John Langford (@JohnCLangford) / X

John Langford

317 posts

John Langford

@JohnCLangford

Solving Machine Learning at Microsoft in New York. icml.cc pandemic past president. vowpalwabbit.org makes RL real. hunch.net for thinking out loud.

New York (mostly)

Joined December 2019

John Langford
@JohnCLangford
Mar 23, 2020
ICML 2020 (icml.cc/Conferences/20…) will be a virtual conference. We've been having discussions with ICLR folks about their plans, and will be learning from their experience (they are up first). We hope to enable as much of the normal ICML experience as possible, virtually.
John Langford
@JohnCLangford
Sep 12, 2020
We've made a decision to plan for a virtual ICML 2021. This is the only choice for which it is possible to make a plan given pandemic-driven uncertainty. If the pandemic dissipates in some places, organization of local meetups may make sense.
John Langford
@JohnCLangford
Jan 4, 2020
I finally gave in and made a twitter account. I'm planning to keep the hunch.net Machine Learning (Theory) blog for longer form discussions, but I've often found myself wanting to discuss relatively short things for which Twitter seems more natural than a blog.
John Langford
@JohnCLangford
Nov 9, 2020
Alekh Agarwal, Akshay Krishnamurthy, and I finished a major tutorial on "Theoretical Foundations of Reinforcement Learning" (hunch.net/~tforl/) for FOCS (focs2020.cs.duke.edu/program/), but potentially of much broader interest. We'll be doing Q&A Friday.
John Langford
@JohnCLangford
Aug 9, 2021
My term as president of ICML has finished---Francis Bach is taking over. The pandemic was the defining element of the last 2 years. This was stressful for everyone, but we coped and learned a fair bit about what is possible remotely. Looking for ICML 2022 in Baltimore...
John Langford
@JohnCLangford
Mar 9, 2022
The ICML board voted today to continue with a 'hybrid' plan for ICML 2022, with an emphasis on in-person and accommodations for remote folks. I really hope to to see many friends in person.
John Langford
@JohnCLangford
Jul 30, 2025
My list of papers I learned about at ICML in case of broader interest: Dynamic Chunking arxiv.org/abs/2507.07955 . Transformer with a learned (de)tokenizer offers plausibly superior per-parameter and per-FLOP performance at small scales.
arxiv.org
Dynamic Chunking for End-to-End Hierarchical Sequence Modeling
Major progress on language models (LMs) in recent years has largely resulted from moving away from specialized models designed for specific tasks, to general models based on powerful architectures...
24K
John Langford
@JohnCLangford
Jan 20, 2020
We are hiring many positions (aka.ms/rl_hiring) to make Reinforcement Learning work for real world applications (see Personalizer aka.ms/personalizer for example). Consider joining us to help make the future and please pass to anyone who might be interested 😀
microsoft.com
Reinforcement Learning - Microsoft Research
The reinforcement learning research group develops theory, algorithms & systems for solving real world problems involving learning from feedback over time.
John Langford
@JohnCLangford
Nov 29, 2022
We (the Real World Reinforcement Learning group at Microsoft Research) are looking to hire, even in a year of serious economic uncertainty: aka.ms/rl_hiring , including intern, postdoc, and researcher positions. Please apply by Dec. 9th if this matches your interest.
microsoft.com
Reinforcement Learning - Microsoft Research
The reinforcement learning research group develops theory, algorithms & systems for solving real world problems involving learning from feedback over time.
John Langford
@JohnCLangford
Jul 14, 2020
In the last year Microsoft has hired RL-related researchers Jordan Ash, Ching-an Cheng, Yonathan Efroni, @ShamKakade6, @criticalneuro, Aadirupa Saha, Karan Singh, @qingyun_wu, Cyril Zhang. If interested, we are still hiring (aka.ms/rl_hiring), looking for complements.
microsoft.com
Reinforcement Learning - Microsoft Research
The reinforcement learning research group develops theory, algorithms & systems for solving real world problems involving learning from feedback over time.
John Langford
@JohnCLangford
Jan 12, 2021
Some decisions for ICML from the board: ICML General Chairs: 2022: Kamalika Chaudhuri @kamalikac 2023: Andreas Krause @arkrause ICML 2022 Program Chairs: Csaba Szepesvari @CsabaSzepesvari, Le Song @dasongle, and Stefanie Jegelka (maybe @StefanieJegelka )
John Langford
@JohnCLangford
Aug 12, 2025
Orthonormal updates appear to roughly double transformer training convergence with Dion paving tractability at the largest scale. Code: github.com/microsoft/dion Paper: arxiv.org/abs/2504.05295
Microsoft Research
@MSFTResearch
Aug 12, 2025
Dion is a new AI model optimization method that boosts scalability and performance over existing leading methods by orthonormalizing only a top rank subset of singular vectors, enabling more efficient training of large models such as LLaMA-3 with reduced overhead:
GitHub - microsoft/dion: Dion optimizer algorithm
From github.com
19K
John Langford
@JohnCLangford
Nov 23, 2020
Many open Reinforcement Learning related positions are here: aka.ms/rl_hiring . Please consider applying if you have RL-complimentary expertise and want to discover and build the future.
microsoft.com
Reinforcement Learning - Microsoft Research
The reinforcement learning research group develops theory, algorithms & systems for solving real world problems involving learning from feedback over time.
John Langford
@JohnCLangford
Aug 9, 2021
We have strong evidence now that the existing virtual conference formats/program/structure are just not as compelling as in-person conferences. Can that gap be closed further? What are the most compelling remote experiences you have had over the last 2 years?