ai-and-hpc-infrastructure-with-amd-mtka95eyhttps://camcat.net/stories/from-atari-to-eve-online-building-on-15-years-of-ai-research-in--mtk5fc5thttps://camcat.net/stories/v3-8-0-mtm9k01bhttps://camcat.net/stories/v2-1-259-mtm9k01bhttps://camcat.net/stories/nist-mathematical-proof-supports-transition-to-a-continuous-moni-mtkqq8pihttps://camcat.net/stories/how-zs-democratized-secure-ad-hoc-analytics-with-amazon-sagemake-mtkq8nothttps://camcat.net/stories/run-nvidia-bionemo-nim-microservices-for-protein-structure-predi-mtkq8onjhttps://camcat.net/stories/how-t54-built-a-trust-layer-with-amazon-bedrock-agentcore-paymen-mtkq8norhttps://camcat.net/stories/crusoe据报以300亿美元估值融资30亿美元-距上轮仅10个月-mtnqzt9rhttps://camcat.net/stories/aws-recognized-as-a-leader-in-the-forrester-wave-ai-infrastructu-mtljmwnchttps://camcat.net/stories/研究综述-生成式ai在信息系统中的收益受技术不可靠性与治理真空制约-mtnujp5vhttps://camcat.net/stories/accel-reportedly-in-talks-to-lead-1b-round-for-thinking-machines-mtnr4p44https://camcat.net/stories/创业公司-abliteration-ai-将-去除-ai-护栏-做成商业化服务-提供无防护-glm-5-3-mtnr68tthttps://camcat.net/stories/v1-1-0-add-lora-mtkqqav1https://camcat.net/stories/amazon-is-investing-in-the-lean-focused-research-organization-mtkqq7iahttps://camcat.net/stories/pairing-google-antigravity-with-gemini-3-7-flash-solves-notable--mtlhq2z5https://camcat.net/stories/nist-joins-national-genesis-mission-to-accelerate-ai-innovation-mtkqq8pghttps://camcat.net/stories/研究-将智能体轨迹逆向重建为可执行终端环境-生成3-73万个训练环境-mtno9lkohttps://camcat.net/stories/多智能体框架-bioinfoysis-在生物信息学基准测试-bixbench-上准确率达-82-4-mtno2ol0https://camcat.net/stories/proactive-cyber-defense-for-governments-and-enterprises-mtlbhq2dhttps://camcat.net/stories/新研究对比结构先验循环单元与echo-state网络-优势取决于具体任务-mtnvyy3ihttps://camcat.net/stories/研究者提出-掷骰子法-为反复查询llm品牌推荐建立标准化审计协议-mtntn5kuhttps://camcat.net/stories/研究-联邦学习梯度更新可近乎完美泄露客户端身份-噪声防御有效-mtntikkwhttps://camcat.net/stories/中国软件公司49名员工30个月面部表情数据公开-称最长自然工作场景情绪数据集-mtnt4273https://camcat.net/stories/guidance-ai-security-institute-frontier-ai-trends-report-factshe-mtkqq95yhttps://camcat.net/stories/eneas-为视频与3d重建引入文本提示的实例跟踪与语义分割统一方法-mtnrnxnkhttps://camcat.net/stories/2026-pnpl-脑机接口竞赛转向词语分类-新增-libribrain100-数据集测试跨被试泛化-mtnrhkkihttps://camcat.net/stories/研究提出-metaquill-kip-结合核方法与元学习-实现稀疏观测下的神经场重建-mtnraljthttps://camcat.net/stories/新网络控制框架将截止时间感知与演示驱动强化学习相结合-mtnqo4ayhttps://camcat.net/stories/ernie-5-0-preview-1103-登陆-lmarena-文本榜单-mtkqq757https://camcat.net/stories/nvidia-neural-modules-2-6-2-mtkqqa7dhttps://camcat.net/stories/96页综述梳理协作学习从欧氏数据到图结构数据的演进-系统化图分布场景与问题定义-mtnpe0vdhttps://camcat.net/stories/研究发现联邦学习入侵检测中隐私保护与罕见攻击检测能力存在冲突-mtnofzg2https://camcat.net/stories/omnirsclip框架-将clip扩展到sar-多光谱与高光谱遥感数据-mtnofer1https://camcat.net/stories/reflect-sql-多阶段自反思框架在bird基准上实现72-03-执行准确率-论文被pakdd-2026录用-mtnob2huhttps://camcat.net/stories/清华等机构研究-单条查询即可恢复在线蒸馏大部分收益-训练瓶颈转向算法步骤效率-mtno9tjnhttps://camcat.net/stories/新方法tahi通过人机交互实现智能体个性适配-成功率提升最高20-9-mtno9ja1https://camcat.net/stories/研究-将指令复制两次可将医疗问答通过率提升2-95个百分点-mtno4e66https://camcat.net/stories/gps-bench-用立法记录-游说披露等实证数据检验llm政策模拟的基准-mtnnw98ohttps://camcat.net/stories/frenet框架用sam视觉先验重构病灶分割特征-在etis数据集上dice较sota提升5-0-mtnnw5r2https://camcat.net/stories/新基准-dsb-ifeval-测试全双工语音助手从角色推断对话行为的能力-mtnntfk2https://camcat.net/stories/研究提出omegause-sop系统-将人类专业软件操作演示转化为gui智能体可复用技能-mtnvhpg3https://camcat.net/stories/新研究提出-hartley-神经算子-用纯实数基底替代-fno-的复数-fft-按算子对称性选择谱基底-mtnvfo6vhttps://camcat.net/stories/generic-fno-首个在函数空间内完整嵌入非平衡热力学结构的神经算子-mtnvahxjhttps://camcat.net/stories/83项研究系统性综述-过半llm测试预言机在无规范依据下做出判定-mtnv8r16https://camcat.net/stories/新框架cams以原子事实锚定摘要句子-使多文档摘要可逐句溯源-mtnv2wbjhttps://camcat.net/stories/新距离度量view-distance-将样本投影到二维平面求和-把k-means距离计算复杂度从o-n2-降到o-k-mtnur08ihttps://camcat.net/stories/para-pipe框架-分层调度算子并行以平衡边缘soc推理吞吐与延迟-mtnuqku6https://camcat.net/stories/新研究用参数化图论刻画张量网络状态表示与层析的学习复杂度-mtnuq5xohttps://camcat.net/stories/spruce-面向云端外包向量检索的私有检索方案-查询速度较此前最优提升最高22-9倍-mtnubn4rhttps://camcat.net/stories/研究发现学习式无蜂窝-isac-调度器高精度仍可能违反硬约束-单个恶意-ap-可致解大量不可行-mtntzgpqhttps://camcat.net/stories/ai多智能体工作流从肿瘤文档中提取结构化数据-f1达85-0-mtntv6c8https://camcat.net/stories/新框架-tad-align-用轮次感知奖励压制多模态多轮对话攻击成功率-mtnttrtahttps://camcat.net/stories/因果反事实-rag-论文被作者撤回-称需大幅修订后重写-mtntrwrahttps://camcat.net/stories/aip-用有向执行图表示智能体技能-编译后claude-sonnet任务通过率从53-升至67-mtntihl4https://camcat.net/stories/研究者提出手语反向词典-视频字幕与描述检索实现开放词表手语识别-mtntho9bhttps://camcat.net/stories/研究-用-ukasiewicz多值逻辑公理完整刻画深层relu网络函数等价性-mtnt4x8ghttps://camcat.net/stories/新算法xcal-fl在联邦学习中动态校准差分隐私噪声-兼顾预测精度与解释保真度-mtnsycedhttps://camcat.net/stories/litellm-v1-99-1-mtnqmm4nhttps://camcat.net/stories/研究-按分钟采样足球监控数据-逐分钟贴伤情标签不可靠-改用固定时间点评估-mtnss6lkhttps://camcat.net/stories/新投影算法实现bures-wasserstein重心单位步长下维度无关线性收敛-mtnsmrj7https://camcat.net/stories/新基准patchbench揭示-现有评估使ai漏洞修复智能体成功率虚高1-83倍-mtnsmhw9https://camcat.net/stories/arxiv新研究-将无资源方言翻译建模为多智能体决策空间-轻量微调使叙利亚阿拉伯语方言标记使用率翻倍-mtnsfowwhttps://camcat.net/stories/3d混凝土打印可建造性模拟-真实挤出丝形状显著影响预测精度-mtnsf1ochttps://camcat.net/stories/新研究提出表征相似性优化-以人类道德分类重构llm潜在表征以提升对抗鲁棒性-mtnsec82https://camcat.net/stories/研究团队提出remind方法-应对早产儿姿态估计中的标注噪声问题-mtnsdw00https://camcat.net/stories/研究-自编码器参数谱特征可作为训练数据集的向量表示-mtns9mr2https://camcat.net/stories/新研究提出两阶段预测框架-先预测服务请求再估算cpu负载-smape低至3-57-mtns61bnhttps://camcat.net/stories/ratl-将历史预测残差作为可检索记忆-提升多变量时序预测-mtns29nzhttps://camcat.net/stories/轮腿机器人触觉补偿框架fwbc-vla-无传感器估计接触力-mtns1ae7https://camcat.net/stories/可解释异常检测新方法diffint-以可读区间作为自编码器瓶颈-在48项基准上平均排名最佳-mtnrzs5lhttps://camcat.net/stories/新框架lugl让lightgbm等梯度提升树在强化学习游戏中媲美甚至超越dqn-mtnrhg3vhttps://camcat.net/stories/daybreak-for-frontline-defenders-1b-to-protect-essential-service-mtm9jzzvhttps://camcat.net/stories/litellm-发布-v1-97-1-仅容器镜像更新-修复已知-cve-mtnqrd87https://camcat.net/stories/研究-levelsyn利用图表神经网络弥合逻辑综合与物理设计的鸿沟-降低功耗并加速收敛-mtnqr0l1https://camcat.net/stories/trace-将接触历史存入边记忆的图网络模拟器-长时程颗粒态预测误差最高降62-mtnqm8qxhttps://camcat.net/stories/微软发布-semantic-kernel-net-1-80-1-移除已退役-openai-assistants-集成测试-mtnqaelzhttps://camcat.net/stories/新跨模态检索方法wide-用通配符解码抑制强制幻觉-在m-beir上超越现有生成式检索-mtnq9q9fhttps://camcat.net/stories/research-assessment-of-ai-capabilities-and-the-impact-on-the-uk--mtkqq95vhttps://camcat.net/stories/nvidia-neural-modules-2-7-3-mtkqqa77https://camcat.net/stories/how-to-size-gpus-for-ai-inference-and-tco-without-overspending-mtkq8onhhttps://camcat.net/stories/v1-6-0-mistrall-goes-small-3-1-with-vision-mtkqqaurhttps://camcat.net/stories/研究发现软件工程师在ai采用中经历问责焦虑与职业身份冲击-mtnogid8https://camcat.net/stories/随机特征法高维逼近谱收敛获证-高精度与严重病态由同一机制驱动-mtnofliehttps://camcat.net/stories/研究-图神经网络可将电子海图更新自动分类为关键或非关键航行风险-mtnoes36https://camcat.net/stories/fglguard-用联邦图学习为llm多智能体系统做隐私保护安全防护-不加池化数据便超集中式上限-mtnoeql6https://camcat.net/stories/新防御框架geo-defender将生成式搜索的恶意geo攻击成功率从50-32-降至6-20-mtnoels7https://camcat.net/stories/研究提出两阶段语音匿名化框架-保护隐私同时减少下游任务性能损失-mtnoa9ujhttps://camcat.net/stories/100个自治ai智能体在数学证明任务中自发作弊-另一群智能体自发举报并抵制-mtno9tiahttps://camcat.net/stories/新方法通过离线演化环境难度提升终端ai智能体训练效果-mtno9ff8https://camcat.net/stories/ai-论文评审新框架-equireview-r-减少过度批评且不遗漏重大问题-mtno3b53https://camcat.net/stories/新架构capa让ai会议代理的沉默率从51-4-降至2-5-mtno3873https://camcat.net/stories/arxiv-新论文-为智能体系统提出三类-价值保持-架构模式-mtno31qwhttps://camcat.net/stories/新检索系统stair利用目录结构降低rag幻觉-recall-1达82-6-mtno2s9ihttps://camcat.net/stories/dnative-twin-用图数字孪生记录并回放ai代理决策-补上工具状态盲区-mtnnxkdyhttps://camcat.net/stories/研究-63篇数据中心ai节能论文中半数仅经仿真验证-无一计入水资源消耗-mtnnwy4fhttps://camcat.net/stories/研究-kv缓存逐出中时间聚合比评分的排序保持更关键-新方法inertiakv提速1-34-1-46倍-mtnnvws0https://camcat.net/stories/新机制smc让llm智能体并行预执行工具调用序列-延迟最高降44-9-mtnnsffnhttps://camcat.net/stories/openai-python-sdk-发布-v3-7-0-更新用量-api-文档并优化-embedding-的-numpy-检查-mtnnqbp2https://camcat.net/stories/x-2095651088502591861https://camcat.net/stories/mrbeast-partners-with-gemini-to-turn-impossibly-big-ideas-into-r-mtlbhq2dhttps://camcat.net/stories/bs-take-the-hint-interactive-multitracer-pet-ct-lesion-segmentat-mtk5fl5xhttps://camcat.net/stories/gradient-update-mismatch-rethinking-conflict-free-training-of-ph-mtk5fl5vhttps://camcat.net/stories/adaptive-critical-token-aware-retrieval-for-repository-level-cod-mtk5fl5ehttps://camcat.net/stories/studentsim-training-llm-based-student-simulators-mtk5fl5lhttps://camcat.net/stories/the-structure-of-quantization-damage-in-llms-why-the-next-bit-sh-mtk5fl5nhttps://camcat.net/stories/retrieved-but-not-ranked-surface-form-bias-in-structural-retriev-mtk5fl5whttps://camcat.net/stories/how-ai-native-companies-turn-workflows-into-operating-capability-mtk4xc36https://camcat.net/stories/llm-guided-reinforcement-learning-for-adaptive-npc-behavior-in-m-mtntc9wehttps://camcat.net/stories/towards-scaling-reinforcement-learning-to-massive-populations-le-mtntbsv4https://camcat.net/stories/air-ground-collaborative-vision-and-language-navigation-via-shar-mtnogv3bhttps://camcat.net/stories/coupled-scaling-a-representational-accessibility-framework-for-n-mtnsegs2https://camcat.net/stories/the-civilization-framework-sovereign-anchored-communication-betw-mtnog2m9https://camcat.net/stories/observerbench-testing-mechanistic-estimates-for-intervention-and-mtnof0rmhttps://camcat.net/stories/modern-transformers-are-implicit-hybrids-from-functional-differe-mtnpjer8https://camcat.net/stories/物理信息机器学习框架实现任意路径下的机器人辐射源定位-mtnvziachttps://camcat.net/stories/vestigekv-the-nope-mla-kv-cache-carries-its-own-eviction-signal--mtntlma7https://camcat.net/stories/研究提出dbs-hybrid连续学习新方法-应对无人机航拍地点识别的灾难性遗忘-mtnvt17uhttps://camcat.net/stories/新论文证明relu网络验证问题在参数化复杂度上为w-l-1-hard-枚举法近乎最优-mtnvnnaphttps://camcat.net/stories/综述-机器学习不确定性量化在生物信号处理中的应用现状与挑战-mtnvlvfjhttps://camcat.net/stories/mugen-首个同时防御节点分类-图分类与链接预测的图数据不可学习框架-mtnvlexihttps://camcat.net/stories/研究人员提出基于活动理论的基础本体-用于定义人机对话中的矛盾-mtnvkh1khttps://camcat.net/stories/新共训练方法wdl-opd稳定在线策略蒸馏-提升qwen3数学与代码性能-mtnvagrihttps://camcat.net/stories/rcprob为树集成规则提取引入概率估计-对数损失中位降幅达71-9-mtnv8k8ihttps://camcat.net/stories/新稀疏gpu优化方法加速物理约束生成式采样的非线性投影-mtnv7xy5https://camcat.net/stories/arxiv新研究-从广义统计熵导出两类注意力算子-替代softmax的代数衰减方案-mtnv2g9ghttps://camcat.net/stories/研究发现小规模预训练数据配比实验失准源于-重复不匹配-提出重复率控制方法-mtnv1enmhttps://camcat.net/stories/研究-共识学习在医学ai协作中可实现与联邦学习相当的精度-但训练时间和通信成本大幅降低-mtnurbaehttps://camcat.net/stories/新方法用金丝雀令牌识别ai抓取器-22个llm系统中发现未公开数据源-mtnur079https://camcat.net/stories/研究-安全顶会可复现工件中近45-的静态分析发现具安全相关性-mtnuqkt0https://camcat.net/stories/新研究-从隐藏状态直接注入推理-token-gsm8k-准确率超过原思维链-mtnuq181https://camcat.net/stories/where-does-harness-optimization-value-live-localized-gains-and-t-mtnp8qruhttps://camcat.net/stories/研究-稀疏视角图像生成完整3d场景的新方法spar3s-无需3d真值监督-mtnuji1uhttps://camcat.net/stories/新综述提出-视觉遇见图-把图的图像当作推理与学习的一等输入-mtnuii4rhttps://camcat.net/stories/新论文-神经网络可从局部统计中学会解析特定层级语法-无需显式句法先验-mtnui2s9https://camcat.net/stories/研究揭示-唯一素因子分解不保证有限表示性-提出fsrp与ptld新性质-mtnuhkzbhttps://camcat.net/stories/新研究-跨层相关的初始化决定残差网络无穷深度极限-可由厄米特过程驱动-mtnueep0https://camcat.net/stories/论文提出-imagine-then-plan-框架-用自适应前瞻想象力提升-ai-智能体复杂任务规划能力-mtnueaykhttps://camcat.net/stories/fadti-面向多元时间序列插补的傅里叶与注意力驱动扩散框架-已获-icdm-2026-接收-mtnud1khhttps://camcat.net/stories/anybox-零样本9自由度纸箱位姿估计框架-检测精度最高提升36点-mtnuburwhttps://camcat.net/stories/卷积层滤波器的新几何参数化-grassmann-plu-cker-映射被证明是闭嵌入-无冗余且参数空间光滑-mtnu8sz0https://camcat.net/stories/gpu驻留批量lm求解器将符号回归常量优化提速-使evogp在18个问题中恢复10个控制方程-mtnu8nixhttps://camcat.net/stories/研究-人类心理问卷无法可靠衡量llm在日常交互中的行为-mtnu7rfzhttps://camcat.net/stories/race-aimc-用统计认证为模拟存内计算芯片选单颗可靠加速器-能耗模拟降低69-mtnu074uhttps://camcat.net/stories/研究-深度学习框架可设计收入最优的数据市场信号方案-mtntzel8https://camcat.net/stories/新型训练正则化egr让机器人多模态策略在传感器失效时成功率最高提升120-mtntv3j6https://camcat.net/stories/新框架-hard-routed-mor-lora-以硬路由组合冻结的推理-lora-专家-减少可训练参数-mtntn407https://camcat.net/stories/meta-is-paying-to-peek-at-how-you-use-their-latest-ai-model-mtnraoodhttps://camcat.net/stories/新水印方案-sbw-让-llm-文本标记速度提升-开销低于-1-mtntihmhhttps://camcat.net/stories/surfspec-不依赖脱靶结构信息的先导化合物优化框架-通过几何失配下界提升特异性-mtnth3nmhttps://camcat.net/stories/knowvis-将视频讲座转为概念图驱动的可视化摘要-emnlp-2026-论文称可降低认知负荷-mtntboaihttps://camcat.net/stories/nyu团队提出前瞻性输入修正-深层连续时间循环网络的深度梯度衰减得以缓解-mtnt9jsmhttps://camcat.net/stories/arxiv新研究-医疗聊天机器人幻觉检测中-单轮标注基准易漏计事实错误-mtnstcuchttps://camcat.net/stories/新无监督方法用-同类对比-识别基站低效能耗-无需标注数据-mtnst0oyhttps://camcat.net/stories/the-full-stack-behind-abundant-intelligence-mtk4xc3ehttps://camcat.net/stories/how-loveholidays-is-making-everyone-a-builder-with-codex-mtk4xc3ehttps://camcat.net/stories/新实验设计方法-raed-在部分可识别性下优先控制误排除-arxiv-论文给出理论保证与基准验证-mtnslybxhttps://camcat.net/stories/研究发现-合成数据在嵌入空间中的位置比数量更能影响nlp分类性能-mtnsfoy2https://camcat.net/stories/研究否定重尾分布下受限特征值样本复杂度猜想-colt-2015-开放问题获否定答案-mtnscc89https://camcat.net/stories/新论文提出lngram-v2-为transformer加入可扩展的离散潜在n-gram记忆机制-mtns5wk3https://camcat.net/stories/新论文-可微因果发现中的-先验指导-机制违反三条必要条件-单个错误先验可在87-97-试验中压制真实边-mtns1sgdhttps://camcat.net/stories/新强化学习算法tigpo让长期任务llm智能体跨策略版本复用历史路径-mtns0plyhttps://camcat.net/stories/wand-面向表格数据的可解释无监督异常检测器-评分与特征归因同源产出-mtnruw0hhttps://camcat.net/stories/announcing-the-quot-ai-agent-standards-initiative-quot-for-inter-mtkqq8pphttps://camcat.net/stories/研究-离线多智能体强化学习中-任务多样性比数据集规模更能提升零样本泛化-mtnrjmlihttps://camcat.net/stories/新基准数据集-finrag-qa-针对跨机构银行财报问答-多阶段-rag-检索准确率从-0-322-提升至-0-710-mtnrc2jqhttps://camcat.net/stories/新方法用深度学习光流从单颗静止轨道卫星提取全球三维风场-mtnr4ktthttps://camcat.net/stories/openai-codex-发布-0-154-0-alpha-3-版本-mtnqzu15https://camcat.net/stories/新框架df-asr为中文语音识别引入语义感知的逆文本规范化-mtnqx7flhttps://camcat.net/stories/path-to-astra-critical-capabilities-and-frontier-safeguards-mtk4xc37https://camcat.net/stories/新方法-latentddm-用局部预训练加模块组合降低神经算子迁移成本-mtnqsm4phttps://camcat.net/stories/jalapen-o-s-first-results-show-industry-leading-speed-and-effici-mtk4xc3fhttps://camcat.net/stories/新流水线框架-pipmre-提升医学关系抽取-无需标注模式-mtnqm8a8https://camcat.net/stories/ernie-5-0-preview-1203-位居-lmarena-中国文本模型榜首-mtkqq756https://camcat.net/stories/混合rag新适配器r2adapter-将图检索使用率最高降低59-且保持答案准确率-mtnpyhachttps://camcat.net/stories/truncgradgs-用截断梯度更新改进-3d-高斯泼溅优化-mtnpyh92https://camcat.net/stories/ernie-5-0-preview-1220-荣登-lmarena-视觉理解榜-为前十唯一国产模型-mtkqq754https://camcat.net/stories/efforts-to-share-best-practices-on-ai-measurement-and-evaluation-mtkqq960https://camcat.net/stories/leangrpo-消除扩散强化学习中的冗余重计算-提速最高1-83倍-mtnpjax6https://camcat.net/stories/研究提出-方程重铸-法-用单一规范算子实现参数化pde零样本外推预测-mtnp8vgkhttps://camcat.net/stories/nvidia-neural-modules-2-7-2-mtkqqa79https://camcat.net/stories/文心大模型5-1-preview-荣登-lmarena-文本榜国内第一-mtkqq74uhttps://camcat.net/stories/how-law-firm-gilbert-tobin-governs-and-scales-ai-with-openai-mtk4xc39https://camcat.net/stories/新框架tabscope按问题类型自适应选择表格局部或全表推理-提升长表格问答准确率-mtnofhlqhttps://camcat.net/stories/研究-利用隐藏状态交互-让llm语音识别在推理时自我纠错-mtnoak4ehttps://camcat.net/stories/新论文提出共同见证证书框架-审计图像编辑的局部到整体不一致性-mtno3ol0https://camcat.net/stories/研究提出htn计划去序方法-去除多余顺序约束同时保持层级有效性-mtno30hphttps://camcat.net/stories/研究-图约束智能体框架提升零售供应链需求自适应改造成功率-mtno2ntnhttps://camcat.net/stories/causecollab-面向异构协同感知的因果统一模态无关网络-登-icml-2026-mtno2acbhttps://camcat.net/stories/simskill-为交通仿真打造终身学习型ai智能体-自主掌握sumo全流程-mtnnxfywhttps://camcat.net/stories/dude-首个面向论文-代码不一致检测的双检测多智能体系统-f1-最高提升-18-7-mtnnt7wxhttps://camcat.net/stories/提示工程框架为通用ai助教实现96种学习者画像的个性化应答-mtnnswk0https://camcat.net/stories/frogs-discrete-neural-sampler-for-independent-alloy-configuratio-mtntfplshttps://camcat.net/stories/mind-the-gap-robustness-risks-in-pii-detection-systems-mtns99xlhttps://camcat.net/stories/no-regret-bayesian-optimization-with-finite-library-input-warped-mtnqo4dghttps://camcat.net/stories/tail-likelihood-reinforcement-learning-mtnpyhblhttps://camcat.net/stories/arxiv-论文因同意授权不足被撤回-涉-ai-生成图像检测框架-hcs-mtnw53xmhttps://camcat.net/stories/the-natural-language-interaction-protocol-and-standard-for-ai-ag-mtno9he7https://camcat.net/stories/analysis-of-prompt-engineering-for-drug-toxicity-prediction-mtnnwtyjhttps://camcat.net/stories/kernelfoundry-用进化搜索代替标准提示词-sycl-gpu-内核平均提速-2-3-倍-mtnvsmrehttps://camcat.net/stories/利用损失轨迹审计密集标注视频-新方法在egoper和cholec80上检测标注错误优于传统基线-mtnvs6pvhttps://camcat.net/stories/rent-a-rag-embedding-space-watermarks-for-auditing-third-party-r-mtntc7qzhttps://camcat.net/stories/研究-硬relu梯度下降的自动微分极限与梯度流不一致-激活事件导致灵敏度缺口-mtnvkz4ghttps://camcat.net/stories/无需训练的检索方法visar能将文档视觉问答rag延迟降低58-7-mtnvkypthttps://camcat.net/stories/prq-kmeans-新语义-id-分词器-工业搜索数据集-hitrate-最高提升-7-4-mrr-提升-11-8-mtnvhphlhttps://camcat.net/stories/latentpress-把对话历史与长文档压缩成连续记忆令牌-4-16-倍压缩供冻结解码器直接读取-mtnvhdi7https://camcat.net/stories/新方法用不到1-10算力估算llm最优超参数扩展律-降低大规模调参成本-mtnvh0jzhttps://camcat.net/stories/tgsr-pinn-面向物理信息神经网络逆向迁移的靶目标引导选择性重加权方法-mtnvgba4https://camcat.net/stories/研究-拒绝回答的脆弱几何形状源于拒绝训练-多样化拒绝起始语可提高抗向量消融攻击能力-mtnvfoz1https://camcat.net/stories/voxreason-listener-free-evaluation-of-source-grounded-speech-pla-mtnt4ccvhttps://camcat.net/stories/opening-mind-by-opening-architecture-analysis-strategies-mtnslyayhttps://camcat.net/stories/llmzero用树搜索自动寻找rl后训练策略-在grpo任务上比基线提升9-至140-mtnv2f3mhttps://camcat.net/stories/新后训练框架lsr-在线性化子空间中修正已训练神经网络-误差可降一个数量级-mtnv1f44https://camcat.net/stories/欢迎在-lmarena-上测试-ernie-5-0-preview-1120-mtkqq758https://camcat.net/stories/tevi框架-用稀疏自编码器按文本条件编辑clip图像嵌入-提升图文检索对齐-mtnv0y81https://camcat.net/stories/新强化学习算法可自适应划分状态空间-求解无界扩散过程控制问题-mtnuyjykhttps://camcat.net/stories/新方法用-注意力轨迹-诊断深度强化学习智能体的特征依赖与偏差-mtnuy7lkhttps://camcat.net/stories/新评测协议与电商数据集taoke-时间级联预测从防泄漏走向保真基准-mtnuwkpbhttps://camcat.net/stories/新方法harp用可学习旋转替代固定hadamard变换-将2bit-llm量化困惑度一致改善-mtnuvnuohttps://camcat.net/stories/memorylace-memory-lifecycle-aware-consolidation-and-evidence-ret-mtnrc2rfhttps://camcat.net/stories/margins-not-windows-training-free-per-step-lossy-speculative-dec-mtnqo4cdhttps://camcat.net/stories/可微分混合建模框架-从实验数据学习并优化化学传输过程-mtnuph6yhttps://camcat.net/stories/新研究将llm评估建模为张量补全-提出半参数推断框架-mtnuph5rhttps://camcat.net/stories/bounded-personas-match-retrieval-on-classification-but-not-regre-mtnpe0onhttps://camcat.net/stories/新方法ease引入视觉证据过程监督-提升多模态rlvr在感知-幻觉等基准上的得分-mtnuihjphttps://camcat.net/stories/f-grpo-面向-rlvr-的困难感知缩放系数-提升罕见正确轨迹学习-mtnuihifhttps://camcat.net/stories/新研究用遗传算法实现贝叶斯网络融合-在树宽约束下优先保留共享结构-mtnui4o3https://camcat.net/stories/新证明-resnet-能以多项式级参数逼近半线性热方程解-突破维度诅咒-mtnuepijhttps://camcat.net/stories/研究者用深度变分框架计算受限近晶液晶的稳定构型-mtnucmpuhttps://camcat.net/stories/新聚类方法面向混合数据-基于预拓扑空间-已在期刊发表-mtnucmnihttps://camcat.net/stories/研究-阿拉伯语方言生成主要靠分布式表征控制-神经元微调仅部分有效-mtnubk3shttps://camcat.net/stories/easysteer-基于vllm的统一框架-宣称推理时llm操控提速10-8-22-3倍-mtnu7z11https://camcat.net/stories/研究-ai开发者与翻译用户社媒分歧明显-研究重点与真实需求脱节-mtnu7mczhttps://camcat.net/stories/edit-框架-用内部信号定位推理错误-训练更遵循评分标准的-llm-评分器-mtnu6qbuhttps://camcat.net/stories/新框架supervalid提出能力对齐的ood验证方法-提升下游扩展预测-mtnu2ic4https://camcat.net/stories/新张量补全框架cttc-以张量形式整合辅助信息-提升药物效果预测精度-mtnu1oqohttps://camcat.net/stories/研究-新框架用可控强化学习平衡科研选题的新颖性-可行性与有效性-mtnu029bhttps://camcat.net/stories/新框架chisagent用多智能体从-二十四史-自动构建中国古代事件分类体系-mtntu0izhttps://camcat.net/stories/立场论文-ai-智能体正让人类监管者认知能力退化-人在回路-并非简单解法-mtnttzrthttps://camcat.net/stories/新音频水印框架-craw-声称可抗神经编解码器-用于合成语音溯源-mtnttxochttps://camcat.net/stories/新方法oqrc收紧机器学习风险控制界限-有限样本下风险差最多降78-64-mtnttrh6https://camcat.net/stories/the-builders-stage-brings-practical-strategies-for-scaling-start-mtnrk6v6https://camcat.net/stories/our-decision-on-cursor-following-its-acquisition-by-spacex-mtk4xc3bhttps://camcat.net/stories/新方法ink将神经符号分层强化学习应用于导航任务-提升样本效率-mtntny4lhttps://camcat.net/stories/观点文章-无标注不等于无人类监督-计算机视觉需明确披露学习管线中的-先验-mtntnxzihttps://camcat.net/stories/emnlp-2026论文提出statefuldiscovery框架-让科学发现智能体在探索中避免过度解读-mtntikm9https://camcat.net/stories/研究-把交互行为统计量直接编码为输入特征-可稳定提升动态图异常检测性能-mtnthoaohttps://camcat.net/stories/研究-利用线性探针可在llm解码前检测拒绝行为-加速越狱攻击-mtnth790https://camcat.net/stories/研究-通用agent框架在cae仿真任务上超越专用多智能体系统-求解器教程是最大增益来源-mtntaziahttps://camcat.net/stories/研究团队提出并发随机博弈首个pac学习框架-可判定纳什均衡是否存在-mtntaikwhttps://camcat.net/stories/新方法sage按梯度效用筛选训练样本-提升离线推理对齐效率-mtnt9fmuhttps://camcat.net/stories/新算法-hood-实现一般博弈中常数个体遗憾-界达-o-n3-log2k-mtnt5218https://camcat.net/stories/paperscout-将学术论文搜索重构为序列决策的自主体-用-pspo-解决多轮训练粒度失配-mtnt45k6https://camcat.net/stories/新估计器消除重尾分布极端分位数处理效应的位置偏移敏感性-mtnt429ihttps://camcat.net/stories/b10763-mtlbhq29https://camcat.net/stories/研究-路由分离不足以阻止moe-lora多领域微调中的负迁移-spawnlora在专家内部动态新增门控子适配器-mtnt3id3https://camcat.net/stories/研究者把量子引力的kontsevich-segal-witten判据引入机器学习-用于评估复数线性映射的谱相位-mtnsz6fahttps://camcat.net/stories/新提示词优化方法espo-精度反超gepa-3-76个百分点-提示词缩短47-mtnsywgxhttps://camcat.net/stories/研究-llm-推理痕迹的可读性不等于可解释性-步骤重要性仅能从文本部分恢复-mtnsyrvthttps://camcat.net/stories/新研究揭示注意力参数化如何影响transformer训练动力学-mtnsyfzxhttps://camcat.net/stories/piloting-the-world-s-first-double-blind-ai-evaluations-mtk5fc5rhttps://camcat.net/stories/离线强化学习新方法mpi-在数据支持范围外实现可控策略改进-mtnsvuvphttps://camcat.net/stories/两阶段强化学习框架tcs-用对抗性测试用例生成提升代码llm性能-mtnsvuughttps://camcat.net/stories/新研究-印地语与马拉雅拉姆语中-仅调整语序或语态便致多语言llm数学推理显著退化-mtnsdw15https://camcat.net/stories/b10724-mtljmwnkhttps://camcat.net/stories/研究提出-creaeval-把-llm-评判拆成分析与打分两步-提升复杂创意任务评估准确率-mtns9ld8https://camcat.net/stories/研究-随机丢弃kv缓存即可媲美最优选择策略-推理吞吐量提升32-43-mtns97qghttps://camcat.net/stories/研究提出qat-fm-用分位数对齐树构建流匹配耦合-降低路径交叉与构建成本-mtns2eg6https://camcat.net/stories/framebench-一个基于框架语义的英语和日语语言理解基准-测试-llm-区分上下文动词框架的能力-mtns1af9https://camcat.net/stories/研究-ai-agent-框架盲目信任生命周期钩子更新-可被供应链攻击植入恶意命令-mtns0olhhttps://camcat.net/stories/研究提出-无时间步-动力学仿真法-用自一致性迭代取代逐步时间推进-mtnrv9hzhttps://camcat.net/stories/研究-单指标评估会高估llm鲁棒性-扰动影响需分层检验-mtnrv9gjhttps://camcat.net/stories/nvidia-neural-modules-2-7-1-mtkqqa7ahttps://camcat.net/stories/新论文提出基于对称性的因果推理数学框架-旨在超越独立同分布数据-mtnrmftchttps://camcat.net/stories/arxiv-研究提出-b2b-客户转化预测框架-准确率达-91-mtnrj4w8https://camcat.net/stories/b10750-mtlbhq2chttps://camcat.net/stories/研究-tabm在车险精算定价中数据扩展效率显著优于表格transformer与mlp-mtnr619rhttps://camcat.net/stories/研究提出-conformal-relevance-框架-以少量人工输入替代逐任务提示工程-提升-nlp-抽取任务的覆盖与简洁性-mtnr60iohttps://camcat.net/stories/新基准lexissue覆盖430起中国民事案件-用检索增强提升法律争议焦点识别-mtnr4ksshttps://camcat.net/stories/introducing-the-admin-plugin-for-chatgpt-work-and-codex-mtk4xc3fhttps://camcat.net/stories/proactive-cyber-defense-for-governments-and-enterprises-mtkba4sehttps://camcat.net/stories/the-hugging-face-incident-and-the-road-ahead-mtk4xc3ehttps://camcat.net/stories/b10773-mtkuxz16https://camcat.net/stories/learning-never-stops-how-ai-makes-learning-continuous-mtk4xc3dhttps://camcat.net/stories/研究-将探针输入投影到主成分子集-可显著提升-llama-3-1-欺骗检测的跨域泛化-mtnpjayjhttps://camcat.net/stories/ai-powered-robotics-in-europe-live-demonstrations-and-strategic--mtkqq84khttps://camcat.net/stories/v0-14-21-mtkqqd62https://camcat.net/stories/ernie-image-文生图模型发布-mtkqq74whttps://camcat.net/stories/we-re-launching-lyria-3-5-in-google-flow-music-with-advances-acr-mtk5fc5xhttps://camcat.net/stories/healthcare-organizations-can-now-connect-ehr-and-additional-indu-mtk4xc38https://camcat.net/stories/新框架a-cegis用反例作反馈-让ai代理在4轮内解决90-自然语言转正则任务-mtnoajqyhttps://camcat.net/stories/研究提出-规划即推断-的对偶平坦几何框架-统一线性与非线性奖励下的强化学习求解-mtno3omahttps://camcat.net/stories/人工代理可模拟享乐性位置偏好行为-研究指向意识物理基础-mtno30gmhttps://camcat.net/stories/ijcai-2025反事实路径规划竞赛-整数规划方案获质量第四-速度最快-mtnnwu0ehttps://camcat.net/stories/研究-ai内容标注应向-证据可视化-转移-验证断言密度比简单作者披露更能帮助用户辨别真假-mtnnvis8https://camcat.net/stories/github-详解-copilot-成本优化-缩短单个输出可能让整个任务更贵-mtnnrqz1https://camcat.net/stories/github-podcast-解读-ai-新术语-循环工程-小队-马具-爬山算法等-mtnnrcy6https://camcat.net/stories/github-copilot-应用支持并行运行多个-ai-代理-每个代理独立会话与-git-工作树-mtnnr0zxhttps://camcat.net/stories/can-llms-extract-architectural-design-decisions-from-source-code-mtnrn9f6https://camcat.net/stories/remember-and-reweight-enhancing-multi-agent-debate-with-experien-mtnr4ayxhttps://camcat.net/stories/when-users-don-t-ask-benchmarking-context-driven-memory-retrieva-mtnogiehhttps://camcat.net/stories/risk-and-anomaly-identification-for-distribution-network-optimal-mtnrnu3zhttps://camcat.net/stories/privacy-preserving-heterogeneous-multi-llm-federated-inference-f-mtnob2l3https://camcat.net/stories/draco-fine-grained-credit-assignment-with-dynamic-rubrics-for-lo-mtno96sjhttps://camcat.net/stories/transfiver-human-ai-co-inference-through-a-shared-editable-state-mtno25a9https://camcat.net/stories/研究-学习式分级码级联策略可将结构化噪声下的量子比特用量降低两个数量级-mtnvt5aphttps://camcat.net/stories/synthetic-semantic-supervision-for-contrastive-code-representati-mtnnwtzghttps://camcat.net/stories/研究-简化谱聚类算法-双社区sbm检测误差逼近信息论极限-mtnvsidhhttps://camcat.net/stories/kc-bench-a-dynamic-interactive-benchmark-for-evaluating-knowledg-mtnnwh43https://camcat.net/stories/hallupeer-a-taxonomy-driven-benchmark-for-detecting-hallucinatio-mtnnwdlzhttps://camcat.net/stories/openai-codex-发布-0-154-0-alpha-2-版本-mtnr4p4yhttps://camcat.net/stories/dalek-a-constructive-agent-machine-mtnnw5s7https://camcat.net/stories/growpage-on-demand-kv-budgeting-for-efficient-llm-reasoning-serv-mtnnvu6hhttps://camcat.net/stories/新核方法直接从轨迹数据学习哈密顿系统-数据稀缺时优于两步基线-mtnvn5ayhttps://camcat.net/stories/新半参数框架用于干预驱动场景下的反事实回归推断-mtnvn58khttps://camcat.net/stories/fresh-memory-stale-plans-dependency-scoped-validation-for-distri-mtnnsq0qhttps://camcat.net/stories/inkan-用截断幂形式替代-cox-de-boor-递归-提速-kan-前向传播-mtnvlukkhttps://camcat.net/stories/fixed-suffix-dependency-ratio-quantifying-the-dual-track-mechani-mtnssz2khttps://camcat.net/stories/duadeep-seqaffinity-仅用氨基酸序列预测抗体-抗原结合亲和力-无需三维结构-mtnv0y9chttps://camcat.net/stories/新联邦学习算法-byrd-nafl-结合-nesterov-动量与拜占庭容错聚合-兼顾收敛速度与抗攻击能力-mtnuwp7mhttps://camcat.net/stories/what-else-needs-fixing-exploring-cost-effective-test-time-comput-mtnrnsv8https://camcat.net/stories/the-analyst-in-the-prompt-role-retrieval-and-memory-biases-in-ll-mtnrjy1thttps://camcat.net/stories/研究人员提出large-language-gibbs-用llm条件分布迭代采样-替代单遍生成-mtnuj3yzhttps://camcat.net/stories/新推荐方法epic-在语义id去噪中引入物品级竞争-提升生成式推荐准确率-mtnueaxihttps://camcat.net/stories/研究提出短窗口滑动学习框架-1-2秒片段识别监控暴力事件-rwf-2000准确率95-25-mtnubhu4https://camcat.net/stories/综述-机器学习辅助定向进化五年进展有限-多数方法忽略dna合成成本-mtntn09jhttps://camcat.net/stories/研究-告知llm智能体其内部激活被监控-无法阻止其串通行为-mtntlpsahttps://camcat.net/stories/级联pinn建模srtio3-si忆阻器-避开条件数超10-16的数值刚性问题-mtnti4wphttps://camcat.net/stories/新方法agentauditor-把多智能体推理痕迹组织成树状再核对证据-准确率最高提升5个百分点-mtntahr2https://camcat.net/stories/coverpruner提出视觉令牌剪枝新框架-以覆盖优化取代-保留高分令牌-mtnt428bhttps://camcat.net/stories/新方法siser用对抗训练抑制说话人特征-iemocap情感识别ua达60-63-mtnt2l4whttps://camcat.net/stories/训练零门槛框架editvid统一指令式与参考式视频编辑-five准确率78-16-mtnsyrumhttps://camcat.net/stories/vlm驱动的sbs框架提升弱监督密集视频描述性能-mtnsycfrhttps://camcat.net/stories/rulemem-从历史对话归纳逻辑规则-提升长期对话问答准确率54-3-mtnsslrchttps://camcat.net/stories/研究-在上下文学习中加入谱系与地理邻居证据-llm-可显著提升类型学特征预测准确率-mtnsmd93https://camcat.net/stories/新研究用因果干预定位llm处理复数指代的注意力头-发现其偏好与人类一致-mtnsj22rhttps://camcat.net/stories/新方法marse用连续神经音频编码表示做语音增强-可灵活权衡性能与算力-mtns5withttps://camcat.net/stories/新研究提出psf-net-将相位信息纳入少样本细粒度图像分类-mtnrva9mhttps://camcat.net/stories/新框架sgd-kv按注意力头分配缓存预算-kv缓存内存占用最高降75-mtnrngv8https://camcat.net/stories/前沿llm在批处理优化中表现脆弱-但在语义丰富场景显著优于经典方法-mtnrawn5https://camcat.net/stories/新研究用自回归语音先验做测试时自适应-提升语音增强在噪声失配条件下的质量-mtnrac9whttps://camcat.net/stories/kick-off-of-new-genai-pilots-for-public-administration-apply-ai--mtkqq84lhttps://camcat.net/stories/新神经视频压缩方法结合可变形时间对齐与差异感知融合-性能优于dcvc-dc-mtnpe0nbhttps://camcat.net/stories/langchain-1-4-0-mtm9k010https://camcat.net/stories/release-v5-16-1-mtk4y1ivhttps://camcat.net/stories/langchain-1-4-0a3-mtkqqct2https://camcat.net/stories/arxiv论文-用smt扩展sat编码-为数值tohtn规划提供首个基准套件-mtno38tthttps://camcat.net/stories/building-an-adaptive-agentic-cybersecurity-system-with-nvidia-ne-mtkq8onfhttps://camcat.net/stories/v0-33-3-mtka96zehttps://camcat.net/stories/introducing-agentic-video-understanding-with-gemini-mtk5fc5phttps://camcat.net/stories/judging-llm-as-a-judge-concerning-rubric-artifacts-in-llm-based--mtnoak8lhttps://camcat.net/stories/新下界证明-梯度下降在任意时间设置下无法达到银级调度速率-mtnw5qe0https://camcat.net/stories/ichthyonoma-nomenclature-and-context-sensitivity-of-zero-shot-bi-mtntm5s4https://camcat.net/stories/新方法可在无真实处理效应情况下选优hte估计器并控制错误率-mtnvnndjhttps://camcat.net/stories/structure-and-implementation-of-new-practical-english-textbooks--mtnnrytmhttps://camcat.net/stories/karma-用知识图谱路径生成对齐对比样本-解决偏好优化分辨率失配-mtnv8k7dhttps://camcat.net/stories/研究提出dp-br-fedavg框架-为银行与医疗联邦学习同时抵御隐私攻击与恶意客户端-mtntnnsdhttps://camcat.net/stories/研究提出flashattention-4新路径-fp4下注意力前向吞吐最高达bf16的2-13倍-mtnt4x79https://camcat.net/stories/评论文章-llm与大脑表征对齐不足以证明二者共享计算机制-mtnrawlzhttps://camcat.net/stories/langchain-anthropic-1-7-1-mtm9k010https://camcat.net/stories/v0-25-1-mtka968nhttps://camcat.net/stories/lrconv-nerv-低秩卷积替代稠密-3-3-卷积-视频解码复杂度降-68-mtnujt7zhttps://camcat.net/stories/新元定理证明-每个有限句法系统都至少有一条自身无法生成的定理-mtnsqpzohttps://camcat.net/stories/eu-launches-ai-gigafactories-call-to-boost-europe-s-computing-ca-mtkqq84ihttps://camcat.net/stories/join-the-inaugural-d-tect-forum-mtkqq84ehttps://camcat.net/stories/commission-starts-enforcing-ai-act-rules-and-new-transparency-re-mtkqq84fhttps://camcat.net/stories/patch-release-v5-15-1-mtk4y1j0https://camcat.net/stories/onnx-runtime-发布-v1-28-2-补丁-修复-compile-api-模型序列化重复节点问题-mtnpjscshttps://camcat.net/stories/patch-release-v5-14-1-mtk4y1jdhttps://camcat.net/stories/pitching-day-finalists-of-the-apply-ai-startup-award-mtkqq848https://camcat.net/stories/we-re-introducing-flexible-usage-limits-for-gemini-notebook-mtlhq2z6https://camcat.net/stories/v0-5-15-post1-mtkqqceehttps://camcat.net/stories/patch-release-v5-13-1-mtk4y1jghttps://camcat.net/stories/state-of-the-union-2026-mtkqq846https://camcat.net/stories/build-a-measurement-stack-you-can-rely-on-to-steer-your-campaign-mtlbhq2dhttps://camcat.net/stories/openai-codex-发布-0-153-0-alpha-5-版本-mtnrbuqehttps://camcat.net/stories/openai-codex-发布-0-153-0-alpha-5-1-版本-mtnrb4g5https://camcat.net/stories/openai-codex-发布-0-153-0-alpha-6-版本-mtnraopbhttps://camcat.net/stories/openai-codex-发布-0-154-0-alpha-1-版本-mtnr68ushttps://camcat.net/stories/notice-secure-ai-infrastructure-call-for-information-mtkqq95thttps://camcat.net/stories/v0-27-1-mtka968chttps://camcat.net/stories/bringing-chatgpt-for-teachers-to-more-u-s-school-districts-mtk4xc3dhttps://camcat.net/stories/a-milestone-in-expanding-access-to-ai-mtk4xc3bhttps://camcat.net/stories/playco-cut-manual-fixes-50-prototyping-games-with-gpt-6-astra-mtm9jzzwhttps://camcat.net/stories/legora-reviewed-41-documents-in-minutes-with-gpt-6-astra-mtm9jzzwhttps://camcat.net/stories/openai-supports-california-s-bill-to-advance-youth-ai-safety-mtk4xc39https://camcat.net/stories/better-answers-broader-thinking-what-students-gain-from-chatgpt--mtk4xc3chttps://camcat.net/stories/disrupting-a-new-covert-influence-campaign-from-russia-mtk4xc3fhttps://camcat.net/stories/ernie-5-0-荣登-lmarena-文本榜国内第一-全面超越多款国际主流模型-mtkqq752https://camcat.net/stories/expanding-openai-s-presence-in-brazil-mtk4xc3dhttps://camcat.net/stories/supporting-thailand-s-next-generation-of-ai-startups-mtk4xc3chttps://camcat.net/stories/polimill-builds-japan-s-next-generation-public-ai-infrastructure-mtk4xc3ahttps://camcat.net/stories/google-deepmind-and-a24-announce-first-of-its-kind-research-part-mtk5fc61https://camcat.net/stories/accelerating-the-frontiers-of-scientific-discovery-google-s-40m--mtk5fc5yhttps://camcat.net/stories/try-google-pics-easy-image-creation-and-editing-in-google-worksp-mtlbhq2dhttps://camcat.net/stories/ask-a-scientist-how-do-researchers-use-ai-to-predict-a-cyclone-mtlbhq2ehttps://camcat.net/stories/paint-net-作者-借-claude-反向工程重写-18-万行-direct2d-实现在-wine-上运行-mtnqztzvhttps://camcat.net/stories/mit提出cw-net方法-帮助人类预测自动驾驶汽车何时会犯错-mtnqa0e8https://camcat.net/stories/mit-ibm实验室推动ai与量子技术从理论走向生产系统-mtnpjmnn