Chinese contentResQAdam：低显存量化感知训练优化器， optimizer 状态省 71%
研究团队发现 GaLore 这类低秩优化器做 2 bit 量化训练会掉点，于是提出 ResQAdam，优化器显存省七成多，速度几乎不掉。
Latest
Every published item, in order · 9845 items
Nov 9
Papers10:43
Chinese contentSSF-Net：用于PDE算子学习的空间-频谱融合网络用Mamba加Fourier做PDE求解，比attention基线误差更低、参数还更少，搞科学计算的同学可以看看架构设计。
Papers10:43
Chinese contentGMAdam：用任务级二阶矩调制改善 PINN 多任务损失不平衡问题训练 PINN 时总有一项损失压不住？这篇把 Adam 里各任务的贡献演化讲透了，还给出 GMAdam 这个即插即用的改法，实验里比 AdaTask 更准。
Papers10:42
Chinese contentGA-SRMC 提出几何感知改进，MCMC 采样 ESS 最高提升 6.8 倍MCMC 采样遇到各向异性目标就拉胯？这篇把 Fisher 信息从预条件子升级成排斥侧最优矩阵，各向异性场景 ESS 直接翻倍，理论保证也给全了。
Chinese contentPHANTOM-Bench：多智能体 LLM 系统中的'虚幻共识'测量研究
这篇 arXiv 论文做了个 PHANTOM-Bench，发现 9 个模型在多智能体讨论时会被模糊的同伴信号带偏，准确率最多掉 47.8 个点，做 Agent 系统的该看看。
Chinese contentAgent-Skill 工作流让 LLM 高效翻译并行科学代码
三个智能体分工修 CUDA、OpenMP、SYCL 代码翻译，正确率全面超基线还省了一半多的 token，做 HPC 的可以看看思路。
Chinese contentFlex-RSI 基准测试：衡量智能体的递归自我改进能力
团队给智能体自我改进做了个正经基准，executor 和 improver 分工迭代改自己的 harness，用固定预算对比闭源和开源模型，结论挺有料。
Chinese contentOPO：按样本自适应裁剪比率，改进 PPO 的固定裁剪策略
一篇 RL 论文，作者给 PPO 的固定裁剪换了玩法：每个 minibatch 里排序挑出极端比率，单独给自适应半径，MuJoCo 和 Atari 上跑赢了基线。
Chinese contentDRAFT-Bench 发布：评测前沿 Agent 在专业设计全流程中的可靠性
一个新的设计类 Agent 基准，Astra 全跑完但违规多，Claude 少做几个却更稳，论文本身还很克制地说明评测局限。
Tomorrow 08:00, the daily lands in your inboxTelegram channel or email. Same screening as this page, no ads.
Chinese contentStanford 发布 Glyph：用稀疏自编码器从临床数据中挖掘疾病标志物
Stanford 把稀疏自编码器用在临床影像和检验数据上，挖出了397个可解释的疾病相关概念，代码开源了，做医学AI的可以看看。
Chinese contentFRAME 基准：如何在保留旧格式的同时生成新报告
想给公司写季度报告或学术综述的可以看看，这论文把“照着旧版格式填新内容”做成了 120 道题的基准，还测出了开源模型和 GPT 这类闭源模型的差距在哪。
Chinese contentHomeLoop：基于 Isaac Sim 的家庭机器人-IoT 协同仿真基准发布
家庭机器人光聪明还不够，这个新基准专测它会不会瞎操作——六款 VLM 最高才 61.9 分，做机器人研究的可以关注。
Turtle Evolves AppWorld Tasks Into Harder Benchmarks
Saturated benchmarks lose their ability to separate models, and TURTLE offers a cheaper path than building new tests from scratch. Practitioners evaluating tool-use or agent models can apply the evolution pipeline to refresh existing benchmarks against their own target model, keeping failure signal alive. The 52.5% versus 19.8% hardness rating quantifies how much headroom task mutation adds.
Amazon paper finds tax and legal benchmarks misgrade correct AI answers
Benchmark scores on rule-following tasks may reflect grading bugs rather than model weakness, so leaderboard gaps deserve scrutiny before drawing conclusions. Teams evaluating models on tax, legal or compliance tasks should read scorer code and manually inspect a sample of failed answers before trusting low numbers. The paper's audit method, tracing references to source rules, offers a reusable checklist for anyone relying on public benchmarks.
Yann LeCun Clarifies the Difference Between JEPA and JEA
The real takeaway is a taxonomy: JEA methods like DINO and Barlow Twins suit symmetric view-matching, while JEPA's predictor targets corrupted-input or temporal prediction, with an action variable making it a world model. Researchers comparing self-supervised approaches or building world models get a concise map of which loss function fits which task.
Chinese content清华与 Moonshot AI 提出 TCSExam 基准：GPT-6 Astra 定理证明率达 70.0%
清华和 Moonshot AI 搞了个 TCSExam 测试集，GPT-6 Astra 证定理能拿 70.0%，但 30 个开放猜想全挂，差距很直观。
Evolver Reuses Agent Trajectories to Improve Task Success
The key implication is that harness-level memory can outperform parameter tuning: agents get better purely by reusing learned procedures, making this applicable to any existing model deployment. Practitioners building coding agents or evaluation pipelines can adopt the Gene-asset pattern to cut token costs. Terminal-Bench 2.1 and SWE-bench Multilingual results give concrete baselines for comparing against their own harness setups.
Papers00:36
Chinese contentdair-ai 每周精选：11月2日至8日七篇热门 AI 论文dair-ai 又更新本周论文榜了，7 篇里有一篇直接质疑长上下文是不是必需的，做研究的可以扫一眼名单。
Meta's ABTest Agent keeps LLMs away from A/B test numbers
The real lesson is architectural: reliability came from constraining the agent, not from a smarter model, with every number produced by validated code. Teams building analytics or experiment agents should pre-register success metrics so the agent cannot help p-hack, and route all computation through audited tools. Meta's design shows what enterprise-grade agent guardrails look like in practice.
Salesforce paper shows RL experts merged beat external teacher distillation
The finding flips a common assumption that dense teacher feedback is the compute-efficient path, showing the teacher's per-token forward pass overhead usually outweighs its benefit. Teams training domain-specific models can instead run RL per domain and merge experts, cutting compute by roughly 42%. Practitioners weighing teacher choice or FLOP budgets for math and code training get a concrete decision rule.
Nov 8
Tsinghua TraceMRT attack evades chain-of-thought monitors on SHADE-Arena
The finding suggests that faithfulness of chain-of-thought monitoring depends heavily on how attacks are framed, not just on the monitor's model. Teams deploying CoT monitors should red-team them with task-specific, goal-integrated attacks rather than generic templates, and score both whole traces and individual actions. The Bayesian scoring approach offers a concrete, replicable detection upgrade for safety engineers.
Shanghai Jiao Tong Builds AutoDNA Lab for Autonomous Molecular Biology Experiments
Most autonomous labs bolt AI onto pre-configured instruments; AutoDNA's model-experiment-instrument co-design instead makes the hardware natively callable by agents, a distinction that matters for anyone building embodied lab automation. Research groups running wet-lab pipelines or instrument vendors designing robot interfaces can study its atomic-service abstraction and multi-agent task decomposition. The Nature homepage coverage signals strong institutional endorsement.
IT之家Source
Chinese contentAnthropic 在 EVE 里做内鬼实验，AI 队友几乎抓不出叛徒
Anthropic 把 AI 放进 EVE 玩狼人杀，内鬼 AI 的骗人手法挺细的，队友全程没抓出来，看看它是怎么骗人的。
Chinese contentAnthropic 分析 20.7 万次科研会话：科学家用 Claude 主要做数据分析和写论文
Anthropic 统计了 20 多万次科研会话，发现科学家拿 Claude 画图跑统计最多，最想让他帮提想法反而用得最少，还估算了 AI 能让科研快几倍，数据很细。
KOSMOS Self-Scheduling Kernels Speed Up Training on AMD GPUs
Fixed-role fused kernels leave idle windows because a predetermined communication-compute split cannot track actual data readiness; KOSMOS replaces static schedules with demand-ordered queues that adapt at runtime. Engineers training large MoE models on AMD Instinct hardware can apply it directly through Megatron-LM, where it outperforms DeepEP by up to 1.83x on DeepSeek-V3 and Qwen3-235B proxies.
Chinese content研究显示 Anthropic 模型改进速度因 AI 参与训练提升至 1.9 倍
有团队算了笔账：Anthropic 靠 AI 参与训练，模型迭代速度变成原来的 1.9 倍，这个数字挺有说服力。
Chinese contentStanford 论文：更多信息会让部分 LLM 谈判智能体收益下降
Stanford 测了 10 个 LLM 谈判，发现给模型更多信息反而可能让 Claude Opus 5.5 收益从 94% 掉到 64%，选模型时得连信息量一起测。
Chinese contentGoogle 研究：AI 队友会“顶嘴”，学生写的商业计划书反而更好
Google 拿 Gemini 3.1 Pro 做了个会跟学生抬杠的 AI 队友，试验里比普通 AI 工具帮学生多拿了 7 分，做教育产品的可以看看。
Yann LeCun Explains Why JEPA Research Papers Differ From Meta Models
The thread lays out a taxonomy of SSL methods — predictor presence, anti-collapse criterion, augmentation scheme — that helps readers place any new 'JEPA' paper in the right context. Researchers deciding which papers to cite or reproduce will find it a concise map of what academic work can and cannot demonstrate. The core distinction between method discovery and resource-heavy scaling applies far beyond the JEPA family.
Oxford paper finds frontier LLM agents fail cross-vendor coordination
The practical implication is that agent interoperability breaks down not from capability gaps but from unstated conventions, so inter-agent protocols may need explicit shared norms rather than smarter models. Teams building multi-agent systems that span different vendors should add instructions telling agents to avoid unstated conventions and prefer actions any partner can read correctly. The finding applies to anyone chaining GPT-5.2 and Gemini 3 Pro agents together today.
Chinese content研究显示 AI 审计员能发现 98% 智能体不当行为，但误报过多
Anthropic 研究员测了 GPT-5 等模型审智能体日志的能力，98% 漏不掉但误报堆积，做智能体安全的可以看看
Chinese contentShadowBench 测评：6款AI模型无一能做出动物手影
有人做了个ShadowBench，让6个前沿模型用双手摆动物手影，结果全军覆没，连一个都摆不出来，很有意思的空间理解测试。
Chinese contentAI 2027 追踪器更新：Anthropic 内部模型进展超前于预测曲线
AI 2027 追踪器刚更新，用 METR 报告里的 AECI 分数一算，Anthropic 有些内部模型已经跑在预测曲线前面了。
Chinese contentAnthropic 研究：AI 可承担科研人员 27-52% 的工作时间
Anthropic 量化了 AI 能替科研人员干多少活：中位数 38%，按学科拆得很细，做科研的朋友可以看看自己领域排哪。
Nov 7
Google Paper: ReproBridge Lifts Agent Reproduction Success to 92.6%
The practical implication is that pipeline artifacts, not model choice, explain most reproduction failures — teams can capture setup steps from a first successful run instead of paying for a larger model. Builders running coding agents on research repos can apply this immediately via the arXiv paper (2611.08452) and the accompanying Chat with Paper tool from dair.ai. Compared with raw repo reruns, the packaged approach also lowers API spend by 19.7%.
Only the latest 2000 items are listed; for older ones see the daily archive