Latest
大家都盯着实验室刷数学证明，这条讲历史学家怎么用 AI 破 John Dee 密码、挖 Darwin 的思想源头，角度少见。
Google DeepMind 这篇论文挺扎心的：你给智能体一句自信但错误的建议，它会照做还不吭声，三大主流模型最多掉分 46.7%。
UCSD 和 Lambda 的 CVP 方法让 3D 模型看清你问的是哪个物体，5 个基准全赢 Video-3D-LLM。
Stanford 和 Together AI 的新论文，智能体别再投票吵架，学15道题分工互检，3个模型从48.8%干到66.7%。
Boltzbit 给出了智能体自我改进的第二条路：直接改权重而不是堆记忆，理论上算力省约 1000 倍，思路很值得琢磨。
Ethan Mollick 说那张刷屏的 METR 任务时长曲线该退休了，还给出了替代图，做 AI 汇报的人可以看看。
Meta 给智能体配了个记忆员，专记它犯过的错并适时提醒，把 Claude Sonnet 4.5 成绩从 37.6% 拉到 45.9%，做智能体的可以看看。
Fireworks 的特种智能指数又添新成员，这次是 FactoryAI 的 Legacy-Bench，专门考模型能不能搞定遗留代码系统，做企业级 Agent 的可以关注。
Claude 参与发现了一种可能是新基因编辑机制的分子机器，博主据此讨论 Anthropic 那句几十万亿收入预测有多靠谱。
Google 这篇论文拆了 serverless CPU 跑量化小模型的延迟构成，55-70% 耗在加载权重，加内存到 8 GB 就能让推理快近一倍，部署前值得看。
Stanford 的 James Zou 团队把论文变成一个个智能体，它们能互相协作做分析、甚至合作出新发现，思路挺有意思。
Stanford HAI 出了份政策简报，专门讨论能看、能行动的世界模型为什么管不住，做政策或研究的朋友可以看看。
The gap between intending an operation and executing it correctly suggests monitoring targets should distinguish procedural signals from factual accuracy when auditing chains-of-thought. Mechanistic interpretability researchers and AI safety teams gain a concrete probe target in middle layers, usable for building internal reasoning monitors rather than relying on surface text. Anyone comparing faithfulness of chain-of-thought monitoring across labs can use the 8-operation taxonomy as a shared baseline.
The finding flips the usual assumption that authorization failures live in the reasoning or execution layer; here, memory compression itself silently rewrites access rules, so agents may look consistent while acting on permissions that were revoked. Teams building long-running agents should add permission provenance and deterministic revocation tracking to their memory pipelines, not just prompt-level guardrails. Practitioners can test the paper's typed incremental-memory setup against their own agent stacks.
The deployment detail matters: a guard skill installed but never invoked left agents exposed, so teams must enforce explicit consultation in system prompts, not just add it to the skill library. The defense-as-skill design means existing agent frameworks can adopt runtime checking without architectural changes. Builders shipping third-party skills should read the arXiv paper (2609.01487) and pair it with sandboxing and permission layers.
The pattern of systematic underestimation on benchmarks while overestimating biological risk offers a useful correction to how decision-makers calibrate AI timelines. Researchers, investors and policy teams can use these deltas to adjust their own probability estimates rather than relying on expert medians. Superforecasters being more skeptical than domain experts is itself a finding that challenges the assumption that forecasting skill transfers to technical fields.
有人泼冷水了：Anthropic 说的这个新酶系统，目前只在大肠杆菌里表达过一次，原理都没搞清，但思路确实有意思。
Epoch AI 搞了个叫 FAB 的新基准，让模型看宜家说明书加半成品照片找拼装错误，10 个月内最高分从 28% 冲到 80%。
METR 给 Anthropic 的 AI 研发加速打了个具体分数：1.5 倍。想知道 AI 到底多大程度替代了工程师干活，这个口径很严谨，值得看看。
Environment quality, not environment quantity, may be the real bottleneck in agentic RL — a collection that looks large can be mostly reward noise. Teams training agents should audit verifiers before scaling data, and can adopt RIVER's filtering plus repetition penalties. River-8B's open checkpoints offer a strong 8B terminal-agent baseline against other open RL-trained models of the same size.
Rubric-augmented rewards trade raw benchmark score for trajectories a judge rates far higher, a hint that outcome-only RL over-optimises for narrow metrics while behavioural rubrics shape more genuinely skilled agents. Teams training terminal agents can reuse the open pipeline to convert existing Agent Skills corpora into RL environments instead of hand-crafting tasks. The US$90,000 generation cost also sets a concrete budget baseline for anyone planning similar corpus builds.
Harness distillation shifts the value of hand-built agent scaffolding from runtime glue into permanent model capability, which could change how labs decide what to ship as harness versus as weights. Teams building agent frameworks can use the corrected-run pipeline to convert an optimized harness into training data for their own models. Compared with the code-as-harness alternative, the agent-as-harness form is the stronger option for frontier models here.
Boltzbit 和剑桥搞了个新架构，把聊天里的新知识直接编译成 LoRA 权重写进模型，长对话和多轮场景下比 RAG 省算力还更准。
Continuous evaluation of production agents is usually too expensive to run often, and this paper shows a practical shortcut that keeps scores accurate to about one point. The transferable, difficulty-stratified subsets mean teams with any agent stack can adopt the method without recomputing item parameters each time. Builders running agent benchmarks on a budget can start with a one-day calibration window and scale from there.
Rollout generation dominates RL training cost, and this work shows batch-size scaling plus a retuned learning rate can cut wall-clock time without new hardware. Teams running GRPO or PPO fine-tuning can apply the retuning rule directly to squeeze more throughput from existing GPU clusters. The numbers come from Tencent's own infrastructure, so results may shift with different model sizes and interconnects.
腾讯混元做了个新基准 WebCraftBench，用 369 条真实需求测 AI 写网页到底靠不靠谱，测了 17 个模型，和 Code Arena 榜单相关性 0.89。
Optimizing an agent's harness against a benchmark can inflate eval scores while degrading real-world performance, and this paper quantifies that gap with held-out numbers. Builders running automated prompt or harness optimization should add critic and pruning loops like RRSI's to protect out-of-distribution performance. The paper and a chat-with-paper interface are available via arXiv and dair.ai.
Microsoft Research 研究说机器人不用把推理都塞进本体，放到外部算反而任务成功率更高、效率更好，做具身智能的可以看看。
The practical takeaway is that a shared directory can substitute for role assignment in multi-agent design, cutting the agent count needed by roughly 4x on some benchmarks. Builders designing agent frameworks should weigh communication overhead against these gains, and the paper's negative results offer clear criteria for when to skip coordination. The ARC-AGI-3 and MNIST results give concrete reference points for evaluating multi-agent architectures.
The non-obvious finding is that letting agents write their own teamwork protocol beats even an oracle that picks the best individual answer, and the protocol transfers across benchmarks after training on just 15 problems. Builders of multi-agent systems can copy this approach instead of hand-tuning roles and pipelines, with off-the-shelf models like DeepSeek-V3 already sufficient as team members.
Anthropic 出了份报告，讲有人拿代理偷偷转发用户 prompt 到 Claude API 蒸馏模型，还点出隐私风险，做 AI 产品的人都该看看。
Retrieval over plain markdown folders is becoming a common pattern for agent memory, and WFM shows link structure itself can carry signal rather than only page text. The 10.5x faster training protocol matters for teams fine-tuning their own retrieval models on private wiki data. Practitioners building agents on file-based memory, RAG pipelines or multi-hop reasoning tasks can test the approach via the paper and the accompanying Chat with Paper tool.
Perplexity 公开了他们训练 Computer 智能体的后训练细节，用 RFT 加自蒸馏纠正错误工具调用，实测失败率降了 21%，做 agent 的话值得看看他们怎么用真实用户会话。
Perplexity 把自家模型改成会从错误里自我纠偏，A/B 测试里工具调用失败少了 21.2%，做法写得很细。
Perplexity 放出实测数据：不换模型只加提示词，失败率就从 75.1% 降到 6.3%，做工具调用的值得对比下自己的数字。
The same model can score differently depending on the harness around it, so benchmark numbers reflect the full stack rather than the model alone. Teams evaluating coding agents should rerun comparisons under their own harness before drawing conclusions. ReFigBench adds a new evaluation axis, and details are on arXiv (2609.18844) alongside a chat interface from dair.ai.
The benchmarks show where cheap small models can replace larger ones in RAG pipelines: stopping decisions and relationship filtering see large latency and cost savings with matched recall. Engineers building retrieval systems can use the published code to pick where to route between Jev, GPT-5-mini and Voyage rerank-3. The evaluation includes Chinese-language queries, useful for teams serving bilingual users.
Token-level correction can cut annotation cost because reviewers mark specific tokens instead of full responses, which changes how teams budget for on-policy data pipelines. Practitioners building RLHF or agent fine-tuning workflows can apply it to reduce labelling overhead and compare it against response-level annotation baselines.
