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.
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.
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。
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。
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。
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。
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.
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.
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。
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。
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。
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%.
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.
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.
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).
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。
微软研究员在 Science 上写了篇文章，把神经网络当成改造社会制度的模板，还起了个新词叫反向对齐，思路挺新颖。
Stanford 出了个新基准 ScholarCatalyst，专门测检索系统在科研里能不能像人类学者一样“直觉”判断哪个方向值得挖，做 RAG 或文献检索的可以看看。
METR 拿 Anthropic 内部数据做了份报告，说研发提速 80% 靠 AI，但有人认为真正加速发生在实验室外面的生态里，两边观点都值得看看。
a16z 给了组数据：医疗勒索攻击 6 年涨了 6 倍，停机从 4 天拖到 18 天，还聊了 AI 在其中的作用。
METR 用测试数据算了下，发现 Anthropic 模型变强的速度快了一倍，主因可能就是拿 AI 来研发 AI。报告细节挺多，连内部员工自评效率提升 12 倍都写了。
BPB has been treated as a neutral yardstick in scaling papers, so correlations this weak undermine cross-tokenizer comparisons that many labs rely on. Practitioners choosing a vocabulary size should benchmark task accuracy on the same text rather than trusting a single loss-based score. The finding matters most to researchers running tokenizer ablations or citing BPB in model cards.
Most open MoE knowledge is scattered across engineering blogs and private Slack channels; a consolidated, reproducible guide lowers the barrier for smaller labs to train MoE models instead of falling back to dense architectures. Training teams, ML engineers and researchers can use the runnable code and benchmarked numbers to shortcut months of ablation work. The handbook is hosted free on Hugging Face and sits alongside DeepSeek's own published MoE techniques as a practical reference.
BenchFlow 团队做了个叫 FrontierPhysics 的物理基准，专测 AI 跑科研长任务，连物理学家关心的判断细节都能打分，很快能在 Artificial Analysis 上看到各家模型成绩。
