这个仓库把开源的 AI 自主渗透测试框架 ARTEX 做了韩语本地化（上游为 Autumn-27/ARTEX，AGPL-3.0），话题标签覆盖自主 Agent、蓝队、检测工程、Sigma 规则与威胁检测^[24]。它出现的时间点很微妙：同一天的安全归因指出，攻击韩国银行的工具正是 ARTEX——同一个人工智能攻防框架，一边被用于攻击，一边被本地化成防守方的检测工具。值得借鉴的是攻防两侧共用同一套 Agent 基础设施已是现实，防御方应当直接研究这些框架的行为特征；限制是这类工具的授权边界完全依赖使用者，仓库本身不提供授权校验，用于非授权环境会直接违法，团队必须在隔离环境与书面授权下使用。

This repository localises ARTEX, the open-source autonomous AI penetration-testing framework, into Korean (upstream Autumn-27/ARTEX, AGPL-3.0), with topics spanning autonomous agents, blue team, detection engineering, Sigma rules and threat detection^[24]. Its timing is pointed: the same day's security attribution identified ARTEX as the tool used against South Korean banks — one AI attack-and-defence framework being used offensively while being localised as a defensive detection tool. Worth borrowing is that both sides of the fight now share the same agent infrastructure, so defenders should study these frameworks' behaviour directly; the limit is that authorisation depends entirely on the user, the repo provides no scope checks, and use outside authorised environments is illegal — work in isolation with written authorisation.

nanoMuse 是一个开源个人 Agent：一个有名字和形象、App 关掉也继续工作、会记住你、在不可撤销的操作前先问你的 Agent，覆盖 Android、iOS/iPadOS、Windows/macOS/Linux 与浏览器，并把手机屏幕和电脑当作它的「手」^[25]。它要解决的是个人 Agent 的形态与归属问题：闭源方案把账号、设备、记忆与跨周对话放在厂商云里，用户既无法审计也无法迁移。做法上用 GPL-3.0 开源实现，把记忆存成用户可读的文件，并让用户自选模型。值得借鉴的是「记忆可读 + 不可撤销操作先确认 + 跨设备同一会话」是个人 Agent 的三个关键设计；限制是跨设备自动化依赖各平台的辅助功能与后台权限，实际稳定性与功耗需要实测，仓库也把「手」的评测套件列为路线图而非已完成项。

nanoMuse is an open-source personal agent: it has a name and a face, keeps working while the app is closed, remembers you, and asks before anything you cannot undo, spanning Android, iOS/iPadOS, Windows/macOS/Linux and the browser, with your phone's screen and computer as its hands^[25]. It addresses the form and ownership of personal agents: closed offerings keep accounts, devices, memory and weeks-long conversation in a vendor cloud where users can neither audit nor migrate. The implementation is GPL-3.0, stores memory as user-readable files and lets the user choose the model. Worth borrowing is that readable memory, confirmation before irreversible actions and one shared session across devices are the three core design choices; the limit is that cross-device automation depends on each platform's accessibility and background permissions, so stability and power use need field testing, and the evaluation suite for the hands is on the roadmap rather than done.

这是一个出行路书工作流技能：联网实查并结合多源交叉验证，产出「可核验、能执行」的旅行攻略，覆盖吃住行游拍避，附美食情报卡、基准骨架与校验工具，支持一键部署在线版^[26]。它要解决的是旅行规划类 Agent 最致命的问题——幻觉。行程里的营业时间、订票规则与交通衔接一旦编造，用户到达现场才会发现。做法上把「多源交叉验证」写进工作流，并配套校验工具与基准骨架，让结论可追溯。值得借鉴的是为高时效性任务专门设计验证步骤，而不是依赖模型常识；限制是交叉验证的有效性取决于源站质量与抓取成功率，仓库未给出评测集与准确率，实际可用性需要按目的地实测。

This is a travel-planning workflow skill: live web checks combined with multi-source cross-validation produce an itinerary that is verifiable and executable, covering food, lodging, transport, sightseeing, photography and safety, with a food-intel card, baseline scaffold and verification tools, deployable online in one step^[26]. It addresses the fatal flaw in travel agents: hallucination. If opening hours, booking rules or connections are invented, users only find out on arrival. The design writes multi-source cross-validation into the workflow with verification tooling and a baseline scaffold so conclusions are traceable. Worth borrowing is designing an explicit verification step for time-critical tasks instead of relying on model priors; the limit is that cross-validation quality depends on source quality and crawl success, and no evaluation set or accuracy figures are published.

harness-mix 的定位从话题标签就能看出来：在同一个工作流里混搭多个 Agent harness，标签覆盖 Claude Code、Codex、Cursor、DeepSeek Harness、Hermes、opencode、Pi、Trae、Zcode 与 worktree 等^[30]。它要解决的是现实中的多工具并存问题：团队往往同时用几套 harness，各有擅长的任务，但上下文、会话与产物互不相通。做法上把不同 harness 的能力拼在一起使用。值得借鉴的是「混搭」本身是一个真实需求，值得用统一的任务与产物格式来降低切换成本；限制是这个仓库没有写描述，本条目只能依据其话题标签判断方向，具体实现与稳定性无法从元数据评估，采用前需要读代码。

harness-mix's positioning is visible from its topics alone: mixing several agent harnesses inside one workflow, with tags covering Claude Code, Codex, Cursor, DeepSeek Harness, Hermes, opencode, Pi, Trae, Zcode and worktree^[30]. It addresses the reality of coexisting tools: teams often run several harnesses with different strengths while context, sessions and artefacts stay incompatible. The approach combines capabilities across harnesses. Worth borrowing is that mixing is a genuine need, and a unified task-and-artefact format would cut switching cost; the limit is that the repo carries no description, so this entry can only infer direction from its topic tags — read the code before adopting, since implementation and stability cannot be judged from metadata.

这个项目强调自己「不是问答机器人，而是能真正干活的数字分身」：共享智能体与项目上下文，让项目所有成员一起协同，并像人一样把资料与经验抽象、总结回项目上下文^[31]。它要解决的是团队内 Agent 重复造上下文的问题：每个人各自喂资料，结论不共享，Agent 之间也无法承接彼此的工作。做法上把「项目上下文」做成一等公民，Agent 在其中读写并沉淀经验。值得借鉴的是共享上下文是团队级 Agent 的关键资产，比单个 Agent 的能力更重要；限制是共享上下文同时意味着错误与偏见会快速扩散到全队，仓库未说明权限、版本与冲突处理机制，规模上去之后治理成本可能超过收益。

This project stresses that it is “not a Q&A bot but a digital twin that actually works”: shared agents and project context let all project members collaborate, with experience and materials abstracted back into that context like a person would^[31]. It addresses duplicated context-building inside teams: each person feeds material separately, conclusions are not shared, and agents cannot take over one another's work. The design makes project context a first-class citizen that agents read, write and accumulate experience into. Worth borrowing is that shared context is the key team-level asset, more important than any single agent's capability; the limit is that shared context also spreads errors and bias across the team quickly, and the repo does not describe permissions, versioning or conflict handling, so governance cost may outgrow the benefit at scale.

这是一个远程 MCP（Model Context Protocol，模型上下文协议）服务清单，收录了一批可远端调用的 MCP 服务^[29]。它要解决的是 MCP 生态的发现成本问题：本地 MCP 需要安装与配置，远程 MCP 只要一个 URL 就能接入，但「有哪些、谁维护、是否可信」缺少索引。做法上以清单形式集中整理，便于对比与试用。值得借鉴的是接第三方 MCP 前先确认维护者与权限范围，清单本身不是安全背书；限制是清单类仓库通常不做安全审计，远程服务的权限、日志与数据去向由各服务方决定，接入前需要逐个核查。

This is a catalogue of remote MCP (Model Context Protocol) servers, collecting services that can be called remotely^[29]. It addresses discovery cost in the MCP ecosystem: local MCP servers need installation and configuration while remote ones only need a URL, yet there is no index of what exists, who maintains it and whether it is trustworthy. The list gathers them for comparison and trial. Worth borrowing is to check maintainer and permission scope before wiring in a third-party MCP server — a catalogue is not a security endorsement; the limit is that such repos rarely audit security, and the permissions, logging and data destinations of remote services are set by each provider, so each must be vetted individually.

alphacode 的定位是免费的 MIT 许可编码 Agent：无需 API Key，内置免费模型，也可以自带 Claude、GPT、Gemini、DeepSeek、Ollama 等 50 多个后端；支持 swarm 模式、40 多个工具、浏览器与桌面自动化，用 Rust 实现^[28]。它要解决的是编码 Agent 的使用门槛与成本问题：对个人开发者与教学场景，订阅与 API 计费是主要阻力，而「内置免费模型」把试跑成本降到零。值得借鉴的是把「零成本试用」做进产品，能显著扩大早期用户基数；限制是内置免费模型的来源、配额与稳定性不在仓库控制之内，一旦上游变更或限流，可用性会骤降，生产环境仍应配置自有后端。

alphacode positions itself as a free MIT-licensed coding agent: no API key required, with a built-in free model or bring-your-own Claude, GPT, Gemini, DeepSeek, Ollama and 50+ backends; swarm mode, 40+ tools, browser and desktop automation, written in Rust^[28]. It addresses the cost and access barrier for coding agents: subscriptions and API billing are the main obstacles for individuals and teaching settings, and a built-in free model drops trial cost to zero. Worth borrowing is that baking zero-cost trial into the product meaningfully widens the early user base; the limit is that the free model's provenance, quota and stability are outside the repo's control, so availability can collapse if the upstream changes or throttles, and production use should still configure a self-owned backend.

这个项目把自己描述为「会学习、会适应、无论什么任务都能把事做完的伴侣型 AI Agent」，属于通用 Agent 框架类^[27]。它要解决的是 Agent 从「单次任务」走向「长期陪伴」时的状态管理问题：伴侣型 Agent 需要跨会话记住偏好与历史，并在任务类型不断变化时保持稳定。做法上以框架形式给出 Agent 与工具的组织方式（具体设计需读代码确认）。值得借鉴的是长期运行的 Agent 更需要状态管理与可解释性，而不是更强的单轮推理；限制是仓库描述极简、没有给出评测与架构说明，无法判断与同类框架的差异，采用前需要投入时间评估。

This project describes itself as “the companion AI agent that learns, adapts and gets the work done no matter the task”, in the general agent framework family^[27]. It addresses state management as agents move from one-off tasks to long-term companionship: a companion must remember preferences and history across sessions and stay stable as task types change. The framework supplies an organisation for agents and tools, though the concrete design requires reading the code. Worth borrowing is that long-running agents need state management and explainability more than stronger single-turn reasoning; the limit is that the description is minimal with no evaluation or architecture detail, so differentiation from similar frameworks cannot be judged without an investment of review time.

Cortico 是一个面向「人格化机器人」的事件流（event-stream）Agent 框架，主要用 TypeScript 实现^[32]。它要解决的是虚拟形象类 Agent 的时序问题：这类 Agent 需要持续接收环境事件、维持人设一致并实时响应，而常见的「请求—响应」结构难以承载持续状态。做法上以事件流作为骨架，让人设与响应逻辑挂在事件处理上。值得借鉴的是把「持续事件」而不是「对话轮次」当作一等抽象，适合直播、虚拟主播等实时场景；限制是仓库以框架与标签为主，未给出延迟、并发与一致性方面的数据，实时场景的真实表现需要自行压测。

Cortico is an event-stream agent framework for building persona bots, primarily in TypeScript^[32]. It addresses the timing problem of avatar-style agents: they must continuously consume environment events, hold a consistent persona and respond in real time, which a conventional request-response structure struggles to carry. The design uses the event stream as its skeleton, hanging persona and response logic off event handling. Worth borrowing is to treat continuous events rather than conversation turns as the first-class abstraction, which suits live and VTuber settings; the limit is that the repo is mostly framework and tags with no latency, concurrency or consistency figures, so real-time behaviour needs load testing.

nautilo 的定位是自托管协作工作区，让人类与「机器人们」在桌面、移动与网页端一起创建与编码，MIT 许可、开源^[33]，并接入 MCP。它要解决的是多 Agent 协作的「人机同场」问题：多数编排工具面向单个操作者，而团队场景需要多人同时看见并介入 Agent 的工作。做法上把人类成员与 Agent 放在同一协作空间中，共享上下文与产物。值得借鉴的是把「人类介入点」设计成工作区的一部分，而不是留在命令行里；限制是多人 + 多 Agent 同时操作会带来权限、冲突与审计问题，仓库未说明这些机制，团队部署前需要自行补齐治理。

nautilo positions itself as a self-hosted collaborative workspace where people and machine people create and code together across desktop, mobile and web, MIT-licensed and open source^[33], with MCP support. It addresses human-agent co-presence in multi-agent collaboration: most orchestration tools target a single operator, while team settings need several people to see and intervene in agent work simultaneously. The design places human members and agents in one shared space with common context and artefacts. Worth borrowing is to design human intervention points as part of the workspace rather than leaving them in a terminal; the limit is that many people plus many agents raises permissions, conflict and audit problems which the repo does not address, so governance must be added before team deployment.

RECAST 指出常规 RAG（Retrieval-Augmented Generation，检索增强生成）与现有 Agent 检索的共同盲点：很多任务需要的证据并不存在于任何单一来源里，而必须通过过滤、聚合或跨条目的计算推导出来，但主流方案仍然把问题当成「检索」^[34]。它要解决的是「证据要算出来、而不是找出来」这一类任务的上下文构造问题。做法上把证据构造建模成在异构检索与计算操作上的序贯决策：一个轻量 RouterLM 迭代地选择并表述基础操作（或为冻结的 CompilerLM 定制操作，编译成可执行代码），一旦判断证据充分，就把被接受的证据交给冻结的 AnswerLM 生成答案；RouterLM 用监督微调加 GRPO 训练。对做检索与数据 Agent 的团队，参考价值是把「算」与「找」统一进同一个决策循环，并按证据充分性决定何时停止；限制是摘要未给出完整的对比基线数字，且流水线依赖 CompilerLM 的代码生成正确性，编译失败时的回退策略没有说明。

RECAST names a blind spot shared by conventional RAG (Retrieval-Augmented Generation) and existing agentic retrieval: for many tasks the needed evidence is absent from any single source and must be derived by filtering, aggregating or computing across items, yet mainstream systems still treat the problem as search^[34]. It addresses context construction for tasks where evidence must be computed rather than found. The design formulates evidence construction as a sequential decision process over heterogeneous retrieval and computation operations: a lightweight RouterLM iteratively selects and formulates primitive operations, or specifies custom operations for a frozen CompilerLM to translate into executable code, and once it judges the evidence sufficient it passes the accepted evidence to a frozen AnswerLM for the final answer; RouterLM is trained with supervised fine-tuning followed by GRPO. For retrieval and data agent teams the reference is to unify computing and searching in one decision loop, stopping on evidence sufficiency; the limit is that the abstract omits full baseline numbers and the pipeline depends on CompilerLM generating correct code, with no fallback described for compilation failure.

CoTrace 针对终端 Agent 训练里一个被忽略的细节：轨迹的价值取决于它是在哪套 harness 下产生的，而现有共同进化方法把 harness 搜索过程中产生的轨迹当作无差别的回放池^[35]。它要解决的是「数据与运行时错配」：相同的动作在不同的提示格式、工具绑定与错误恢复策略下意义不同，混着用来训练会把噪声当信号。做法上建立一个交替共同进化框架，把 harness 搜索与策略训练解耦成组件级的晋升决策，并提出 CoTrace 这一「harness 感知」的数据配方，明确管理轨迹路由、来源匹配与课程刷新：反复出现的执行失败驱动 harness 合成，而策略训练严格使用与已采用运行时匹配且经过验证的 rollout（SFT 用这些，RL 用新的在线交互）。效果上在 Tmax 晋升划分上，CoTrace 把 Qwen3.5-9B 在监督微调下从解出 78 个任务提升到 88 个，在线强化学习变体达到 90 个。对训练终端或编码 Agent 的团队，参考价值是为每条训练轨迹记录它对应的运行时，并按运行时分组使用；限制是提升来自单一模型与单一划分，跨 harness 与跨任务族是否同样成立需要复现。

CoTrace targets an overlooked detail in terminal-agent training: a trajectory's value depends on the harness under which it was generated, yet existing co-evolution methods treat trajectories from harness search as an undifferentiated replay buffer^[35]. It addresses data-runtime mismatch: the same action means different things under different prompt formats, tool bindings and error-recovery policies, so mixing them treats noise as signal. The work establishes an alternating co-evolution framework that decouples harness search from policy training through component-wise promotion decisions, and introduces CoTrace, a harness-aware data recipe governing trajectory routing, provenance matching and curriculum refresh: recurring execution failures guide harness synthesis while policy training is strictly conditioned on verified rollouts matched to the adopted runtime — supervised fine-tuning on those, reinforcement learning on fresh online interactions. On the Tmax promotion split, CoTrace lifts Qwen3.5-9B from 78 to 88 solved tasks under supervised fine-tuning, with an online RL variant reaching 90. For terminal or coding agent training teams the reference is to record the runtime behind every training trajectory and group by it; the limit is that gains come from one model and one split, so cross-harness and cross-task replication is still needed.

这篇论文问的是可解释性中的一个硬问题：强化学习教会模型一个新行为之后，我们能不能指出是哪条训练轨迹教的？而那些声称能做到的归因方法，怎么知道答案是真的？^[36] 做法上构造了一个「已知病因」的受控实验：在 GRPO 的在线强化学习微调中植入一个行为，然后发布 BehaviorTrace——一个结合全梯度 sketching、植入行为设定，并对梯度幅度、流畅度、可提升空间以及跨随机种子与生成抽样变化做控制的评测工具。结果对现有归因方法是坏消息：在 Qwen2.5-1.5B 的三个随机种子上，相当一部分看似有效的归因信号来自混淆因素——一个只按梯度大小给训练步排序、完全没有行为目标的对照方法就达到了随机水平的 4.2–4.5 倍，并在三分之二种子上追平或超过了最好的定向估计器；在饱和检查点上，模型流畅度预测行为标签的效果不比任何梯度方法差；控制流畅度之后，逐条 rollout 的结果随种子与生成抽样而变，说明单次实验无法定论。唯一在三个种子上都成立的信号是：触发 token 的梯度与「行为实际发生处」构造的目标对齐。对做 RL 微调与可解释性的团队，参考价值是归因结论必须配对照与多种子复现，否则「梯度对齐」很可能只是流畅度与幅度的代理；限制是实验局限于 GRPO、单一模型规模与植入行为，真实训练场景的结论需重新验证。

This paper asks a hard interpretability question: after reinforcement learning teaches a model a new behaviour, can we identify which training rollout taught it — and when an attribution method claims to, how do we know the answer is real?^[36] The approach builds a controlled experiment with a known cause: a behaviour is planted during GRPO online RL fine-tuning, and BehaviorTrace is released — an evaluation harness combining full-gradient sketching with the planted-behaviour setup and controls for gradient magnitude, fluency, headroom, and variation across seeds and generation draws. The results are bad news for existing attribution: across three seeds on Qwen2.5-1.5B, much of the apparent signal comes from confounds — a control ranking training steps by gradient size alone, with no behaviour target, reaches 4.2 to 4.5 times chance and matches or beats the best targeted estimator on two of three seeds; at saturated checkpoints fluency predicts the behaviour label at least as well as any gradient method; and once fluency is controlled, per-rollout results change from seed to seed and draw to draw, so one run cannot settle the question. One signal holds on all three seeds: the gradient of the trigger tokens aligns with a target built where the behaviour actually occurs. For RL fine-tuning and interpretability teams the reference is that attribution claims need controls and multi-seed replication, or gradient alignment may just proxy fluency and magnitude; the limit is that the study is confined to GRPO, one model scale and a planted behaviour.

这篇论文处理评测中最尴尬的一类问题：很多丢给大模型的问题根本没有正确答案可供打分——某项政策值多少、用户该选哪个方案、互相冲突的价值如何权衡^[37]。它借来的是「陈述偏好经济学（stated-preference economics）」应对同样困境几十年的工具箱：不依赖真值也能判断回答是否有效的框架，包含内容效度、构念效度、效标效度、信度、激励相容性与后果性；作者把每个概念逐一映射到 LLM 评测，并用一份已发表的水质陈述偏好估值调查（Vossler et al. 2023）在六个模型上做示范——这里效度检验表现为经济学理论的预测：需求曲线应当向下倾斜，支付意愿应当对物品范围与收入有反应。效果上这些检验把模型区分得很清楚：两个较老的模型在最基础的检验上就失败（家庭收入 75,000 美元这一档），两个最新的模型通过了所有可打分理论效度检验，但在收敛效度上出现分歧；作者特别强调，通过效度检验说明回答是自洽的，而不是正确的。对做评测与产品决策的团队，参考价值是在缺少标准答案的场景改用「理论预测是否成立」来检验模型，而不是强行编造基准；限制是示范只覆盖一份经济学问卷与六个模型，而且这套框架依赖「回答有成本、被认真对待」等前提，聊天式使用场景未必满足。

This paper tackles the most awkward class of evaluation questions: many questions put to LLMs have no correct answer to score against — what a policy is worth, which option a user should choose, how to weigh competing values^[37]. It borrows the toolbox that stated-preference economics has used for decades on the same problem: a framework of validity that judges responses without knowing the true value, covering content, construct and criterion validity, reliability, incentive compatibility and consequentiality. The authors map each concept to LLM evaluation and demonstrate with a published water-quality stated-preference valuation survey (Vossler et al. 2023) administered to six models, where validity tests take the form of theoretical predictions: demand should slope down, and willingness to pay should respond to the scope of the good and to income. The tests separate the models sharply: two older models fail the most basic test at a household income level of $75,000, while the two newest pass every scorable theoretical-validity test but diverge on convergent validity — and the authors stress that passing validity tests shows a model's answers are coherent, not that they are correct. For evaluation and product-decision teams the reference is to test whether theoretical predictions hold when ground truth is unavailable instead of inventing a benchmark; the limit is a demonstration on one economic survey and six models, resting on assumptions such as responses being costly and taken seriously that chat usage may not satisfy.

这篇论文提出了一个前瞻但已经很现实的问题：研究型 Agent 的部署正在走向「上千个 Agent 共享一个算力池」的规模，而现有系统要么一次只组织一个项目，要么干脆不组织^[38]。作者的核心论断是：这样一群 Agent 无论设计者是否提供组织，都会自行长出组织，所以不如显式地设计它，而多智能体社区恰好握有相应工具。做法上提出「Agent 社会」：一群持久存在的 Agent 在明确的制度下运行，并在科学场景中落地为「研究者社会」，建立在六条原则上——课题负责人通过提案、独立评审与资助来竞争算力，一位人类「市长」负责分配资源且不指派任务。效果上，在一个运行中的「一万名研究者」社会里，只被要求改进语言模型预训练，其中一间实验室报告能用约少 30% 的算力达到同等质量，但参与验证的实验室尚未就这一结果达成一致。对做多 Agent 系统的团队，参考价值是规模上去之后，「制度」比单个 Agent 的提示词更决定产出质量，值得提前设计评审与资源分配规则；限制是这是一篇立场与设计论文，核心结果尚未被独立复现（论文本身也如实说明了分歧），六条原则的选择依据需要进一步论证。

This paper raises a forward-looking but already real problem: research-agent deployments are moving toward populations of thousands sharing one pool of compute, while current systems organise one project at a time or not at all^[38]. The central claim is that such a population will acquire an organisation whether or not its designers provide one, so designers should provide it explicitly, and that the multi-agent community already holds the tools. The proposal is a society of agents — a population of persistent agents under explicit institutions — developed for science as a society of researchers built on six principles: principal investigators compete for compute through requests for proposals, independent review and grants, while a human governor, the mayor, allocates resources and assigns no tasks. In a running society of ten thousand researchers asked only to improve language-model pretraining, one lab reported reaching the same quality with about 30% less compute — a result the labs that tested it do not yet agree on. For multi-agent teams the reference is that at scale, institutions matter more than any agent's prompt, so review and resource-allocation rules deserve early design; the limit is that this is a position and design paper whose central result is not yet independently replicated, as the authors themselves note.

EngramEdit 针对「知识编辑」的结构性难题：条件记忆架构（如 DeepSeek Engram）用输入 n-gram 查表扩展模型容量，理论上可以把事实存储与通用计算解耦，从而在不改 Transformer 主干的前提下更新事实知识^[39]。但这条路有两个实际障碍：同一个事实的不同表述会激活不同的 n-gram 嵌入，而更新共享嵌入又会意外改变模型对其它事实的预测。做法上分两步：先计算能让模型在多种表述下都输出更新后事实的目标记忆表示，再联合更新共享 n-gram 嵌入去匹配这些目标（跨表述与跨编辑），并对「被频繁复用的嵌入」施加更强的惩罚以保护无关知识。效果上，EngramEdit 实现了独立的事实更新，编辑成功率达到接近完美，且修订后的知识能在多种表述下生效。对做知识更新与检索增强的团队，参考价值是当模型改成条件记忆架构后，「改知识」终于可以不碰主干权重，这会影响增量更新的架构选择；限制是结论建立在特定条件记忆架构与事实类任务上，对推理型知识与跨语言迁移的效果仍需验证。

EngramEdit addresses a structural difficulty in knowledge editing: conditional memory architectures such as DeepSeek Engram look up learned embeddings by input n-grams, expanding capacity with limited extra computation, and in principle decouple factual storage from general computation so facts can be updated without touching the Transformer backbone^[39]. Two obstacles block that path: different expressions of the same fact activate different n-gram embeddings, while updating shared embeddings can unintentionally change the model's predictions about other facts. The method proceeds in two steps: compute target memory representations that make the model predict the updated fact across multiple expressions, then jointly update shared n-gram embeddings to match those targets across expressions and edits, penalising updates to frequently reused embeddings more strongly to preserve unrelated knowledge. As a result EngramEdit enables independent factual updates with near-perfect editing success, and the revised knowledge holds across expressions. For teams doing knowledge updates or retrieval augmentation the reference is that with conditional memory, editing knowledge no longer requires touching backbone weights, which matters for incremental-update architecture choices; the limit is that results rest on one conditional-memory architecture and factual tasks, leaving reasoning-type knowledge and cross-lingual transfer unverified.

Long-WAM 处理实时机器人控制里的一个矛盾：要推断运动与任务进度就需要足够长的视觉历史，但处理历史会推迟动作^[40]。它的中心发现可以概括成一句话：「能访问历史」不等于「会使用历史」——只有当视频基座是自回归（autoregressive）预训练时，更长的历史才真正带来收益。做法上先从机器人与第一人称视频中学习不含动作标签的因果预测，再在世界-动作模型适配过程中保留这种「历史→未来」结构，并用流式观测编码、异步执行与硬件相关加速来保证实时性。效果很具体：在 RoboCasa GR-1 上，把上下文从 0.0 秒增加到 19.2 秒，成功率从 63.3% 提升到 78.7%，而双向预训练的初始化没有净增益；机器人领域的自回归预训练还能进一步提升 GR-1 与 LIBERO-Long 的峰值成功率，且在 LIBERO-Long、RoboTwin 2.0 与 DOMINO 上优于对比方法；部署侧可在 RTX 5090、DGX Spark 与 Jetson AGX Thor 上运行而不放弃未来预测，在 RTX 5090 上每个动作块（含未来视频潜变量预测）耗时 107.4 毫秒。对做机器人策略与实时系统的团队，参考价值是上下文长度必须与预训练目标匹配，堆历史只有在自回归基座上才划算，同时把延迟预算当硬约束；限制是结论建立在特定仿真基准与自研加速实现上，跨本体与真实高动态环境的可迁移性需要进一步验证。

Long-WAM addresses a tension in real-time robot control: inferring motion and task progress needs enough visual history, but processing that history delays action^[40]. Its central finding fits in one line: access to history is not the same as using it — longer histories pay off only when the video foundation is pretrained autoregressively. The method first learns causal prediction from robot and egocentric video without action labels, then preserves that history-to-future structure during world-action adaptation, using streaming observation encoding, asynchronous execution and hardware-specific acceleration for real-time behaviour. The results are concrete: on RoboCasa GR-1, increasing context from 0.0 to 19.2 seconds raises success from 63.3% to 78.7%, whereas a bidirectionally pretrained initialisation shows no net gain, and robot-domain autoregressive pretraining raises peak success further on GR-1 and LIBERO-Long, also beating compared methods on LIBERO-Long, RoboTwin 2.0 and DOMINO; it deploys on RTX 5090, DGX Spark and Jetson AGX Thor without dropping future prediction, taking 107.4 ms per action chunk including future-video latent prediction on an RTX 5090. For robot policy and real-time systems teams the reference is that context length must match the pretraining objective — piling on history only pays on an autoregressive base — and latency budget is a hard constraint; the limit is that results rest on specific simulation benchmarks and a bespoke acceleration stack, leaving cross-embodiment and high-dynamic real-world transfer to verify.

这篇论文解决的是一个花钱问题：在一轮昂贵的 Agent 后训练之前，怎么判断哪个基础检查点值得投入？^[41] 现有的 pass@K 在这里不好用，因为很多基础模型根本产不出完成端到端任务所需的规范工具调用；而单步或短程任务虽然绕开了工具调用失败，也丢掉了真正要考察的能力——在仓库持续演化时跨多步维持连贯状态。做法上把成功的后训练 Agent 轨迹当作基础模型潜力的「前瞻信号」：重放每条轨迹并在每次改代码的步骤后重跑测试，找出「决定性步骤」——即累计补丁第一次把仓库从失败翻转为通过的那一步，它证明该动作在既有上下文下解决了任务；随后在这一步上构建三个筛选指标（例如 Decisive-Action BPB，用每字节比特数衡量模型对该动作的概率），都不需要基础检查点冷启动驱动 harness。对做 Agent 训练与选型的团队，参考价值是用「能否复现关键一步」来代替端到端奖励作为早期筛选，可以省下大量训练算力；限制是摘要未给出各筛选指标与最终训练收益的相关性数字，且方法依赖已有成功轨迹，冷启动场景如何适用还不清楚。

This paper solves a spending problem: before an expensive round of agentic post-training, how do you tell which base checkpoint is worth it?^[41] End-to-end pass@K is a poor fit because many base checkpoints cannot reliably produce the well-formed tool invocation needed to complete a task end-to-end, while single-shot or short-horizon tasks avoid those failures only by discarding the capability that matters: maintaining coherent state over many tool-using steps as the repository evolves. The method treats successful post-trained agent trajectories as a lookahead signal of base-model potential: replaying each trajectory and rerunning tests after every code-changing step identifies the decisive step — the first step whose cumulative patch flips the repository from failing to passing, certifying that the recorded action solves the task given the prior context; three screens are then built at that step (for example Decisive-Action BPB, bits per byte measuring the model's probability of that action) and none requires the base checkpoint to drive the harness from a cold start. For agent training and model-selection teams the reference is that screening on whether the model can reproduce the decisive step replaces end-to-end reward as an early filter and saves substantial training compute; the limit is that the abstract omits correlation figures between screens and downstream training gains, and the method depends on existing successful trajectories, leaving cold-start applicability unclear.