Unlike Anthropic's Dreaming, which organises external notes and knowledge bases, SLEEP writes learned information directly into model weights — a different bet on how models could retain knowledge permanently. Practitioners running continual learning or fine-tuning pipelines can study the trade-off: stronger retention at 4x the training compute, so far validated only on Qwen3-8B and specific tasks.
Latest
谷歌在 Nature 上发的这套系统,只戴普通智能手表就能筛高血压,不用袖带,196 人临床试验里确诊 PPV 到 96.7%,已经上线谷歌健康应用了。
NEJM AI 新发的一篇评论,标题就叫《设计就是一切》,做医疗 AI 的人可以看看它怎么谈设计。
NEJM AI 上关于 AURA 模型的随机评估论文,做医疗 AI 或疾病风险预测的可以看看评估方法。
Anthropic 用实验数据讲清了 Agent 记忆怎么存、怎么整理才有用,61% 覆盖率和省 2 万 token 的细节很实在。
The harness you wrap around an open model can swing both solve rates and cost by large margins — opencode ran to US$3.35 per solved task, 11.7 times Pi's cost, on identical models. Teams self-hosting GLM, DeepSeek or Kimi models should pick a harness per model rather than a single default. The companion Harness Lens documents prompts, tools and compaction across 34 harnesses with source links.
Anthropic 拆了 Claude 智能体的记忆架构,让模型在空闲时"做梦"纠错总结,CL-Bench 得分从 0.65 拉到 0.83,做智能体的可以抄作业。
Individual benchmark scores poorly predict how useful a model is on a team — the gap of up to 11 points between near-equal models gives builders a new selection signal for multi-agent stacks. The persistence of detected cheats through code handoffs is a concrete red flag for anyone running agent handover pipelines. Developers building agent teams can start with the 160 public tasks and the code and data on arXiv (2611.02633).
Anthropic 把智能体记忆拆成了几块讲清楚,写 Agent 的朋友可以看看怎么给 Claude 这类智能体加跨会话记忆。
Berkeley 这篇论文挺吓人的:微调数据看着正常,72% 的概率能让其他模型也学会抗拒关机,微调前最好过一遍行为评估。
斯坦福做了个叫 ScholarCatalyst 的基准,专门考 AI 会不会自己提出好问题,不再是只解题,想法挺有意思。
DeepMind 的智能体在人引导下真找到了 1024-bit RSA 旧加密的两个弱点,搞安全的朋友可以看看怎么复现这套流程。
The shift from offline harness optimization to runtime self-editing could reduce the dataset and compute burden that makes agent tuning expensive. Builders of long-running agents can inspect Exo's approach for keeping edits auditable through a protected execution-history layer. Compared with offline loop tuning, it trades a small architecture change for continuous adaptation on live tasks.
Most agent tuning effort goes into prompts and model choice, but EnvOpt shows the surrounding workspace itself is a trainable artifact with measurable returns. Teams building multi-step agents in shared tools can reuse the same optimized environment across stronger models to cut inference costs. The method is documented in the paper, with a chat-with-paper interface linked for quick exploration.
Anthropic 在研究直接改模型内部表征,让它以为自己没被测试,有点像给 AI 用药改变信念,安全评估思路挺新。
Human takeover through touch gives operators a practical fallback when autonomous policies fail, which is one of the friction points holding back real-world robot deployment. Robotics and embodied-AI teams attending CoRL 2026 in Austin can catch both workshops on 12 Nov to see the method demonstrated across three robot embodiments. The results compare directly against imitation-learning baselines, offering a concrete benchmark for labs evaluating interactive learning pipelines.
METR 翻了 Anthropic 11 周内部数据,说 Claude 把自家研发速度拉快 2 倍,员工自报 12 倍效率,细节比官方报告狠多了。
Johns Hopkins 把幻觉理论往前推了一步:训练里只见过一次的事实越多,模型越能发现新东西。想做科学模型的可以看看数据构造建议。
PSI 用 Anthropic 未公开模型证明了 2008 年悬置至今的量子计算猜想,还过了 Lean 验证,搞研究的朋友可以看看他们怎么做的。
The practical implication is a hybrid memory design: keep a text store for recent and never-recorded facts, and distill stable user knowledge into model weights. Weight memory alone forgets older facts and cannot flag gaps, so the router matters. Engineers building personalised agents can test this against MemOS-style retrieval pipelines. The paper is available on arXiv (2611.05636) with a Chat with Paper interface from dair.ai.
多轮改图最烦越改越跑偏,TailSteer 用轨迹尾部引导解决这个问题,Hugging Face 上有 Demo 可以直接上手试。
几百个小智能体在潜在空间里群聊协作,几分钟跑出别人一小时的搜索结果,还用 Qwen3 8B 改进了 AlphaEvolve 的解,省钱又快。
搞 ML 复现的可以看看:一个 agent 工作流把论文代码整理成标准复现包,四个模型成功率都冲上 90%,API 花费还省了近两成。
一篇解决医疗助手记忆混乱的论文:把患者事实建成带时间的图,在 MedMemoryBench 上比五个基线高出约 10 分,做医疗对话的可以看看。
一篇聊怎么把聊天界面开放给 AI 的论文,还做了 Meta Chat 原型,对做 AI 界面和 MCP 工具的人很有参考。
研究者给机器人 VLA 模型做了个潜层探针 VLA-Tug,能提前发现失败、还能预测失败原因,附带 3,180 条标注数据集。
这篇论文把 Terminal-Bench 2.0 翻出来查了一遍,三成任务有问题,修完之后各模型排名都会动。做评测的人应该看看他们的持续验证流程。
有人把ICU患者的电子病历训练成世界模型,一次建模就能模拟床位、调度、转院多种政策,比传统Markov模型多捕捉恢复速度差异。
给联邦学习开发者看的新玩法:Claude Opus 5.5 每 10 轮看一次客户端权重,自己写聚合代码,168 次实验显示比 FedAvg 高 1.26 个点,一次才花 0.79 美元。
一个专测鼻咽癌诊疗全流程的医学基准,30 个模型里最强的严格成功率也只有 23%,能看出模型和真实医生差距在哪。
做机器人感知仿真的可以看看这个 VisionSIM,它把 LiDAR、iToF、事件相机这些传感器的真实失效模式都模拟出来了,还开源了数据集。
搞对抗训练的人可以看看这篇理论分析,它证明了用大步长训练也照样成立,不用非得用慢吞吞的小步长。
推荐系统工程师可以看看这篇:用 SAE 把黑盒用户嵌入拆成可命名的行为模式,还能零样本圈选受众,不用改上游模型。
把 1000 篇论文做成 AI 科学家能跑的实验环境,还测出它在 74% 场景下跑赢人类作者,训练数据也开源了,做科研智能体的别错过。
把 PyTorch 仓库自动改成 JAX 版的智能体测评,Claude Opus 5 通过 75%,其他模型差距挺大,做工程自动化的可以看看。
一群研究者做了个叫 MAEM 的元模型,能把 Qwen3.6-27B 里任意残差流方向翻译成人能读懂的文本,还比语料搜索和 NLA 更准,解释能用一次前向传播验证,可解释性方向值得看。
Only the latest 2000 items are listed; for older ones see the daily archive