Latest
统计学者看过来：这篇把随机森林思路搬进多状态生存分析，不依赖 Markov 假设也能算转移率，还有现成 R 包 JumpForests 可以直接上手。
一个 6B 激活 0.5B 的 MoE 模型，跑循环层数就能切 5 档推理深度，小预算也能拿到 89% 的深度增益，挺实用的思路。
物理版科研能力测试题，63 道复现加 43 道发现题，最强模型闭卷也只有七成多正确率，还有合成数据能白嫖涨 11 个点。
A 12x reduction in training data to beat a frontier SSL checkpoint suggests the data appetite of self-supervised pretraining is not fixed, which matters for teams with limited compute budgets. Researchers doing vision pretraining or fine-tuning DINO-family backbones should review the JEM method and regularization design for reuse in their own pipelines. Compared with DINOv3, it trades scale for a simpler, more principled objective.
这篇论文问了个挺实际的问题：药物设计 Agent 花对接算力学到的经验，能不能存下来给下一个蛋白口袋用。答案是能复用的部分主要是编辑动作本身，不是靶点信息。
给做智能体的朋友：这篇测了 MCP 工具挂掉时各家 agent 框架的表现，结论挺反直觉，模型变大不顶用，框架的错误处理才关键。
一篇无线通信基础模型的研究：把信道配置信息放进预训练，在 LoS 分类和 beam selection 等任务上，小样本场景效果比基线更好。
评测 GUI 智能体老盯着静态截图，这篇造了 601 个目标先藏起来的任务，人能拿 94.6%，GPT-5.5 只有 79.4%，差距到底卡在哪写得很清楚。
有人拿 BERT 分析了奥地利议会 22 年的演讲，发现联合执政的政党说话越来越像，退党议员也是换了党才改口风。
NVIDIA 把 Audio Flamingo Next 微调成能模仿具体真人的音乐品味和说话风格，10 个 persona 测试有效，做个性化 AI 的可以看看。
一个不用训练就能做视频异常检测的框架 SAAF，比之前的零样本方法快两倍多，还能超过监督模型，代码已开源。
同一套任务，换不同 harness 排名就变。这篇用 52,168 条轨迹告诉你怎么区分模型和 harness 谁在背锅，做智能体的都该看看。
arXiv 上新论文 CASE，专门治多模态模型分不清相近情绪的问题，把 Qwen3-VL-8B 的 CA-MER 准确率从 46.80% 拉到 59.13%，还不用动模型参数。
一篇做多模态嵌入的论文，把推理当成隐变量来训练，MMEB-V2 上 78 个任务全赢了基线，视觉文档检索涨了 6 分多，做检索的可以看看。
让 LLM 照着护士用的分诊指南一步步执行，566 个场景测出 0.95 准确率，直接问 LLM 只有 0.47，过度分诊也从 0.5 压到 0.02。
做 Web3 安全评测的可以看看：这套路子用协议变异生成新任务，照抄链上公开攻击代码没用，3035 个任务还覆盖检测到修补全流程。
一篇医疗预测的论文：DIGOM 用图同时建模患者间和跨模态关系，在 MIMIC-IV 等数据集五个任务上都拿了最高 AUROC，模态缺失时也更稳。
训练科研智能体最缺的就是能跑、能验证的环境，这套流水线从代码仓库自动造出近 10 万个，还把 Qwen3.5 训出了跨基准提升。
The striking implication is that an AI lab produced mathematics that leading PDE researchers describe as possibly beyond what a human would have devised, with the mechanism lying outside prior literature. Researchers in mathematical fluid dynamics and anyone tracking AI's role in formal proof work should watch for the expert digests Armstrong expects, then compare the construction against Elgindi's blow-up results and the Albritton–Brué–Colombo framework.
有位研究者聊了聊她读 Anthropic 机制可解释性论文的真实感受，还吐槽了里面拟人化的 belief 写法，观点挺直接。
dair-ai 每周论文清单来了，7 篇都围绕智能体，从 Gemini 打 Minecraft 到缓存工程，做 Agent 的可以按名字去搜原文
Swapping LLM calls for a lightweight classifier in memory pipelines cuts latency by an order of magnitude without losing answer quality, a design pattern applicable to any RAG or agent system that overuses the LLM for trivial choices. Builders working on chat memory, retrieval routing or stopping criteria can replicate this by scoring candidates in batches and reserving the LLM for final generation. The numbers also set a concrete benchmark point against LoCoMo baselines.
Stanford 和伯克利搞了个 TokenGate，往模型里一插就能把长文档 prefill 快最多 11.3 倍，内存还省 3 倍多，跑长上下文的可以看看。
Google 研究员在 EMNLP2026 摊位现场演示一套评估 AI 潜在能力的方法，还把真实用户查询失败案例拿来 mapping，做评测的可以去看看。
Nathan Lambert 转了一条讲闭源实验室 RL 训练细节的推文串，OpenAI、Anthropic 内部怎么做强化学习，这里能挖到不少线索。
GPT-6-Astra 在新基准里解出 49/50 道研究级物理题，每道不到 3 美元，还能对比联网和不联网的差距。
GEO 方法都想让网页被 AI 引用，但各家测试规则不同，分数根本没法比。这篇论文给出统一的打标签框架，做内容优化的人可以照着核对。
DeepMind 跟了 5 个月、20 多个团队的真实数据，同一个智能体有人爱有人怕，结论很落地：部署前先征得同意。
Scoring bias in judge models flows directly into training signal and leaderboard rankings, so skewed calibration quietly shapes which outputs get rewarded. Teams using MLLM-as-a-judge pipelines should audit scale usage and split alignment reports by reference culture rather than averaging. Anyone building evaluation or RLHF infrastructure for products serving both Western and Chinese audiences has a concrete methodology to copy here.
Fixed macro execution works only when the skill matches the task; making granularity adaptive turns brittle agent failures into recoverable ones. Developers building coding agents can apply the macro-to-micro fallback pattern to improve reliability without hand-tuning heuristics. The learned router beating a hand-written heuristic on the 12-task study shows the approach scales beyond scripted workflows.
Cost and privacy trade-offs in agent execution do not move together — once task completion converges, strategy choice is mostly a privacy decision, and the best-scoring setup is also the most exposed. Teams deploying on-device agents can pick a collaboration mode based on edge model size: route execution to the cloud for small models, use cloud advisory for 9B-class ones. The comparison against Tsinghua-affiliated GLM-5.3 makes the findings directly applicable to Chinese open-source model stacks.
Only the latest 2000 items are listed; for older ones see the daily archive