Homepage
Delving into the latent unknown.
Mao Song's technical blog covering machine learning, large language models (LLMs), deep learning research, and AI innovations.
START HERE
Featured Tutorials
Foundational guides for building a reusable mental model.
- Learning Notes Ready to ReadReinforcement Learning for Large Language Models: An Overview
从强化学习基础到大模型训练，串起价值估计、策略优化与关键推导。
- Learning Notes Ready to ReadLLM Resource Computation: Parameters, FLOPs and Memory
理解大模型参数量、计算量与显存开销，建立训练资源估算的基础。
- Learning Notes Ready to ReadHands on LLM (1) Tokenizer
从文本到 token：理解分词方法，以及词表与编码的基本取舍。
Deep Dives
Question-driven derivations and comparisons. These articles are in progress; their status reflects the current reading experience.
- Deep Dives In ProgressKL divergence from machine learning to reinforcement learning
围绕 KL 散度的估计与优化展开推导，辨析不同形式的含义和适用条件。
- Deep Dives In ProgressTraining Inference Mismatch
分析训练与推理不一致如何影响策略优化，梳理误差来源与修正思路。
Explore by Content Type
What do these statuses mean?
Status describes reading readiness within the current scope. Ready to Read articles may still receive updates.
Ready to Read A coherent treatment of the current scope, ready for end-to-end reading.
In Progress Substantial content is available; sections, derivations, or validation are still being developed.
Outline An outline, source collection, or early notes.
News
- I have joined the SenseTime as a Researcher
I have joined the SenseTime as a Researcher. I will continue to focus on post training of LLMs.
- Make this blog framework as a blog template
A blog template for research articles and blog posts.
- Update the blog framework from Hugo to Astro
I update the blog framework from Hugo to Astro. Now the blog is more flexible and easier to maintain.
- Investigating Redundancy in Multimodal Large Language Models with Multiple Vision Encoders
We invalidate the hypothesis that 'adding more vision encoders will always improve the performance of multimodal large language models'. Accetpted by ICLR2026.