Latest

Every published item, in order · 9845 items
Nov 11
Chinese content论文提出 SAT:让强模型先诊断再教学弱模型 Agent

这篇论文有点意思:Claude 教 DeepSeek 时不是直接写攻略,而是先跑诊断探针找出弱模型卡在哪,再针对性写技能,成功率从 75% 拉到 86%。给做 Agent 框架和多人协作系统的开发者提供了可复用的技能准入机制思路。

arXiv: DeepSeek@Chang Liu, Bo WuSource
Chinese content论文实测:Odia 文本在 GPT-4o 上的 token 消耗是英文的 5 倍

印度团队实测了 9 个主流 tokenizer 处理 25 种印度语言的开销,Odia 在 GPT-4o 上要多花 5 倍 token,Sarvam 的 tokenizer 差距最小。做多语言客服或按 token 预算的团队,选模型前值得查一下自己的语言在哪个 tokenizer 上最省。

Event
arXiv: DeepSeek@Johin Johny Arimbur et al. (3)Source
Papers10:45
Chinese content条件熵正则化的 Evidential Deep Learning 用于青光眼检测

做医学影像或不确定性建模的值得看:作者把条件熵加进 EDL 正则项,在 RIM-ONE-DL 等数据集上校准误差比 MC Dropout 和 Deep Ensembles 都低,还能用 Grad-CAM 看模型是否盯错了视盘。

arXiv cs.LG@Hai Siong Tan, Kuancheng WangSource
Chinese content论文对比 Alibi-Detect、NannyML、PyTorch-OOD 检测遥感影像分布偏移

无人机上跑模型的团队值得看看这篇:它实测了 Alibi-Detect、NannyML、PyTorch-OOD 三个开源工具,不只看能不能检测出偏移,还测了耗时,这俩在机载设备上经常冲突。做模型监控选型前可以当对比清单用。

arXiv cs.LG@Abdullah Almansour, Ozan K. TonguzSource
Chinese content研究揭示 Transformer 组合泛化评估中的等价性与模块覆盖问题

这项研究解释了为什么不同论文测组合泛化结论互相矛盾:很多基准里的模块序列其实功能等价,模型是在走捷径。做基准设计或评估模型泛化能力的开发者应该看看,结论会直接影响你怎么切训练测试集。

arXiv cs.LG@Purva Pruthi et al. (4)Source
Chinese contentBradley-Terry 模型下的差分隐私排名统计推断方法

做 LLM 评测榜单的人可以看看这篇:在 Bradley-Terry 模型上加边级差分隐私,不只给排名,还给每个名次的置信区间和 top-K 检验,模拟实验和创意写作排名案例都有。要保护人类评委偏好数据时很实用。

arXiv cs.LG@Xintao Xia, Jia GuSource
Chinese contentCoco 提出深度个性化:桌面主动智能体的本地上下文观察方案

做主动式智能体的开发者可以看看这篇。Coco 用独立的 Observer Agent 在本地观察桌面行为,判断什么时候帮你、什么时候别打扰,介入判断的 F1 提升了 2.4-3.0 倍。数据不出设备这点对企业场景很关键,还附带 OSWorld-Persona 基准可直接用来测自己的方案。

arXiv cs.LG@Yijia Shao et al. (8)Source
Tomorrow 08:00, the daily lands in your inboxTelegram channel or email. Same screening as this page, no ads.
Chinese contentSLIC:用强化微调 4B 小模型做多轮模拟电路优化

模拟电路调参不再需要云端大模型:4B 小模型本地跑仿真闭环,Pass@1 反超 DeepSeek 的 763B 模型近 49 个百分点,还能跨工艺节点迁移。做 EDA 工具或芯片设计的工程师值得把它加进工具链,论文附带 41 个拓扑的开源基准可直接复现。

arXiv cs.LG@Yan Xu et al. (9)Source
Papers10:41
Chinese content研究:核机器无需特征学习也能完成 in-context learning

这篇论文用核机器的谱分解把 ICL 的泛化和记忆讲成同一解的不同截断,思路很新颖。它还证明没有特征学习的模型也能做 ICL,并用高阶统计量解释了 transformer 的记忆行为。做理论研究和理解 ICL 机制的研究者应该看看。

arXiv cs.LG@Joseph Turnbull, Jamie SimonSource
Chinese contentParticle解码免费估计序列级Rényi熵,自适应推理降低算力开销

做推理部署的可以看看:SMC 解码顺手就能算出序列级熵,不用额外跑补全。论文在 MATH500 上验证过,两遍策略用少 34.2% 的粒子就能保住 64 粒子的准确率,算力紧张的团队直接受益。

arXiv cs.LG@Erfan Baghaei Potraghloo et al. (9)Source
PapersBuildersSelected10:41
SlackTier Moves DeepSeek-V4.1 Engram Tables Off Host DRAM

The non-obvious insight is that per-layer prefetch slack, not raw bandwidth, decides whether lookup tables fit on NVMe, DRAM, or remote memory — static hot sets transfer poorly for agent traffic. Inference engineers running DeepSeek-V4.1 or similar n-gram-augmented models on SGLang can reclaim roughly 180 GiB of host memory and convert it into KV cache, directly lifting throughput. Compared with simply scaling up DRAM, the RDMA option lets one table copy serve two instances, cutting per-node memory cost.

arXiv cs.LG@Wen Huang, Jack LiuSource
Chinese content论文揭示前缀缓存的智能体服务存在亚稳态:一次突发可致吞吐腰斩70分钟

跑过 vLLM 智能体服务的人应该都见过那种莫名的性能崩塌,这篇论文把它解释清楚了:一次几小时的突发流量就能让命中率掉到 10%,吞吐腰斩还自己好不了。论文给了能用的预算器方案,不用调参,直接把 KV 池撑爆的可能性掐掉。

arXiv cs.LG@Wen Huang, Jack LiuSource
Papers10:40
Chinese contentPBPK模拟揭示生成式药物设计的剂量暴露盲区

做 AI 制药或分子生成的朋友可以看看,这篇用 PBPK 模拟给生成管线的筛选流程补了剂量暴露这一环,60 个 PARP1 候选里近三分之一会在这一步被筛掉或翻盘。如果你只依赖预测效力和毒性打分来选分子,这篇会改变你的评估思路。

arXiv cs.LG@Maria Franz, N. Benjamin ErichsonSource
Chinese contentMicrosoft Research 发现:休眠神经元可被重编程为 AI Agent 后门

Microsoft Research 指出 Agent 里少量休眠神经元就能埋后门,触发前一切正常,很难察觉。防御反而老套:管好模型来源和权限。自己在企业里部署 Hugging Face 开源模型的团队,这篇值得当供应链安全检查清单用。

Event
Official accountMicrosoft Research@MSFTResearchSource
PapersBuildersOfficial07:00
Chinese contentAgentCore Evaluations 评测手册:内置、DeepEval 与 AutoEval 三种方案

AWS 这篇手册把智能体评测拆成了任务完成、跑题、数据泄露这些具体维度,教你怎么在 AgentCore 上选对评测器。写智能体的开发者可以直接照着搭回归测试,不用再靠肉眼验收。

Primary sourceAWS Machine Learning Blog@Bharathi SrinivasanSource
PapersBuildersSelected02:53
LingBot-Touch: tactile foundation model transfers across robot sensors

A frozen tactile model matching a sensor-specific adapter is the notable result — it suggests touch data can become reusable infrastructure rather than per-robot sensor data, similar to what ImageNet-style pretraining did for vision. Robotics teams building dexterous hands should evaluate it as a drop-in feature extractor before collecting sensor-specific datasets. Weights and code are open, so benchmarks can be replicated directly on any tactile hardware.

elvis@omarsar0Source
PapersBuildersSelected01:44
Berkeley study finds model routing saves 8.82X cost per quality

The gap between the theoretical 5.4-point router ceiling and the 0.3-point real-world gain reframes routing as a cost tool, not a quality tool. Teams building LLM applications, especially those balancing API budgets against output quality, can use the 25.5X coding-efficiency figure to justify routing to cheaper models for routine requests. The cross-domain benchmark coverage beyond language models makes the findings useful for robotics and multimodal pipelines too.

rohanpaul_ai@rohanpaul_aiSource
PapersBuildersSelected01:19
UC Berkeley's Agency Cuts Agent Harness Costs by Up to 35%

The practical shift here is that cost optimisation moves from model selection to infrastructure — a middleware layer can be reused across Claude Code, Codex and other harnesses. Engineering leads running production agents should audit trace logs for repeated context reads and unnecessary tool calls before upgrading to a costlier model. Teams in cost-sensitive markets, including ASEAN startups paying USD-denominated API bills, gain a lever that compounds across every harness they deploy.

DAIR.AI@dair_aiSource
PapersBuildersSelected01:18
InfraRSI Lifts LLM Training MFU from 17.4% to 30.0%

The non-obvious takeaway is that the hard part of agent-driven infra optimisation is verification, not code generation — the contract-and-recheck design is what makes the 2.75x speedups real. ML engineers and platform teams running multi-GPU training can apply the same contract-based evaluation to their own agent workflows. The cross-GPU memory of learned models suggests reusable optimisation knowledge across heterogeneous clusters.

DAIR.AI@dair_aiSource
Chinese contentAI Futures Project 更新时间线:超级人类 AI 研究员或于 2027 年出现

AI 2027 那份预测报告的作者们把时间线提前了 5 个月,认为 2027 年就会出现能自己做 AI 研究的超级系统。不管你信不信这个预测,这份报告此前引发了大量讨论,看看他们为什么改主意挺有意思。

Event
Andrew Curran@AndrewCurran_Source
Chinese contentStanford AI100 发布第三份报告《AI Goes Mainstream》,回顾 2021 年以来 AI 变化

Stanford AI100 这个百年研究计划刚出了第三份报告,专门梳理 2021 年以来生成式 AI 怎么一步步走进科研和日常生活。跟前两份比,这版把生成式 AI 的规模化当成主线来写,做政策或研究的人可以直接拿来当参考。

Stanford HAI@StanfordHAISource
PapersBuildersMulti-sourceSelected00:52
Nvidia paper identifies intrusion heads behind RAG distraction

The non-obvious finding is that distractor damage is localised in a handful of attention heads, so a lightweight patch can match most of what a costly retraining effort would achieve. Teams running RAG or agent systems on open-weight models can implement the mask as an inference-time intervention without touching training pipelines. Compared with context-cleaning heuristics that filter noisy documents before inference, head masking operates inside the model and needs only a one-time identification of the offending heads per model. Practitioners in ASEAN serving models under cost constraints benefit because the fix adds no training expense.

Event
rohanpaul_ai@rohanpaul_aiSource
Chinese contentApple 论文:Claude Opus 5.5 直接控制机械臂,零样本成绩超 VLA 专用模型

苹果没训练任何机器人模型,只把 Claude Opus 5.5 接上摄像头和机械臂,RoboLab-120 上零样本就拿到 85.6%,比专门训过的 VLA 高了近一倍。做机器人或具身智能的值得看看这个思路,缺点是实时性还跟不上。

rohanpaul_ai@rohanpaul_aiSource
PapersBuildersMulti-source00:32
Chinese contentJack Clark 推荐一类超细分场景的 AI 应用论文

Jack Clark 分享了他爱看的一类 AI 论文:专攻某个极窄场景的应用研究,像 CucumberVision at the Edge 这种标题就很妙。想做 AI 落地、关心模型在真实环境里会卡在哪的人,可以顺着这个思路找论文看。

Event
Jack Clark@jackclarkSFSource
Nov 10
PapersBuildersSelected23:24
Microsoft harness lifts Llama3-8B long-context reasoning without training

The key insight is that evaluation and inference-time scaffolding can substitute for expensive data synthesis and fine-tuning on long-context tasks. Teams serving small open models like Llama3-8B or Qwen2.5-3B can apply the approach to extend context handling at under half the cost of training-based methods like NExtLong or PolicyLong. Engineers building local or low-budget deployments should test harness transfer before commissioning post-training runs.

Event
elvis@omarsar0Source

Only the latest 2000 items are listed; for older ones see the daily archive