Lukas Schäfer

19 June 2025

Our textbook “Multi-Agent Reinforcement Learning: Foundations and Modern Approaches” has sold out! Another print round with minor corrections is in production with @mitpress.bsky.social and coming 👏 An errata of these corrections and the updated book PDF can already be found at www.marl-book.com

Lukas Schäfer@lukaschaefer.bsky.social

📚🧵1/7 It is finally here!! Only one more week until the print release of our textbook “Multi-Agent Reinforcement Learning: Foundations and Modern Approaches” with @mitpress.bsky.social! What you get, why you should be interested and more, all below in a short 🧵👇

Excited to share some news! After two wonderful years at MSR, I will be returning to academia as a tenure-track assistant professor at the University of Southern Denmark (SDU) for their new Vejle campus in January 2027. Until then, I'm taking some time to recharge to dive into it with fresh energy!

A big thank you to everyone I worked with at Microsoft, it's been a lot of fun and I learned a lot during these two years. Special thanks to Sergio Valcarcel Macua and @katjahofmann.bsky.social for their continued support and mentorship, and for helping me navigate faculty applications!

14 July

This looks awesome! We need benchmarks that consider coordination between agents. Also super cool to see cross comparison between LLM/ VLM and MARL agents! Congrats to @kale-ab.bsky.social & co. 👏

Kale-ab Tessera@kale-ab.bsky.social

Can LLM agents coordinate in long-horizon, open-ended worlds? We test 13 LLMs in Alem, a new benchmark. Most struggle, averaging ~6% normalised return. Yet on Hard, zero-shot Gemini 3.1 Pro matches the best MARL agent after 1B training steps. Our ablations show communication matters most. 🧵

9 July

At @icmlconf.bsky.social and still some energy left? If you’re interested in imitation learning/ world action models, checkout our poster this afternoon 2:30pm where we explain - theoretically and empirically - how future modelling introduces an explicit trade off that can improve performance! 👇

Lukas Schäfer@lukaschaefer.bsky.social

I will be at @icmlconf.bsky.social in Seoul over the next week - first time in South Korea and at ICML so looking forward to an exciting week! I'll be co-presenting our work on PIDMs for offline imitation learning - Poster Session 7 - Hall A #304 - Thu, Jul 9 at 2:30pm - details 👇 See you there!

msft.itRethinking imitation learning with Predictive Inverse Dynamics ModelsThis research looks at why Predictive Inverse Dynamics Models often outperform standard Behavior Cloning in imitation learning. By using simple predictions of what happens next, PIDMs reduce ambiguity...

1 July

Lukas Schäfer@lukaschaefer.bsky.social

Super excited to be able to share more about this project! At MSR, we have looked into PIDMs as alternative architectures for imitation learning to understand when and why they offer promise! Learn more at the blog post: msft.it/6047QMkd3 and the arXiv preprint: arxiv.org/abs/2601.21718

luma.comAI for Games - ICML Social · LumaAI for Games [Notice: You must be registered for ICML Tutorials to join this event.] Games have been a proving ground for AI since the earliest game-playing…

1 July

On Monday evening, I'll also be at the AI for Games social where I'll give a talk about our research in games at MSR and join for the panel! First panel discussion so that should be exciting 👀 You can register via 👉 luma.com/wf7pe0ee

1 July

Lastly, you can find me, Sergio Valcarcel Macua, and Pallavi Choudhury on Wed July 8, 11-11:30am at the Microsoft booth! We'll have a Q&A and we can talk world models, world action models, imitation learning, and more. Please come by and chat if this sounds interesting to you!

9 Feb

Microsoft Research@msftresearch.bsky.social

This research looks at why Predictive Inverse Dynamics Models often outperform standard Behavior Cloning in imitation learning. By using simple predictions of what happens next, PIDMs reduce ambiguity and learn from far fewer demonstrations. Learn more: msft.it/6018QMkdO

9 Feb

In short, our paper derives new theory that explains when and why PIDMs show promise, especially in the low-data regime where BC often fails, and validates this theory in small-scale and realistic experiments. Checkout the preprint for more details!

cohere.comCohere Labs - Lukas Schäfer, Postdoc Researcher, Microsoft ResearchCohere Labs - Lukas Schäfer - Decision-Making in Modern Video Games: From Human Play to World Models

19 Jan

I'll be giving a talk at Cohere Labs about my research on decision making agents and world models at Microsoft Research! 📅 January 20th (Tomorrow!) 5pm GMT/ 6pm CET/ 9am PST 📍 Online (link below) Everyone is welcome! Link via calendar invite and more details: cohere.com/events/coher...

7 Dec 2025

It’s a wrap for me. Thanks a lot to the organisers for putting #EurIPS together, it’s been a fantastic week and one of my favourite conference experiences. Looking forward to seeing how this experiment will continue!🤞

apply.careers.microsoft.comResearch Intern - Machine Learning - People Centric AI | Microsoft CareersIn collaboration with your mentor and a diverse team, contribute to solving an ambitious research challenge and translate your results into actionable insights that are relevant to potential applicati...

2 Dec 2025

📣We're hiring new research interns for 2026 at MSR Cambridge! If you're interested in ML research (esp. generative AI and/ or decision making agents), please consider applying. It's a great collaborative environment with a very kind and capable team! apply.careers.microsoft.com/careers/job/...

I'm in Copenhagen for #EurIPS this week, please reach out if you'd like to chat about decision-making research! I'm also on the academic job market for assistant professorship positions in Europe - if you believe I could be a good fit, please reach out, I'd love to connect and chat over☕!

View full thread

Tmrw (Tue 2nd Dec, 13:50-14:10), I'll be giving a talk on this topic titled "Exploiting State and Action Uncertainty for Imitation Learning using Inverse Dynamics Models" at the Interactive Learning and Interventional Representations (ILIR) workshop of the ELLIS UnConference!

On Saturday (6th Dec), I'll be at the EurIPS workshop on Epistemic Intelligence in Machine Learning where I'll be presenting a poster on the same topic! If this sounds interesting to you, please pop by and let us know what you think!

I'll also be presenting work from my postdoc at MSR Cambridge in collaboration with the Game Intelligence team and MSR NYC. In our work, we dug deeper into recent imitation learning algorithms to theoretically & empirically understand what makes them work. You can find me talking about this 👇

Lukas Schäfer reposted

Kale-ab Tessera

@kale-ab.bsky.social

26 Nov 2025

Thrilled to present HyperMARL at #NeurIPS2025 in San Diego next week! 🚀 (Amos will present at @euripsconf.bsky.social too.) TL;DR: Coupling obs and agent IDs can hurt performance in MARL. Agent-conditioned hypernets cleanly decouple grads and enable specialisation. 📜: arxiv.org/abs/2412.04233

Kale-ab Tessera@kale-ab.bsky.social

📜🤖 Can a shared multi-agent RL policy support both specialised & homogeneous team behaviours -- without changing the learning objective, requiring preset diversity levels or sequential updates? Our preprint “𝘏𝘺𝘱𝘦𝘳𝘔𝘈𝘙𝘓: 𝘈𝘥𝘢𝘱𝘵𝘪𝘷𝘦 𝘏𝘺𝘱𝘦𝘳𝘯𝘦𝘵𝘸𝘰𝘳𝘬𝘴 𝘧𝘰𝘳 𝘔𝘶𝘭𝘵𝘪-𝘈𝘨𝘦𝘯𝘵 𝘙𝘓” explores this!

@eugenevinitsky.bsky.social

20 Oct 2025

It was great to visit @sheffielduni.bsky.social today to give an invited talk on work from my PostDoc at MSR, and have a chance to talk to students and faculty at the CS department. Thanks a lot to Robert Loftin for the kind invitation and for hosting me!

Lukas Schäfer reposted

Kale-ab Tessera

@kale-ab.bsky.social

3 Aug 2025

🇨🇦 Heading to @rl-conference.bsky.social next week to present HyperMARL (@cocomarl-workshop.bsky.social) and Remember Markov (Finding The Frame Workshop). If you are around, hmu, happy to chat about Multi-Agent Systems (MARL, agentic systems), open-endedness, environments, or anything related! 🎉

Kale-ab Tessera@kale-ab.bsky.social

Lukas Schäfer reposted

Eugene Vinitsky 🍒

10 July 2025

Will be at ICML and looking to hire a postdoc to help us scale up and deploy RL in self-driving. So, hit me up to chat.

10 July 2025

The Edinburgh RL Reading group is back with a fresh new website 👏 Anyone is welcome to attend!

RL & Agents Reading Group@rl-agents-rg.bsky.social

Hello world! This is the RL & Agents Reading Group We organise regular meetings to discuss recent papers in Reinforcement Learning (RL), Multi-Agent RL and related areas (open-ended learning, LLM agents, robotics, etc). Meetings take place online and are open to everyone 😊

arxiv.orgRobust Autonomy Emerges from Self-PlaySelf-play has powered breakthroughs in two-player and multi-player games. Here we show that self-play is a surprisingly effective strategy in another domain. We show that robust and naturalistic drivi...

23 June 2025

Eugene is awesome! — if you are interested in autonomous driving and RL, and New York sounds like an exciting place for a postdoc then this is an amazing opportunity! 👇

Eugene Vinitsky 🍒@eugenevinitsky.bsky.social

Hiring a postdoc to scale up and deploy RL-based planning onto some self-driving cars! We'll be building on arxiv.org/abs/2502.03349 and learn what the limits and challenges of RL planning are. Shoot me a message if interested and help spread the word please! Full posting to come in a bit.

Lukas Schäfer reposted

Eugene Vinitsky 🍒

@eugenevinitsky.bsky.social

21 June 2025

19 June 2025

Lukas Schäfer@lukaschaefer.bsky.social

github.comGitHub - marl-book/marl-book-exercises: Code exercises for the MARL TextbookCode exercises for the MARL Textbook. Contribute to marl-book/marl-book-exercises development by creating an account on GitHub.

19 June 2025

We also updated and moved our code exercises which we had built for a summer school to the marl-book github page so they can be easily found at github.com/marl-book/ma... Thanks @sacha2.bsky.social for the reminder on that! They had previously been hidden inside my GitHub account 😅

Lukas Schäfer reposted

21 May 2025

Today, I’ll be presenting our work on exploration in MARL using ensembles here! 👇 Multiagent Learning Session Where: Ambassador Ballroom 1 & 2 When: 14:00 - 14:13 I’ll also present the poster later at the Learn track of the poster session at 15:45 - 16:30

Lukas Schäfer@lukaschaefer.bsky.social

At the main conference, I'll be presenting our work on using ensembles of value functions for multi-agent exploration! I'll be presenting the oral at the Multi-agent Learning 1 session on Wednesday (2:00 - 3:45pm), and the poster after 3:45pm! Paper: arxiv.org/abs/2302.03439

On my way to Detroit for @aamasconf.bsky.social! Looking forward to presenting the last work from my PhD at the main conference, and work from @msftresearch.bsky.social at the Adaptive and Learning Agents Workshop. More info 👇 If you'd like to chat, feel free to DM me!

View full thread

github.comGitHub - microsoft/imitation_learning_in_modern_video_games: Accompanying code for "Visual Encoders for Data-Efficient Imitation Learning in Modern Video Games" publicationAccompanying code for "Visual Encoders for Data-Efficient Imitation Learning in Modern Video Games" publication - microsoft/imitation_learning_in_modern_video_games

Paper: arxiv.org/abs/2312.02312 It took some time but I'm super excited that now our code is also open-source and available for everyone to use at: github.com/microsoft/im...

Thanks to all the co-authors and collaborators! Logan Jones, Anssi Kanervisto, Yuhan Cao, Tabish Rashid, Raluca Georgescu, David Bignell, Siddhartha Sen, Andrea Treviño Gavito, and first and foremost Sam Devlin It's been an absolute joy working with this group of kind folks 👏

View full thread

This has been a long time coming, thanks a lot for my collaborators for all their help! Oliver Slumbers, Stephen McAleer, Yali Du, Stefano V Albrecht, and David Mguni It's actually going to be my first ever oral presentation, so excited (and nervous) about that 👀

At the Adaptive and Learning Agents Workshop I'll be presenting our comprehensive study on the efficacy of different visual encoders for imitation learning in modern video games. I'll be presenting the work as a short talk and poster at the ALA workshop on Monday!

@msftresearch.bsky.social

Lukas Schäfer reposted

Microsoft Research

7 Apr 2025

We are making available an experimental and interactive real-time gameplay experience in Copilot Labs, powered by our Muse family of world models. Learn more about the research underpinning this experience: www.microsoft.com/en-us/resear...

microsoft.comWHAMM! Real-time world modelling of interactive environments. - Microsoft ResearchToday we are making available an interactive real-time gameplay experience in Copilot Labs, powered by our Muse family of world models. This blog post provides a deeper look at the research underpinni...

Lukas Schäfer reposted

Bálint Gyevnár

@gbalint.bsky.social

14 Mar 2025

Call for papers! 🎉 If you have felt confused about evaluation in explainable AI, this workshop is for you! We are excited to go to Bologna @ ECAI'25 as we try to build better methods for evaluating XAI. Can't wait to see what ideas the community will bring! sites.google.com/view/excd-20...

Lukas Schäfer reposted

Jakob Foerster

@jfoerst.bsky.social

12 Mar 2025

My group @FLAIR_Ox is recruiting a postdoc and looking for someone who can get started by the end of April. Deadline to apply is in one week (!), 19th of March at noon, so please help spread the word: my.corehr.com/pls/uoxrecru...

my.corehr.comJob Details

Lukas Schäfer reposted

Katja Hofmann

@katjahofmann.bsky.social

7 Mar 2025

What happens when we train generative AI models to predict everything that happens in a modern video game? In my TED AI talk I share insights from how AI can learn the physics of a game, and how this research could empower game creatives: www.ted.com/talks/katja_...

ted.comWhy we're training AI on video gamesAI is already a powerful tool for collaboration — but this is just the tip of the iceberg, says AI researcher Katja Hofmann. She describes her team's work training AI on years of human gameplay data f...

Lukas Schäfer reposted

Katja Hofmann

@katjahofmann.bsky.social

19 Feb 2025

Today in Nature: our research on world and human action models (WHAM) - generative ai models of video games, aimed towards supporting game creatives in gameplay ideation : nature.com/articles/s41... - huge congrats to everyone who made this happen, I couldn't be more proud 🥳

nature.comWorld and Human Action Models towards gameplay ideation - NatureA state-of-the-art generative artificial intelligence model of a video game is introduced to allow the support of human creative ideation, with the analysis of user study data highlighting three neces...

@katjahofmann.bsky.social

19 Feb 2025

Awesome to see this released and out in the world now. Huge congrats to the entire team! It was super cool to see the vision come to life as an intern and everything coming together now after re-joining the team 👏

Microsoft Research@msftresearch.bsky.social

Nature published Microsoft research detailing our WHAM, an AI model that generates video game visuals & controller actions. We're releasing the model weights, sample data & WHAM Demonstrator on Azure AI Foundry to enable researchers to build on this work. www.microsoft.com/en-us/resear...

Lukas Schäfer reposted

Katja Hofmann

13 Jan 2025

An exciting new internship opportunity in my team: Generative AI for Human Creativity - in the context of a multi-disciplinary internship with a fantastic set of mentors 🤩

jobs.careers.microsoft.comSearch Jobs | Microsoft Careers

aamas2025.orgCall for Contributions to the Doctoral Consortium – AAMAS 2025 Detroit

10 Jan 2025

I had an awesome experience at the AAMAS doctoral consortium (despite virtual format) and would highly recommend it to any PhD student. @sharky6000.bsky.social has been an amazing mentor who gave me lots of motivation and advice for my journey forward! Deadline in 7 days so apply soon 👇

Marc Lanctot@sharky6000.bsky.social

As a PhD student, would you like the opportunity to interact closely with established researchers and other students, receive feedback on your work, and get advice on managing your career? Check out the Doctoral Consortium at #AAMAS 2025! Deadline: Jan 17th. aamas2025.org/index.php/co...

Lukas Schäfer reposted

Karl Tuyls

@karltuyls.bsky.social

18 Dec 2024

The first textbook on multi-agent reinforcement learning is out - a landmark for the field, the first textbook covering game-theoretic foundations with state-of-the-art deep learning! Congrats to its authors Stefano Albrecht , LukasSchaefer and Filippos Christianos More details: www.marl-book.com

marl-book.comMulti-Agent Reinforcement Learning: Foundations and Modern ApproachesTextbook published by MIT Press (2024)