Overview
This reading group studies frontier research in large language models. Topics include machine learning theory, reasoning, verification, post-training, optimization, empirical scaling, and the interaction between AI and quantitative research.
The group is organized by Columbia Quant LLM and is currently inactive Columbia University Computer Science. The main purpose is to learn, understand LLMs, and approximate truth. Serious research projects may emerge naturally from the discussion.
We welcome self-introductions and active sharing of work related to the group’s ultimate goal. Any relevant theory work, empirical result, engineering note, seminar page, benchmark, replication attempt, or frontier research discussion is welcome.
Motivation
The group was partly motivated by discussions around the Anthropic Fellowship Program and by the need for a Columbia-centered community that studies LLMs seriously across theory, empirical practice, and frontier systems.
The group takes inspiration from open course communities and public course pages, including Stanford CS336, Stanford CS229, and Columbia machine learning theory pages.Specially acknowledgement Columbia Theory Group
Focus Questions
We are especially interested in understanding what is currently correct and reliable in frontier LLM practice. Here, “correct” means not merely elegant, popular, or locally successful, but verified at meaningful scale in industry or supported by strong empirical evidence.
- Where is the frontier of OPD?
- What should a serious post-training harness contain?
- Which post-training methods have been validated at large scale?
- How should theory tools help explain reasoning, alignment, verification, and scaling?
- How should academic research distinguish toy success from industry-verified behavior?
Research Taste
The group aims for serious, technically grounded, and publication-oriented discussion. Our reference standard for research taste and clarity includes strong top-conference work, especially the ICLR oral paper listed below.
- Reference standard: ICLR oral paper
- Featured reading: Ruiqi Zhang et al. — a major reading object for the group and a Citadel Paper Award work.
The final ambition is to do serious and cool research. If no faculty supervisor or clearly dominant contributor is involved in a project, contributors should be treated with equal academic respect. Authorship and credit should follow actual intellectual contribution.
Model Systems
In addition to academic papers, the group follows frontier model systems and technical reports from major industrial labs and open model communities.
Kimi / Moonshot AI
Reasoning, coding, agentic intelligence, long-context, and post-training practice.
Moonshot AI GitHubOpen Course Communities
Course pages and public notes are treated as serious infrastructure for learning.
Stanford CS336Site Structure
This site is organized like a public course page. Schedule, notes, projects, and people are separated into linked pages for easier maintenance.
- Schedule — weekly topics, readings, and discussion themes.
- Course Notes — external links to notes, slides, seminar pages, and reading summaries.
- People — consultants, contributors, and community background.
- Projects — possible project directions and contribution model.