Zhang Hammer

[Matt Dancho (Business Science) @mdancho84]

Agentic Context Engineering

Key takeaways (and get the 23 page PDF):

[meng shao @shao__meng]

再看 Manus 这篇深度好文「Context Engineering for AI Agents: Lessons from Building Manus」还是很有收获和启发

@ManusAI 怎么在不改变模型本身的前提下，通过极致优化输入给模型的上下文信息结构，解决 AI Agent 变慢、变笨、变贵的问题。实际上是在构建一个 “以 LLM 为核心的操作系统”，梳理了这几个方面 🔽

性能优化：像保护数据库一样保护“KV-Cache”
痛点：Agent 运行慢、成本高，本质上是因为每次请求都在做大量重复计算。
技术洞察：大模型推理时，会产生 KV-Cache 临时缓存。如果你的输入提示词的前半部分保持不变，这部分缓存就能被复用，推理速度能快 10 倍以上。
Manus 的解法：“前缀冻结”策略。
· 绝不要在 System Prompt 的开头插入动态变化的信息（比如精确到秒的时间戳）。因为开头只要变一个字符，整个 KV-Cache 就会失效，系统就得从头计算。
· 这就好比写代码时，要把“静态常量”放在最前面，把“动态变量”放在最后面，最大化缓存的命中率。

状态管理：对抗大模型的“健忘症”
痛点：随着任务链变长，模型容易出现“Lost-in-the-middle”，忘了自己最初的目标或中间的某个状态。
技术洞察：Transformer 架构对长文本首尾的注意力最强，中间最弱。单纯把任务历史堆在中间，模型很容易“看漏”。
Manus 的解法：“显式状态复述” (Explicit State Recitation)。
· 不仅仅是记录日志，而是强制模型在每次输出的末尾，重新生成一遍当前的 Todo List 和 当前状态。
· 这利用了 Transformer 的注意力机制特性：把最重要的状态信息，强行搬运到模型视线的最新处，相当于在每一次推理前，都做了一次“注意力校准”。

错误处理：把“报错信息”当作一种训练数据
痛点：传统软件遇到 Error 通常会 catch 掉并重试，但对于 Agent，如果把错误日志删了，模型不知道自己刚才错了，很可能会重蹈覆辙。
技术洞察：大模型具备 In-Context Learning 能力。它不仅能学“怎么做对”，也能从“怎么做错”中吸取教训。
Manus 的解法：保留“负样本”上下文。
· 当 Agent 执行工具失败时，保留完整的报错堆栈。模型看到“路径 A -> 失败”，在下一次推理时，其内部概率分布会自动降低路径 A 的权重。
· 这是一种运行时强化学习。你不需要训练模型，只需要把“失败的尸体”留在现场，模型自己就会学会绕路走。

样本设计：防止模型进入“自动补全模式”
痛点：如果你给模型的 Few-Shot 格式太完美、太统一，模型会变笨。
技术洞察：大模型底层有很强的“模式复制”倾向。当它发现输入全是重复格式时，它会倾向于机械地复制格式，而停止思考 内容 的逻辑。
Manus 的解法：引入“结构熵”（噪音）。
· 不要让历史交互记录看起来像是一个模子里刻出来的。在构建 Context 时，故意保留一些异构的、不那么完美的记录。
· 这种微小的“混乱感”，能打破模型的机械惯性，迫使它每次都必须真正“读懂”当前的内容才能生成回答，而不是简单地做文本补全。

阅读原文
https://t.co/Aw7NcB20zl

[Rohan Paul @rohanpaul_ai]

This paper turns traditional algorithmic trading into a network of AI agents that design, trade, and check strategies.

In tests, that setup beats S&P500 stocks and Bitcoin buy-and-hold while keeping price swings and worst losses noticeably lower.

The authors start from a trading pipeline where different blocks clean data, build signals, control risk, construct portfolios, execute orders, and evaluate performance, which usually needs specialist teams.

They turn each block into its own agent, then use a planner and orchestrator to send task messages between data, alpha, risk, portfolio, execution, backtest, audit, and memory agents.

Large language models like GPT-4o or Llama3 suggest signal ideas and plans, but code tools handle all numbers, training, and backtests so models never touch test returns.

A memory agent keeps short summaries, file links, and safe statistics under hashed IDs, giving later runs context and an audit trail without storing raw prices.

In a 7-stock test and a fast Bitcoin test, the agent system earns more than benchmarks while also reducing return swings and worst peak-to-trough loss.

Overall, this makes trading workflows reusable and safer.

----

Paper Link – arxiv. org/abs/2512.02227

Paper Title: "Orchestration Framework for Financial Agents: From Algorithmic Trading to Agentic Trading"

[elvis @omarsar0]

What's missing to build useful deep research agents?

Deep research agents promise analyst-level reports through automated search and synthesis. However, current systems fall short of genuinely useful research.

The question is: where exactly do they fail?

This new paper introduces FINDER, a benchmark of 100 human-curated research tasks with 419 structured checklist items for evaluating report quality. Unlike QA benchmarks, FINDER focuses on comprehensive report generation.

The researchers analyzed approximately 1,000 reports from mainstream deep research agents. Their findings challenge assumptions about where these deep research systems struggle.

Current agents don't struggle with task comprehension. They fail at evidence integration, verification, and reasoning-resilient planning. They understand what you're asking. They just can't synthesize the answer reliably.

The paper introduces DEFT, the first failure taxonomy for deep research agents. It identifies 14 distinct failure modes across three categories: reasoning failures, retrieval failures, and generation failures.

This systematic breakdown reveals that the gap between current capabilities and useful research isn't about smarter search or better language models. It's about the reasoning architecture that connects retrieval to synthesis.

(bookmark it)

Paper: https://t.co/gAA7feYHm1

[Yangyi @yangyi]

这事儿有点意思…

50 位来自字节、阿里、腾讯等机构的 AI 研究员，联合发布了一份 300 页的论文，里面有些关于代码模型和 Agent 的结论，挺出人意料。

关键点：

> 小模型也能打败大厂的专有模型

RL（特别是 RLVR）让小开源模型在推理方面超越大模型。

一个��� RLVR 在高质量验证问题上训练的 14B 模型，可以匹敌 OpenAI 的 o3。

> 模型学 Python 有点吃力

预训练时混合语言模型是好事，但 Python 和静态类型语言不太一样。

语法相似的语言（Java 和 C#，或者 JavaScript 和 TypeScript）能产生很好的协同效应。

如果把 Python 大量混入静态类型语言的训练中，反而可能因为 Python 的动态类型而适得其反。

> 语言之间有差异（代码 Scaling Laws）

让模型精通一门语言所需的数据量，很大程度上取决于这门语言本身。

论文认为 C# 和 Java 更容易学习（所需训练数据更少）。

Python 和 JavaScript 实际上更难学，有点讽刺（但你会发现 AI 最常被用于这些语言 😂）。

> MoE vs Dense（能力 vs 稳定性）

MoE 模型容量更大，但在 SFT 期间比 Dense 模型更脆弱。

训练中的超参数对 MoE 模型的影响更大，而 Dense 模型更稳定。

MoE 模型还需要持续调整学习率，以避免路由不稳定。

> 代码模型默认是「不安全」的

在公共代码库上训练，会让模型学习到多年积累的不安全编码模式。

安全微调通常效果不佳。

模型可能拒绝写仇恨言论邮件，但会很乐意生成一个存在 SQL 注入漏洞的函数，因为它「能用」。

> 结构比内容更重要

在 CoT 微调中，模板和逐步推理模式比推理的事实性更重要！

模型实际上学习的是分解问题的「认知模板」。

为推理创建合成训练数据时，逐步结构（「如何做」）比确保每个中间事实都完美更重要。

> 「Lost in the middle」问题仍然存在

新模型支持更大的上下文，但当重要代码位于 Prompt（或上下文）中间而不是开头或结尾时，准确性仍然会下降。

> 多 Agent 辩论减少幻觉

对于复杂的软件工程任务来说，这明显更好。

我认为设置辩论非常关键，如果调整不当，甚至可能导致 Agent 强化幻觉。

> 代码安全上的「对齐税」

对齐可能会降低一般的编码能力。

论文实际上提出了一种解决方案，以减轻这种税收的方式进行对齐。

想了解更多细节，可以阅读完整论文。
https://t.co/sXyz0sBiEI

[howie.serious @howie_serious]

if 你关注 ai，then 推荐观看这部 deepmind 纪录片。

尤其适合家庭观看，顺便聊一聊 ai、游戏、思考等话题。每天吃饭看一点，三次就能看完。

这部纪录片去年就电影节放映了，但昨天google 把全片在youtube 上免费放出来了。

非常有趣的是纪录片的名字：《思考游戏》（the thinking game）。

可以说，最让人敬佩的，不是 demis 得了诺贝尔奖，不是 demis 的��商，而是那种“把世界当成一场思考游戏”的眼光。

======

如果把整个人类文明看成一场漫长的“思考游戏”，这部片子讲的，就是一群人如何把这场“思考游戏”本身拆开、重写规则的故事。

对他们来说，棋盘、街机游戏、星际争霸、蛋白质折叠，甚至整座虚拟世界，都只是训练一种东西的道具：通用智能。

Demis 的人生像一条被同一问题穿线的轨迹：少年国际象棋天才，在教堂里熬到第十个小时，突然意识到，也许把几百个顶尖大脑锁在一间大厅里只为分出输赢，是在浪费心智；于是转向游戏设计，用主题公园里的“呕吐连锁反应”去逼真模拟人类行为；再到剑桥研究大脑，从神经科学里���师，最后拉上 Shane，干脆在伦敦创立 DeepMind，明目张胆地说：我们要做 artificial general intelligence。

他们选择了一条看起来很“不务正业”的路径：先让系统学会在 Atari 里打 Pong、挖 Breakout 的侧翼隧道，再去下围棋、打星际。强化学习、奖励信号、环境与智能体的闭环，在这些游戏里被一次次打磨：系统一开始连挡球都不会，只知道分数越高越好，最后却能在围棋上下出连职业棋手都无法想象的招法，能在星际里练出堪比职业选手的多线操作。这里有一种残酷的坦诚：你不告诉机器规则，只给目标，它会自己找到通向高分的“人类从没想过”的路径。

但真正让这条路线显出重量的，是当他们把同样的思���搬到现实科学问题上。蛋白质折叠这个自 20 世纪 60 年代就存在的难题，几十年里无数聪明人试过，都卡在实验慢、数据少上。第一次参加 CASP 比赛，他们的 AlphaFold 确实赢了同行，却只是在“一个没人解好的问题上解得稍微好一点”。团队后来回看那一届，很坦白：我们知道自己还很烂，这只是最高的梯子，但月亮还在很远的地方。

转折来自第二次“总攻”：重写数据管线，把物理与生物学的结构知识真正喂进模型，让系统不再只是模仿，而是内化约束。最终在 CASP14 上，他们的结构精度几乎等于按下按钮得到实验级结果，主持人说，经过半个世纪，这个问题可以宣布被解决了。接下来的一步更激进：既然可以在一个月里预测全已知物种的蛋白质，那就别等别人排队提交序列，直接把全地球的蛋白质都“折完”，免费公布，让它成为人类生物学的基础设施。

技术从来不是中性的玩具。AlphaGo 之后的“Sputnik moment”、AI 军备竞赛的隐喻、与曼哈顿计划的对照、对 autonomous weapon 的明确拒绝，以及对 move fast and break things 的公开反感，都在提醒一个事实：当你试图造出一种比人类更强的认知系统时，你不是在做一个新 app，而是可能在改写历史的分界线。连 DeepMind 自己的人都开始担心：也许我们将来要做的，不是说服世界我们已经造出了智能，而是解释我们还没到那一步。

这部片子最有价值的地方不在于追星 DeepMind，而在于它把几个关键问题扔在桌面上：什么叫“通用”的学习能力？在一个只看奖励的系统里，价值是如何被编码、��放大的？科学探索和工程冲刺之间，节奏如何拿捏，才不会“死在时间轴上”？当智能被外包给机器，人类该把注意力放在哪里——是新的问题意识、新的制度与治理，还是对自身价值观的重新书写？

也许真正值得我们模仿的，不是某个算法或产品，而是那种“把世界当成一场思考游戏”的眼光：敢于承认自己暂时还很烂，敢于在失败里调整节奏，也敢于在成功时把成果毫无保留地推向全人类。

https://t.co/OzJmSucWt9

[小樱💞｜实用工具分享 @xiaoying_eth]

一位 Google 的高级工程师，
刚刚发布了一份名为 Agentic Design Patterns 的 424 页文档

不是 PPT，不是文章，
是真·课程级别的大工程

有兴趣的可以查看🔗 https://t.co/fi6tSUwynD

[howie.serious @howie_serious]

memory/记忆，它是什么？它如何运作？

不论对人类，还是对llm，这都是最重要的问题。

周末，一起来读一篇，要么没啥用/有大用的文章～📖

[Sauers @Sauers_]

Big if true

[Dongxi 东锡 NLP @dongxi_nlp]

仔细读了 Agent KB，太好了！我觉得完全可以创造一个新词：experience engineering（经验工程）。

这里的 experience 就是指 Agent 在解决不同问题时积累的经验。

相比于 context engineering（上下文工程），experience engineering 的抽象层次更高。因为 experience 不仅包含了 context，还涵盖了问题模式、问题解决的 workflow、元数据（meta data）以及关系图谱（graph）等信息。

Experience 可以像知识库一样被存储起来，不同的 Agent 在面对不同问题时，都可以通过 Reason -> Retrieve -> Refine 的过程学习已有的经验，极大地提高 Agent 解决问题的能力。

优秀的工作，来自 OPPO agent team。Salute！

[James Zou @james_y_zou]

🏆Thrilled that #CollabLLM won the #ICML2025 Outstanding Paper Award!

We propose a new approach to optimize human-AI collaboration, which is critical for agents. Congratulations to my fantastic co-authors; great job @ShirleyYXWu and Michel Galley driving the project!👏

[Zichen Liu @zzlccc]

Though not attending #ICML2025 in person, I'm super excited to share 3 accepted papers:

1.🎊Best Paper Honorable Mention @ AI4MATH workshop: Understanding R1-Zero-Like Training: A Critical Perspective (a.k.a Dr. GRPO but I think the paper is more than this loss fix)

2. Main Conference Spotlight: Continual Reinforcement Learning by Planning with Online World Models; I worked on the continual learning topic for one year before diving into RL for LLMs https://t.co/zuUy0BV3ny

3. AI4MATH Workshop Poster: Optimizing Anytime Reasoning via Budget Relative Policy Optimization; a dense reward training framework that controls thinking budget (a cool feature of Gemini 2.5 Pro)

[凡人小北 @frxiaobei]

挺好的一本书，没想到竟然能把 Agent 做得这么��、这么落地。建议收藏，常看常新。
阿里云这份《金融行业 Agent 百景图》真的太全太实了，把过去一年在银行、证券、保险里做的大模型落地实践梳理成册，200 多个 Agent 案例，干货密度超高。

比如智能投研怎么跑，客服怎么加 AI，风控怎么做合规验证，甚至连监管辅助 Agent 都写出来了；全程围绕 MoA（Mixture of Agents）+ RAG 架构，搭配工具链 + 多模态，完全就是一本金融 AI 工程化说明书。

如果你是已经在做金融 AI 的，看这个能找到方向参考；
如果你是对金融感兴趣、但不知道 AI 能干嘛、或者 Agent 怎么落地的，这份可以直接抄，照着做都行。
哪怕你是其他行业，只要在想“我这行业 AI Agent 能干嘛”，这份也能提供非常多可迁移的思路，值得好好读一遍。

电子版链接放评论区，建议收藏一份，做 AI 的别错过，做金融的更得看。

[johann.GPT @ProgramerJohann]

看到个接近 10K🌟的多智能体 LLM 金融交易框架，，突破传统单Agent限制。

架构包含：
- 分析师团队：基本面分析师评估财务状况、情绪分析师分析社交媒体情绪、新闻分析师监测宏观经济、技术分析师检测交易模式；
- 研究团队：看涨看跌Agents通过结构化辩论评估风险收益；

交易Agent整合多方报告做决策；还有风险管理和投资组合Agent。多Agent协作实现全方位市场分析和智能交易。

[Together AI @togethercompute]

Announcing DeepSWE 🤖: our fully open-sourced, SOTA software engineering agent trained purely with RL on top of Qwen3-32B. DeepSWE achieves 59% on SWEBench-Verified with test-time scaling (and 42.2% Pass@1), topping the SWEBench leaderboard for open-weight models.

Built in collaboration with the @Agentica_ team.

💪 DeepSWE is trained with rLLM, Agentica’s modular RL post-training framework for agents. rLLM makes it easy to build, train, and deploy RL-tuned agents on real-world workloads — from software engineering to web navigation and beyond.

🤗 As always, we’re open-sourcing everything: not just the model, but the training code (rLLM), dataset (R2EGym), and training recipe for full reproducibility.

🔥 Train DeepSWE yourself. Extend it. Build your own local agents. No secrets, no barriers.

DeepSWE and rLLM mark our major shift: from training language reasoners to building language agents that can truly learn from experience.

We believe the future of AI lies in experience-driven learning — and we’re here to democratize it.

Welcome to the era of experience. 🌍

[Carlos E. Perez @IntuitMachine]

OpenAI self-leaked its Deep Research prompts and it's a goldmine of ideas! Let's analyze this in detail!

[Zengzhi Wang @SinclairWang1]

Finally had a bit of time to jot down some thoughts on this solid, open data engineering work from @essential_ai.

This work brings Essential-Web, a 24T-token pre-training corpus, to the open-source community. I've always appreciated open-source research, as it can significantly promote AI democratisation. Beyond the data release, this work also provides guidance on building a systematic Taxonomy of Categories for web documents to support data governance, with impressive levels of detail—including even scripts. This technical report, in my view, deserves multiple careful readings.

Notably, it also—finally—acknowledges our contributions to curating math pre-training corpora, such as MegaMath. I sincerely appreciate that☺️. Especially given that several orgs have used our data from our recent work or referenced our work without extending the appropriate credit. Let’s be honest: conducting research and doing real engineering work on data is far from trivial—yet it’s often dismissed as lacking novelty. 😅

That said, I do have some respectful disagreements regarding the experiments on data quality comparisons—particularly in the math domain (cc @AndrewHojel @ashVaswani). In our MegaMath paper, we showed that even the full MegaMath-Web corpus outperforms OpenWebMath in a 55B-token continual pretraining setup (see Figure 2). Also, I’d recommend clarifying how the top 10% of MegaMath-Web documents were selected. Furthermore, I observed that several existing domain-specific datasets (e.g., code and medical) show performance comparable to the DCLM baselines reported in this paper. I believe this might raise similar concerns for others as well.

There’s also a common misunderstanding—especially among folks who aren’t hands-on in pretraining data engineering—about types of pretraining corpora.

In my view, there are two major types:
1. True pretraining corpora, meant to lay the foundational knowledge for LMs.
2. Mid-training corpora, used in later stages (e.g., during LR decay) with focused curation, smaller in scale, and tailored for specific capabilities or benchmarks.

For instance:
- FineWeb is a broad-coverage true pretraining corpus.
- FineWeb-Edu is curated for high educational value, ideal for mid-training and great for benchmarks like MMLU.

In the context of the math domain:

- MegaMath-Web = a true pretraining corpus (about 100 Common Crawl dumps from 2014–2024).
- FineMath (3+, 4+) = a mid-training corpus, filtered via edu-style classifiers.

So what’s the “educational” version of MegaMath-Web? That would be MegaMath-Web-Pro—we used the same edu classifier as FineMath to extract high-ed value docs, followed by LLM-based refinement for noise reduction.

But to be clear: simply filtering MegaMath-Web by math_score isn’t equivalent to using the edu classifier. These are different metrics. It’s important for fair comparisons.

Given that EAI-TAXONOMY Math w/ FM and FineMath-3plus are reported in Table 3, I believe MegaMath-Web-Pro also deserves inclusion.
(cc @youjiacheng —thanks for the kind mention today and recognition of our work!)

Another way to evaluate math corpus quality?
Check out our recent work: OctoThinker.

We found that mid-training on MegaMath-Web-Pro (and soon, MegaMath-Web-Pro-Max) significantly boosts RL scaling—outperforming FineMath-4+. (See third figure attached.)

Blog here: https://t.co/pgHe4yCA3I

The tech report + MegaMath-Web-Pro-Max open release is coming late this week or early next. Still working hard on it—stay tuned!

Finally, I want to shout out to the amazing data engineering work from @huggingface (cc @LoubnaBenAllal1). Their contributions—FineWeb, FineWeb-Edu, FineMath, Nanotron—are hugely appreciated. Their curation, technical depth, and open-source spirit inspired much of our own work, including ProX (https://t.co/WPvYLONFG3) and MegaMath. Thank you!

If you’re building models and need high-quality corpora—feel free to explore ours:
MathPile: https://t.co/ciuGaFvJDe
DCLM-Pro: https://t.co/7749Pc5xfk
FineWeb-Pro: https://t.co/njdto6ia65
MegaMath: https://t.co/VIdyPlEx30

More is coming—let’s keep brainstorming & building. 🚀

[Infini-AI-Lab @InfiniAILab]

🚀 Excited to introduce our latest work GRESO: a method that identifies and skips millions of uninformative prompts before rollout and achieves up to 2.0x wall-clock time speedup in training.

More rollouts lead to better model performance, but they’re also a major bottleneck in RL ⚠️—and not every prompt is worth the cost. Why waste resources on the rollout of uninformative data?

🔗 Blog: https://t.co/Yspx6AT7rv
📄 Paper: https://t.co/Uqedh29GhA
💻 GitHub: https://t.co/mEtluCagUD

[wh @nrehiew_]

An extremely underrated paper