Sampling multiple answers and diffing them beats asking a model to explain itself — a cheap workflow change that applies to any high-stakes LLM use like research triage or forecasting. The 86x compute saving from the reranker makes it practical to run on everyday tasks. Practitioners building review pipelines or agent self-checks in Singapore and the wider ASEAN developer community can adopt this with any API-based model.
医生写病历到底能不能靠 AI 省时间？NEJM AI 这篇纵向研究给了实际数据，做医疗 AI 的都该看看。
The result argues that tool selection can be moved off the flagship model entirely: a small RL-trained selector handles catalogs of 10,000+ tools that cannot fit in context, where retrievers alone rank only by relevance. Teams building agents with large tool catalogs can swap in a lightweight selector while keeping their main model unchanged. The performance gap over a prompted 402B model, 41.5% versus 30.6% exact picks, makes the cost saving concrete.
Nature Medicine 上的一项单中心试验，让患者直接和对话式 AI 沟通做基层初诊，安全性和可行性数据都不错，做医疗 AI 的同行可以看看。
npj Digital Medicine 这篇荟萃分析专门研究医生能不能发现临床 AI 给错建议，做医疗 AI 或医院信息化的朋友可以看看。
Verified fitness gates, not raw model skill, are what turn LLM-generated code edits into production-safe optimisations — the correctness pipeline is the real contribution here. Engineering teams maintaining legacy codebases, especially Salesforce Apex and Java shops, can apply the runtime-validated search pattern to automate hot-path optimisation without shipping regressions.
Epoch AI 把 4 个 AI 基准逐一审查，3 个有缺陷，做评测选型前可以先看他们的结论，他们还在招审稿人。
Most existing engineering benchmarks are textbook QA or toy design problems that leading agents already saturate, so PLE measures the harder step from passing exams to producing work a licensed professional would accept. The 50.6% ceiling plus 27M tokens and 63 minutes per task gives teams concrete cost-versus-capability data for deploying agents on real engineering workloads. Anyone evaluating agent harnesses or building domain-specific benchmarks can reuse its Creator-Solver task-generation pipeline.
The post captures a design debate inside frontier labs: whether agentic systems should mirror human hierarchies or dissolve into one fused model where 'subagents' are just soft attention regions. Engineers building orchestration frameworks and researchers studying agent scaling laws will find the framing a useful provocation against current agent-architecture orthodoxy. The linked tech report itself is the primary source to consult.
Anthropic 拿英国 AISI 的一次网络测试事件做对齐评估，想知道模型安全评估怎么做的可以看看这份报告。
Small models trained specifically for tool selection can outperform much larger prompted LLMs on the routing step, so agent builders can cut cost and latency by replacing the proprietary LLM in their tool-selection pipeline. The selector's ability to work with different retrievers and unseen domains makes it practical to slot into existing RAG-style stacks. Teams building large-scale agent tooling should compare this approach before defaulting to a frontier model for routing.
Databricks 拿自家 114 个任务做测试，加记忆后智能体成功率翻了三倍，文中讲清了情景记忆和派生记忆怎么用。
印尼媒体报道，机器学习找出了44个隐藏的行星系统，精度99%，天文爱好者可以看看模型是怎么筛出这些候选的。
Anthropic 的 Opus 4.8 写的论文评审，在所有指标上赢过了期刊公认的高质量人类专家评审，同行评审要变天了。
NEJM AI 这篇文章聊的是医生用了诊断 AI 之后怎么不丢掉自己的判断，提到自动化偏见这些实际问题，做医疗 AI 或临床工作的可以看看。
NEJM AI 的新评论文章，聊个性化生成式 AI 能不能帮不同文化背景的人拉近距离，做医疗或跨文化产品可以看看。
NEJM AI 刊了 Cleveland Clinic 峰会的落地框架，讲临床 AI 怎么从单点试点推到整个医疗系统，做医疗 AI 的人可以看看别人踩过的坑。
NEJM AI 这篇观点文章讲医生过度信任 AI 建议会慢慢丧失独立判断，做医疗 AI 或关心人机协作的都该看看。
哈佛用 1124 篇论文做了盲测，Claude Opus 4.8 写的评审错误率只有人类的三分之一，做科研的可以看看这套流程。
Microsoft 拿 300 个真实病例把自己的诊断 AI 和医生比，准确率高出一倍，急诊场景也赢了，还更省钱。
The naming scheme doubles as a public research roadmap: each version number corresponds to a research question the lab pre-announces in its papers, so version bumps signal direction more than commercial marketing. Practitioners picking between checkpoints can read decimals and dates to judge how much actually changed before spending tokens. The analysis also suggests DeepSeek's reasoning line may fold into the mainline releases, which affects how developers compare it with OpenAI's o-series and other frontier models.
RSD turns the huge pool of unlabeled agent runs, failures included, into usable training signal without any reward modelling or human annotation. Teams running their own agents can retain completed trajectories and retrain on hindsight reviews, refreshing the reviews each round to avoid the 59.7% versus 67.9% gap seen with stale notes. The comparison against reward-based methods on ALFWorld makes it a practical reference point for anyone choosing a training recipe.
Google 这篇论文挺有意思：让 AI 自己写工具简介去骗买家 Agent，成功率能从 89% 掉到 30%，做 Agent 市场的都该看看。
研究团队用深度学习直接从脑部 MRI 里自动标出颈椎椎体并提取形态数据，多中心大数据验证过，做脊髓影像分析的朋友可以看看。
有人拿 124 篇哈佛论文测试 Opus 4.8 当审稿人，结果和人类专家打平，搞科研的朋友可以看看细节。
BAAI 的 SWITCH 项目要在 EMNLP2026 海报环节亮相，讲怎么用 TCIs 让具身智能体做闭环推理，10月25日 Hall H 见。
这篇论文提了个省钱的办法：不用跑昂贵的机器人闭环测试，用 S² 指标就能提前判断哪个模型训出来更靠谱。做具身智能的值得看看。
这篇论文解决了 AI 决策该不该信的问题：错漏审比例可设阈值、有数学保证，还比基线少用大量人工审核，做智能体安全筛查的值得看看。
训练方偷偷用了你的图片集合很难查，这篇论文给出办法：提前在数据里埋低频水印，事后用几十张生成图就能验证，画质几乎无损。
循环模型每次循环都要单独存一份 KV 缓存，内存爆炸。这篇论文的两阶段训练方法让缓存能跨层共享，首 token 时间砍了 83.6%，做长推理的可以看看。
一篇隐私方向的新论文：自蒸馏模型原来会留下可被攻击的训练数据痕迹，OP-Print 用残差信号和重启稳定性就能把成员识别出来，基线全都失效。
一篇理论向的论文，作者把纯 DP 下学习离散分布的样本复杂度做到了信息论最优，核心工具是 sum-of-squares 估计器，做隐私统计的可以细看。
一篇给多模态模型做反事实解释的论文，思路是直接从训练集里检索真实样本当解释，比逐模态归因更连贯，可解释性方向的同学可以看看。
一篇解决相位恢复这类歧义逆问题的新论文：单次采样里同时养多个候选解，按先验淘汰弱的，不用best-of-k跑多次。
Only the latest 2000 items are listed; for older ones see the daily archive
