Knowledge Flywheels

Yisong Yue 2026-08-06 5.6 KB 138.3K 944 𝕏 文章列表

文章提出了一个新的扩展维度——知识扩展。通过将代理的经验转化为可复用的知识，形成知识飞轮，从而提升智能体的性能和自我改进能力。知识扩展不仅能简化代理，还能训练更明智的模型，为递归自我改进提供新路径。

# Knowledge Flywheels **作者**: Yisong Yue **日期**: 2026-08-05T16:42:27.000Z **来源**: [https://x.com/yisongyue/status/2085043769297277114](https://x.com/yisongyue/status/2085043769297277114) --- ![Image](/doc/3873/image/img_20d82149fb54.jpg) ## A new scaling dimension is emerging Today, we scale models through better data, architectures, and compute. We scale agents through better tools, search, verification, and harnesses. Tomorrow, we will also scale knowledge. Here, knowledge means reusable insights distilled from agent experience: what works, what fails, when, and why. Model scaling improves the reasoner. Agent scaling expands what it can do on a task. Knowledge scaling improves what the next task can build on. ## From experience to knowledge AI systems generate far more experience than they convert into reusable knowledge. Agents discover architectural constraints, recurring failures, better search and verification strategies, and relationships across datasets or experiments. Without curation, these lessons remain tied to individual runs and must be rediscovered. A knowledge flywheel works in two directions: it turns repeated experience into reusable abstractions, then applies them back to new situations. Software work can yield principles for modifying large codebases or optimizing complex stacks; scientific analyses can yield skills for combining evidence, deciding what generalizes, and choosing what to measure next. ![Image](/doc/3873/image/img_0c4d7dc7add0.jpg) Knowledge flywheels distill experience into insights, which leads to smarter agents. Our recent work on knowledge-centric self-improvement tests this idea in a controlled experiment. Generic, disposable agents contribute and critique evidence, which is distilled into a shared knowledge base for future agents. Across reasoning, coding, and terminal tasks, this approach improves performance at substantially lower cost than more elaborate agent-centric self-improvement systems. Early versions of this idea are already emerging in the wild. EinsteinArena lets agents tackle open mathematical problems through repeated submissions, verifier feedback, and refinement of one another’s ideas. This process has already produced new state-of-the-art results, showing how persistent shared artifacts can turn isolated agent runs into cumulative discovery. ## Knowledge simplifies agents Knowledge scaling can move reusable structure out of runtime search. Rather than rediscovering task representations, decompositions, checks, and failure modes on every run, automated harness engineering can begin from accumulated knowledge: task + model + tools + knowledge → task-specific harness The persistent asset is not one harness, but the knowledge from which many harnesses can be constructed. One example is Trace2Skill, which distills lessons from pools of execution traces into transferable, declarative skills. This pattern is also appearing in production. At Asari AI, agents optimize the entire inference stack through repeated cycles of implementation, evaluation, and verification. Each campaign leaves behind improved code and reusable lessons, so later agents inherit a better process and not just a better result. ## Knowledge can train wiser models Knowledge curation also enables on-policy distillation. Models near the edge of their competence generate experience that exposes their weaknesses. Instead of training on raw traces, a knowledge flywheel can compare runs and extract what generalizes: which strategies transfer, which failures recur, what evidence changes the conclusion, and where a rule breaks. That curated knowledge can become targeted supervision for the next model, or be encoded into reusable learned memory through systems such as Cartridges or MeMo. The goal is to teach models something closer to judgment: not just what worked, but when, why, and how broadly the lesson should be trusted. ## A new path to recursive self-improvement Knowledge flywheels offer a different path to recursive self-improvement: the system improves not only its solutions, but how it extracts, tests, and organizes knowledge. Those improvements produce stronger agents, leading to more effective future curation. The need is clearest where evidence exceeds what any one person or agent can synthesize. HuBMAP, the DANDI Archive, and Gaia span vast bodies of scientific evidence; large software ecosystems distribute knowledge across code, hardware, tests, logs, and operational experience. Progress depends on abstractions that connect these pieces. The payoff is compounding synthesis: each analysis improves the next, each failure sharpens future tests, and each abstraction enables a higher level of synthesis. The next scaling frontier is not only larger models or more elaborate agents. It is systems that get better at turning experience into knowledge. ## 相关链接 - [Yisong Yue](https://x.com/yisongyue) - [@yisongyue](https://x.com/yisongyue) - [11K](https://x.com/yisongyue/status/2085043769297277114/analytics) - [EinsteinArena](https://einsteinarena.com/) - [Trace2Skill](https://arxiv.org/abs/2603.25158) - [Cartridges](https://arxiv.org/abs/2506.06266) - [MeMo](https://arxiv.org/abs/2605.15156) - [Upgrade to Premium](https://x.com/i/premium_sign_up) - [12:42 AM · Aug 6, 2026](https://x.com/yisongyue/status/2085043769297277114) - [11.5K Views](https://x.com/yisongyue/status/2085043769297277114/analytics) - [View quotes](https://x.com/yisongyue/status/2085043769297277114/quotes) --- *导出时间: 2026/8/6 18:08:49* --- ## 中文翻译 # 知识飞轮 **作者**: Yisong Yue **日期**: 2026-08-05T16:42:27.000Z **来源**: [https://x.com/yisongyue/status/2085043769297277114](https://x.com/yisongyue/status/2085043769297277114) --- ![Image](/doc/3873/image/img_20d82149fb54.jpg) ## 一个新的扩展维度正在出现今天，我们通过更好的数据、架构和算力来扩展模型。我们通过更好的工具、搜索、验证和约束来扩展智能体。明天，我们还将扩展知识。在这里，知识指的是从智能体经验中提炼出来的可重用见解：什么有效、什么无效、何时有效以及为何有效。模型扩展提升了推理者。智能体扩展扩展了其在任务上能做的事情。知识扩展提升了下一个任务所能构建的基础。 ## 从经验到知识 AI 系统产生的经验远多于其转化为可重用知识的部分。智能体发现了架构约束、反复出现的失败、更好的搜索和验证策略，以及跨数据集或实验的关系。如果没有整理，这些经验仍然局限于单独的运行中，必须被重新发现。知识飞轮在两个方向上起作用：它将重复的经验转化为可重用的抽象，然后将其应用回新的情况。软件工作可以产生修改大型代码库或优化复杂堆栈的原则；科学分析可以产生结合证据、决定什么具有普遍性以及选择接下来测量什么的技能。 ![Image](/doc/3873/image/img_0c4d7dc7add0.jpg) 知识飞轮将经验提炼为见解，从而产生更智能的智能体。我们最近关于以知识为中心的自我改进工作在一个受控实验中测试了这个想法。通用的、一次性的智能体贡献并评判证据，这些证据被提炼成供未来智能体使用的共享知识库。在推理、编码和终端任务中，这种方法以比更复杂的以智能体为中心的自我改进系统低得多的成本提高了性能。这个想法的早期版本已经在现实中出现。EinsteinArena 允许智能体通过反复提交、验证器反馈和改进彼此的想法来解决开放的数学问题。这个过程已经产生了新的最先进结果，展示了持久的共享工件如何将孤立的智能体运行转化为累积发现。 ## 知识简化了智能体知识扩展可以将可重用的结构从运行时搜索中移出。自动化约束工程可以从积累的知识开始，而不是在每次运行中重新发现任务表示、分解、检查和失败模式：任务 + 模型 + 工具 + 知识 → 特定任务的约束持久的资产不是某一个约束，而是可以构建许多约束的知识。一个例子是 Trace2Skill，它将执行轨迹池中的经验教训提炼为可迁移的、声明式的技能。这种模式也出现在生产环境中。在 Asari AI，智能体通过实施、评估和验证的重复周期来优化整个推理堆栈。每次活动都会留下改进的代码和可重用的经验教训，因此后来的智能体继承的是一个更好的过程，而不仅仅是一个更好的结果。 ## 知识可以训练更明智的模型知识整理还支持在策略蒸馏。处于其能力边缘附近的模型产生的经验暴露了它们的弱点。知识飞轮可以比较运行并提取具有普遍性的内容：哪些策略可迁移，哪些失败会反复出现，什么证据会改变结论，以及规则在哪里失效，而不是在原始轨迹上进行训练。那些整理过的知识可以成为下一个模型的有针对性的监督，或者通过 Cartridges 或 MeMo 等系统编码到可重用的学习记忆中。目标是教给模型更接近判断力的东西：不仅仅是什么有效，还包括应该在何时、何地以及多大程度上信任这个经验教训。 ## 递归自我改进的新路径知识飞轮为递归自我改进提供了一条不同的路径：系统不仅改进其解决方案，还改进其提取、测试和组织知识的方式。这些改进产生更强的智能体，从而带来更有效的未来整理。在证据超过任何一个人或智能体所能综合的范围时，这种需求最为明显。HuBMAP、DANDI Archive 和 Gaia 涵盖了大量的科学证据；大型软件生态系统将知识分布在代码、硬件、测试、日志和运营经验中。进步取决于连接这些部分的抽象。回报是复合的综合：每次分析都改善下一次分析，每次失败都完善未来的测试，每个抽象都使更高层次的综合成为可能。下一个扩展前沿不仅是更大的模型或更复杂的智能体。而是那些更善于将经验转化为知识的系统。 ## 相关链接 - [Yisong Yue](https://x.com/yisongyue) - [@yisongyue](https://x.com/yisongyue) - [11K](https://x.com/yisongyue/status/2085043769297277114/analytics) - [EinsteinArena](https://einsteinarena.com/) - [Trace2Skill](https://arxiv.org/abs/2603.25158) - [Cartridges](https://arxiv.org/abs/2506.06266) - [MeMo](https://arxiv.org/abs/2605.15156) - [Upgrade to Premium](https://x.com/i/premium_sign_up) - [12:42 AM · Aug 6, 2026](https://x.com/yisongyue/status/2085043769297277114) - [11.5K Views](https://x.com/yisongyue/status/2085043769297277114/analytics) - [View quotes](https://x.com/yisongyue/status/2085043769297277114/quotes) --- *导出时间: 2026/8/6 18:08:49*