
Reinforcement Learning in LLMs - Why and How
From imitation to optimization: when LLMs need RL, how verifiable rewards unlock reasoning, and a minimal GRPO playbook.

From imitation to optimization: when LLMs need RL, how verifiable rewards unlock reasoning, and a minimal GRPO playbook.
How we built a custom markdown pipeline that handles LaTeX math, image galleries, and rich embeds while keeping content in plain .md files—no MDX required.

A living collection of advice from mentors, friends, and books.

Concentration of measure pushes Gaussian samples onto a thin shell—here's the intuition, the math, and why typicality matters for generative models.

This note provides a high-level summary of the progress in large language models (LLMs) covering major milestones from Transformers to ChatGPT. The note serves as a fast-paced recap for readers to catch up on this field quickly.

Exponential-min and Gumbel-max tricks for reformulating sampling from a discrete distribution as argmin and argmax, making the sampling operation differentiable.

A quick walk-through of Expectation-Maximization (EM) algorithm and its cousins.

PPO made RLHF work; DPO made it simple. This post derives DPO from PPO, explains why it’s a supervised alternative (not RL), where it shines, and where RL/GRPO still helps.

You know how to differentiate through a function—but how do you differentiate through a sampling step? Two estimators: score‑function (REINFORCE) and pathwise (reparameterization); pathwise backpropagates through the sampling transform with lower variance.

Can we speed up generation without changing the distribution? A small draft model proposes, the big model accepts/rejects—yielding exact samples, faster.

Why that particular sigmoid in logistic regression? This short post shows how simple moment constraints lead to exponential families (MaxEnt chooses the model) and how MLE fits them.

A living list of books, essays, and videos that helps me keep perspective on life.

Music notation gives us a tidy grid of notes, but physics delivers a messy spectrum of vibrations. Here's why tuning is always a compromise.

A short list of interview preparation resources for Data Scientists, Machine Learning Engineers, Machine Learning Scientists, Quant Developers and Quant Researchers.

A quick riff on Hume's is–ought gap—why facts don't dictate values, and how the leap from 'is' to 'ought' rests on sentiment.

There has been a lot of confusing information about dopamine. I finally found a literature review-style article, and here is what I learned.

A January dash to London to finally see Jay Chou live, with museums, parks, and good meals along the way.

In October 2021, we spent two weeks traveling to various cities in Italy, including Rome, Cinque Terre, Florence, Tuscany, and Venice. This was our first trip to Italy, and we have documented our journey with a report and photos.

Trip report (itinerary and photos) from our recent trip to southern Utah (Zion, Arches, Canyonlands and Bryce).

This is the first post of hopefully a series of post walking through diffusion models. This post will introduce the foundations, focusing on two foundational papers, that many other papers built upon.

This is a quick note to discuss a few topics below related to building LLM-powered products and applications, such as how to let LLM use tools and become autonomous agents, how to incorporate domain adaptation, and the production hurdles.

In this note, we'll take a look at how Auto-GPT work and discuss LLM's ability to do explicit reasoning and to become an autonomous agent. We'll touch upon a few related works such as WebGPT, Toolformer, and Langchain.

Building Next.js app with Firebase authentication on the client-side, as well as using it on the server-side with a middleware pattern similar to Express.js.

This page is a high-level summary / notes of various recent results in language modeling with little explanations

A list of starter resources for Natural Language Processing (NLP), mostly with deep learning.

A literature survey of recent papers on Neural Variational Inference (NVI) and its application in topic modeling.

A high-level summary of various generative models including Variational Autoencoders (VAE), Generative Adverserial Networks (GAN), and their notable extentions and generalizations, such as f-GAN, Adversarial Variational Bayes (AVB), Wasserstein GAN, Wasserstein Auto-Encoder (WAE), Cramer GAN and etc


A quick reference tying each hiragana and katakana character back to its Chinese origin, plus the patterns and mnemonics I lean on while studying.

How to add a table of contents in Ghost without editing the site template

EmailOctopus form is a script tag, this post shows how to make that work with React (using useEffect and useRef).

My plan and progress updates on learning web frontend development more or less from scratch. Will be semi-regularly updated.

My takeaways from State of JS 2020 survey.

Automatically add preview / teaser content for Member-Only posts in Ghost.

This is a demo for "How to Enable Preview for Member-Only Content in Ghost"

Local development setup for Ghost themes.

Making featured posts show up first in Ghost Casper theme (instead of the default reverse chronological order).

I recently switched to Ghost to host my blog, Here are some editing tips as I learn to use this platform.

My onboarding experience with Ghost and some wishlist items for future improvements.
Testing error handling for malformed gallery and link-preview blocks
Testing core markdown syntax including headers, lists, tables, links, and formatting
Validates that backticked HTML comments and tags render literally as code.
Testing code syntax highlighting with various programming languages and code block types
Testing iframe embedding for videos, maps, and other external content
Comprehensive test suite for both auto-detected sequential image galleries and explicit gallery markers, testing various image counts and edge cases.
Comprehensive test suite for MathJax rendering with various equation types, LaTeX syntax, and edge cases to verify the markdown pipeline's math processing.
Testing link preview card functionality for external URLs

End of posts • 49 posts