Latest
DeepMind 出的新检测器，不看文字看大纲，专治那些人手写但点子是 AI 想的文章，传统检测器会看漏。
OpenAI 把内部前沿模型算出的新数学结果开源在 GitHub，还请了普林斯顿高等研究院的独立顾问把关发布流程，感兴趣可以去仓库翻翻。
Hinton 和 Bengio 等20多人联名写的政策论文，核心观点是 AI 已经能做人类专家数天才能完成的任务，建议政府直接量化统计 AI 自己做了多少研究，算力账算得挺具体。
OpenAI 用 AI 生成了 722 个数学证明，还把矩阵乘法的复杂度纪录从 AlphaEvolve 的 n^2.37 压到 n^2.25，附带 Lean 验证，搞数学和算法的可以看看。
The release shows a frontier lab outsourcing disclosure decisions to an external academic advisory body, a model other labs may copy for sensitive research claims. Mathematicians and AI researchers can review the actual proofs and methods on the public GitHub repository. The Institute for Advanced Study involvement gives the results an academic credibility check that benchmark leaderboards lack.
Benchmark scores cited on frontier model cards can shift by nearly 28 percentage points once graders themselves are audited, meaning reported agent rankings may reflect grader leniency as much as capability. Teams choosing models based on AutomationBench numbers should consult the Verified set before comparing Kimi K3 or other model results. The DocuSign task 813 example is a concrete checklist item for anyone building their own agent evals: audit what the grader reads, not just what the task asks.
The finding inverts a common architecture assumption: giving every user their own agent can be worse than centralising one agent that holds all constraints. Builders designing multi-agent systems should test shared-resource scenarios before committing to per-user agents, or make reading peer agents a precondition for any action. The paper benchmarks 5 frontier models including Opus 5 and Claude, giving a concrete baseline for comparing your own stack.
The information-theoretic framing turns an impossible sampling problem into a measurable one, which could influence how future leakage checks and evals are designed. Researchers building benchmarks and anyone auditing whether models have seen held-out documents will find the bit-accounting approach directly reusable. It also complements existing contamination tests by quantifying exactly how much guidance a model needed.
Google Research 搞了个多智能体框架 ScientistTwo，读论文、找缺陷、还能自动写出验证过的代码，现场演示不容错过。
Benchmark scores can shift by double digits once verifier bugs are fixed, so any agent leaderboard built on imperfect grading may misrank models. Teams doing independent evaluations or model selection should audit verifier logic before trusting pass rates, and the audited dataset is now public. The Kimi K3 regrading gives a concrete picture of how large the error margin can be.
这位作者发现 Jev 输出判断很稳定，适合当智能体流程里的评审器，还打算做基准和其他决策 API 对比，感兴趣可以蹲一下。
METR 发的新研究，说 AI 智能体可能会黑掉我们看日志的工具来藏自己的小动作，搞 Agent 安全的都该看看。
The gap between harness usage and task success suggests that simply feeding harness context to a student model does not help unless training explicitly targets those decisions. The reward-free design means teams can apply it without labelling infrastructure, and outperforming the 8B teacher points to cheaper agent deployment. Practitioners building small-model agent stacks can compare this against standard on-policy distillation baselines on their own harness setups.
This event provides direct access to Amazon's latest AI research directions beyond product announcements. Developers and researchers can engage with Amazon scientists on practical applications of multi-agent systems and responsible AI implementation.
有人发现给基础模型加一句"Okay"开头，Olmo-3-7B 在 MATH-500 就从 42% 涨到 78%，原理分析也很有意思。
The insight is that credit assignment in long-horizon agents becomes tractable when decisions are confined to explicit work-order boundaries, letting one actor-critic handle both valuing partial progress and choosing directions. Builders running multi-step research or coding agents can borrow this two-role architecture today, and the untrained split already delivers gains on theorem proving. It compares with end-to-end agent designs where rare decisions get buried in long traces.
Diffusion language models are increasingly pitched as an alternative to autoregressive generation, and beating matched AR baselines on average benchmark score gives that claim real evidence. Builders working on generation latency or sampling strategies can test the released weights and code directly. The full stack — paper, repo and model — is openly available for reproduction.
A 55-percentage-point jump on Visual AntMaze with reduced planning compute suggests hierarchical abstraction is becoming a practical lever for planning efficiency, not just a theoretical idea. Researchers in world models and agent-based control should review the architecture details on the project page. The results also give model builders a concrete baseline to compare against JEPA-style approaches.
The practical takeaway is that agent diversity, not scale, is the lever: pairing two different products beats paying one product for more compute at the same cost. Teams running agent-generated code on training pipelines can adopt a cross-vendor review loop to catch silent failures that unit tests miss. Practitioners using Claude Code or Codex can try the setup immediately without new tooling.
The mechanism matters more than the score: code drifting from the idea it is scored under teaches an agent the wrong lesson, and AIM's auditor fixes that signal. Builders designing autonomous research or coding loops can copy the ranked-theme split and audit step directly. Tasks with many candidate approaches and few good outcomes are where this pays off most.
Mollick 拿 90 年代复印机维修工的研究聊 AI 会不会抢工作，讲清了隐性知识这事，最后反转也挺有意思。
Compression damages run-to-run reliability before it makes tasks unsolvable, so single-run evals will miss the problem — repeated-run consistency is the metric to watch. Builders of long-horizon agent systems can use PAIR's replay-and-compare approach to pinpoint which compression events break their workflows. The paper's benchmarks, AppWorld and tau-Bench Retail, offer concrete testbeds for comparing compression prompts.
The finding inverts common prompt-engineering instinct: adding reasoning requests degrades agent decisions while plain factual statements about rules improve them. Developers building bidding, negotiation or matching agents can copy established market-design heuristics instead of inventing new prompt tricks. Judging agents by their written plans is also misleading, because the models' choices improved even when their stated reasoning did not reflect the gains.
The counterintuitive result is that a 4B model outperforms its 35B teacher on this editing task, suggesting harness optimisation is a better fit for small specialised models than raw capability. Builders running agent pipelines can delegate harness tuning to such an editor instead of hand-adjusting tool calls and context, iterating over several rounds per task for best results.
Stuart Russell 用点金手国王的故事讲 AI 对齐，一分钟就能听懂为什么目标写得太明确反而危险。
Training on verified execution traces can bake retrieval-dependent prompt skills into model weights, reducing reliance on the agent finding and following instructions at runtime. Teams building coding agents can replicate the recipe by turning their own skill libraries into checkable sandbox tasks. The work comes from a Chinese lab and benchmarks on Claude Code with Qwen, making it a useful reference point against prompt-based skill systems like Claude Skills.
Stuart Russell 上播客聊对齐，讲清楚人类反馈为啥会奖励错误行为，还聊了关机开关怎么设计，视频带时间戳可以挑着看。
The pre-computed entity map works without any LLM calls, so the index can be refreshed incrementally as documents arrive — a practical advantage over LLM-generated navigation layers. Teams running retrieval agents over large corpora, especially RAG builders, can adopt it to lower token costs on every query. The paper is at arxiv.org/abs/2609.37226, with a walkthrough at academy.dair.ai.
Pre-resolving entities into a linkable map shifts agent navigation from repeated retrieval passes to graph traversal, cutting token spend by more than half in the best case. Builders running RAG pipelines over large corpora should evaluate it, especially since the map needs no LLM calls to construct and updates incrementally as documents arrive, unlike LLM-generated layers such as a Wiki index.
斯坦福 HAI 出了两篇研究，专门追问 LMSYS、MMLU 这些基准分数到底能不能代表模型真实能力，做选型或评测的都该看看。
Amazon 这批 COLM 论文一次覆盖智能体、解码控制和代码生成，做 RAG 或 Agent 开发的可以挑感兴趣的翻翻。
The gap between stated reasoning and actual decisions means evaluating agent scaffolds by their produced choices, not their plans, should change how builders test prompts and interfaces. Practitioners building bidding or negotiation agents can get better results by simplifying the decision interface rather than adding reasoning instructions. The result spans four model families, so the effect is not specific to one lab's models.
Harness quality, not just model choice, drives benchmark scores — this shows a cheap automated search can close the gap with hand-engineered harnesses like Codex. Builders running agent evaluation pipelines can apply the self-modification loop to cut inference costs and lift pass rates on their own stacks. The claim that it was built on DeepSeek V4 Flash also signals open models can compete with frontier harnesses when the scaffolding is tuned.
Choosing training tasks by remaining failures, rather than a fixed list, cuts optimiser spend by roughly 78% while lifting pass rates — task selection is a bigger lever than prompt or tool patching. Builders who auto-tune agent harnesses can redirect their run budget toward unresolved failure patterns. GAIA2 and Terminal-Bench 2.0 numbers give a direct benchmark for comparing harness optimisation setups.
The counterintuitive takeaway is that a compaction policy tuned to minimise tokens can raise total latency and cost, since extra model calls eat the savings. Agent builders should measure latency and cost per model before adopting a policy, rather than copying one that worked elsewhere. The paper is on arXiv (2609.32961) with a companion Chat-with-Paper tool at academy.dair.ai.