ScholarPulse 日报 2026-10-01

2026-10-01 学术简报：2 篇。本文提出了由AI智能体协助构建、基于JAX的模块化、可微且多尺度的地球系统模型legoESM。

今日速览

序号	标题	来源	日期	主题	推荐等级
1	legoESM: a modular, differentiable, multiscale, AI-ready Earth system model built with AI agents	arXiv	2026-10-08	AI-Agent	高
2	Memento 3: Model-Based Recursive Self-Improvement through Reflective Rulebooks	arXiv	2026-10-08	AI-Agent	高

重点论文与技术动态

1. legoESM: a modular, differentiable, multiscale, AI-ready Earth system model built with AI agents

来源：arXiv
日期：2026-10-08
作者/机构：Pierre Gentine, Dhruv Balwada, Aytaç Paçal, Linnia Hawkins, Alistair Adcroft, Hang Fan
主题标签：AI-Agent, arXiv
推荐等级：高
分类：physics.ao-ph

一句话结论

本文提出了由AI智能体协助构建、基于JAX的模块化、可微且多尺度的地球系统模型legoESM。

核心内容

模块化与多尺度：物理方案或机器学习模拟器等组件可像积木般自由替换，单一代码库即可覆盖米级至全球、天气至气候尺度的模拟。
端到端可微：支持基于梯度的校准、变分数据同化与在线训练，便于系统评估不同模型变体以探索结构不确定性。
优异性能：跨尺度模拟逼真，利用梯度校准有效降低了陆面温度偏差，且在GPU上可高效扩展至公里级分辨率。

方法与数据

方法：在人类设定的科学契约下，由AI编程智能体将社区长期积累的参数化与数值方法重构为统一的JAX框架，并经基准测试验证。
数据：摘要未明确。

价值判断

值得关注：用AI智能体成功将传统物理模型重构为适配加速硬件（GPU）与现代AI的可微架构。
可复用点：提供地球科学开源社区基础设施，以及多尺度物理系统的建模模板。
局限/待核查：摘要未明确其在解决云辐射反馈等气候响应不确定性方面的具体量化成效。

摘要

Earth system models (ESMs) have grown tremendously in realism, yet key uncertainties persist in the climate response to greenhouse-gas forcing, particularly due to cloud radiative feedbacks. In addition, their software architecture was not designed for accelerator hardware or modern artificial intelligence (AI). Here we present legoESM, a composable, differentiable, multiscale ESM written in JAX. It builds on decades of community-developed parameterizations and numerical methods, recast in a unified framework by AI coding agents under a human-specified scientific contract and verified through benchmarking. Dynamical cores, physics schemes, grids, complexity levels and components are swappable like building blocks, and can use conventional physics or machine-learned emulators. A single code base spans metre-scale large-eddy simulation to global simulations and weather to climate. End-to-end differentiability enables gradient-based calibration, variational data assimilation and online training. legoESM modular architecture enables systematic evaluation of diverse model variants to explore structural uncertainty and test hypotheses. legoESM produces realistic simulations across scales, reduces land-surface temperature bias through gradient-based calibration, and scales efficiently on GPUs to kilometer-scale simulations. It offers an open, community infrastructure for hypothesis testing, research and teaching in Earth sciences and a template for multiscale physical systems.

2. Memento 3: Model-Based Recursive Self-Improvement through Reflective Rulebooks

来源：arXiv
日期：2026-10-08
作者/机构：Haoyu Zhao, Zhengxu Yu, Zhiyuan He, Meng Fang, Rasul Tutunov, Haitham Bou-Ammar
主题标签：AI-Agent, arXiv
推荐等级：高
分类：cs.AI, cs.CL, cs.CV, cs.LG

一句话结论

提出 Memento 3，使冻结的 LLM 智能体通过规则书与代码的反思验证，持续学习显式世界模型并实现递归自我改进。

核心内容

维护自然语言规则书记录环境动态假设，并编译为代码用于预测与规划。
经观察、反思、修订、编译和验证循环，仅在符合规则且精确重放转换时更新代码。
支持种群扩展，并行维护多个世界模型以共享证据并指导探索。

方法与数据

结合固定参数的 LLM 与外部记忆闭环；底层模型具体名称摘要未明确。
测试于 ARC-AGI-3（通关全部25个公开游戏，仅用人类44%动作数）和 Atari Pong（零后续LLM调用以21:0获胜）。

价值判断

值得关注：免微调实现递归自我改进，在 ARC-AGI-3 上动作效率超越人类。
可复用点：“文本规则+编译代码+精确重放验证”的世界模型构建机制。
局限/待核查：摘要未明确底层基座模型型号及种群扩展的具体实验数据。

摘要

Learning to act in unfamiliar environments requires agents to infer how the world works and revise that understanding as new evidence arrives. Yet limited observations can support multiple world models that explain past interactions but predict different outcomes in unseen states. We introduce Memento 3, building on the Memento series to enable frozen LLM agents to continually learn explicit world models through external memory. The agent maintains a natural-language rulebook as persistent semantic memory, recording revisable hypotheses about environment dynamics while leaving unknown aspects underspecified. It compiles this rulebook into executable code for prediction and planning. Through a continual loop of observation, reflection, rule revision, compilation, and verification, the agent uses prediction errors to refine both the rulebook and its code. Updated code is accepted only when the LLM judges it faithful to the rulebook and cell-exact replay reproduces the observed transitions. We investigate this process as a model-based route to recursive self-improvement (RSI): the agent autonomously explores the environment, revises its world model, and uses verified updates to guide subsequent interaction and learning, while the underlying LLM remains fixed. A population extension maintains multiple world models in parallel, sharing interaction evidence and using their predictions to guide exploration. On ARC-AGI-3, the single-model agent clears every level of all 25 public games, achieves a mean Relative Human Action Efficiency (RHAE) of 100.0, and uses 44% of the human action count. In an Atari Pong case study, a learned feedback controller wins 21:0 in each of three evaluated episodes with different openings, without further LLM calls.