陆昱宽

Most modern world models represent environment dynamics using neural networks, including recurrent models, Transformers, and diffusion models. An alternative research direction is to represent world models as executable programs.

Such models, referred to as code world models (CWMs), are constructed through LLM-based program synthesis. In this paradigm, the LLM does not directly act as the world simulator. Instead, it synthesizes executable code that implements the simulator, including the environment's state transitions, observations, rewards, and other relevant dynamics.

We will discuss three representative CWM works:

1. WorldCoder (2024)
2. DeepMind CWM (2025)
3. PoE-World (2025)

We focus primarily on WorldCoder, while the latter two are used to complement the discussion with more advanced formulations and design choices and we don't talk about the details of them.

最近，一篇名为 Stealing Reasoning Traces from Proprietary LLM APIs 的文章引发了广泛讨论。顾名思义，这篇文章提出了一种从闭源 LLM API 中提取 reasoning trace（思维链）的方法。

这篇文章挺有意思的，虽然有些标题党，但给很多人增长了见识，这里就简单聊聊。

介绍英语发音，包括音标、英语元音以及特殊发音规则。本文以通用美式英语（General American, GA）为例。

参考资料：

1. General American from teflpedia
2. General American from wikipedia

Youtube channels are also helpful, such as [@rachelenglish](https://www.youtube.com/@rachelsenglish).

Sources:

1. DiT 2022 paper

Also, it's recommended to read my DDPM and Improved DDPM post to fully understand the context of diffusion models.

Sources:

1. original diffusion model 2015 paper by Jascha Sohl-Dickstein
2. DDPM 2020 paper
3. Improved DDPM 2021 paper

这两年代理软件管得比较严，但 ChatGPT 对科研和学习又不可或缺。通常的解决办法是找更稳定的代理节点，也就是俗称的“机场”。不过，如果你手上有一台本身已经能访问外网的远程服务器（别管是怎么来的），也可以换一种思路：让本机通过 SSH 连接这台服务器，再把它作为本机的网络出口。这样一来，本机访问网站时，流量会先走到服务器，再由服务器访问目标网站。

Straight-through estimator, or STE, is a trick for training models that contain non-differentiable operations, notably the quantization operation commonly used in VQ-VAE.

The core idea is:

Use the non-differentiable operation in the forward pass, but replace its gradient with a simple surrogate in the backward pass.

Model Predictive Control (MPC) is a model-based planning method widely used in model-based reinforcement learning.

Sources:

1. Sutton and Barto. Reinforcement Learning: An Introduction.
2. Hansen et al. Temporal Difference Learning for Model Predictive Control.
3. Scannell et al. Discrete Codebook World Models for Continuous Control, ICLR 2025.

Straight-through Gumbel-softmax is a trick for training models with discrete latent variables, similar to the famoud straight through estimator (STE) trick used in VQ-VAE.

The problem is simple:

We want to sample a discrete category in the forward pass, but we still want gradients to flow through it in the backward pass.

本文介绍 DJI Mic Mini 一代的使用方式，对二代及其他版本也适用。

Sources:

1. 大疆官方的 Mic Mini 使用方法讲解
2. DJI Mic Mini
3. DJI Mic Mini 常见问题