·
Researchers & Thinkers
Edward Beeching
Research Scientist |Hugging Face
RL and LLM alignment researcher. Co-created the Open LLM Leaderboard, co-authored Zephyr and the Alignment Handbook, and built Godot RL Agents.
7papers
GitHubedbeechingResearch Papers
7 papersarXiv
Optimizing Test-Time Compute via Meta Reinforcement Fine-Tuning
2503.07572·2025-03-10
arXiv
Jack of All Trades, Master of Some, a Multi-Purpose Transformer Agent
2402.09844·2024-02-15
arXiv
Zephyr: Direct Distillation of LM Alignment
2310.16944·2023-10-25
arXiv
Godot Reinforcement Learning Agents
2112.03636·2021-12-07
arXiv
Learning to Plan with Uncertain Topological Maps
2007.05270·2020-07-10
DOI
EgoMap: Projective Mapping and Structured Egocentric Memory for Deep RL
10.1007/978-3-030-67661-2_31·2020-09-14
arXiv
Deep Reinforcement Learning on a Budget: 3D Control and Reasoning Without a Supercomputer
1904.01806·2019-04-03
Biography
Deep Research
Edward Beeching is a Research Scientist at Hugging Face in Lyon, France, specializing in reinforcement learning for LLM alignment and embodied learning. He holds a PhD from INSA Lyon / INRIA CHROMA on deep RL for robotic navigation and planning. At Hugging Face he co-created the Open LLM Leaderboard, co-authored the Zephyr alignment model and the Alignment Handbook, integrated the Decision Transformer into the Transformers library, and co-authored TRL (Transformer Reinforcement Learning). He built Godot RL Agents (1,400 stars), an open-source framework connecting the Godot game engine to deep RL algorithms. His recent work includes SmolLM3, a 3B dual-mode reasoning model, the JAT multi-purpose transformer agent, and a survey of 16 open-source async RL training libraries. He has published 381 models and 103 datasets on Hugging Face.
Reinforcement Learning from Human Feedback (RLHF)LLM Alignment & Preference TuningOpen LLM Evaluation & LeaderboardsDecision Transformers & Offline RLAsync RL Training InfrastructureSmall Language Models (SmolLM)Deep RL for Robotics & NavigationGame Engine RL (Godot RL Agents)Multi-Purpose Transformer AgentsTest-Time Compute Optimization
Timeline
12 Research12 total
2026
2025
2024
2023
2022
2022-03Research
Joined Hugging Face as a Research Scientist in Lyon2021
Key Contributions
Godot RL Agents
Open-source framework connecting the Godot game engine to deep RL algorithms (StableBaselines3, Sample Factory, Ray RLLib, CleanRL). 1,400 stars on GitHub.
Open LLM Leaderboard
Co-created the Hugging Face Open LLM Leaderboard for benchmarking open-source large language models on IFEval, BBH, MATH, GPQA, MUSR, and MMLU-PRO
Zephyr: Direct Distillation of LM Alignment
Co-authored Zephyr-7B, a distilled chat model that surpassed Llama2-Chat-70B on MT-Bench using distilled direct preference optimization without human annotation
The Alignment Handbook
Co-authored robust recipes for aligning language models with human and AI preferences, including fine-tuning recipes for SmolLM2-Instruct and Zephyr
TRL (Transformer Reinforcement Learning)
Co-author of Hugging Face's open-source library for training transformer language models with RL techniques including SFT, GRPO, and DPO
Decision Transformer in Transformers
Integrated the Decision Transformer model into Hugging Face's transformers library, enabling offline RL via conditional sequence modeling
SmolLM3
Co-authored SmolLM3, a 3B parameter dual-mode reasoning model with multilingual support, 128k context, and state-of-the-art performance at 3B scale
JAT: Jack of All Trades Multi-Purpose Transformer Agent
Co-authored the first fully open-sourced multi-task RL agent capable of sequential decision-making, CV, and NLP with a single set of weights
Mixture-of-Thoughts Dataset
Published the Mixture-of-Thoughts dataset on Hugging Face with 349k items and 197 likes, used for reasoning model training
Notable Quotes
“
I had the opportunity to spend the last month building an open-source, state of the art, dual-mode reasoning model at the 3B scale. Building on the amazing work of Hugging Face's Pretraining team. It was tough but we managed to get on the Pareto front with the Qwen3 models.
“
A month ago I joined Hugging Face as a Research Scientist. They're great: opening an office in Lyon, allowing me to work on open-source projects and trusting me to define my own schedule. I am proud to have added the Decision Transformer to transformers.
“
For those of you who don't have time to read 5,000 words about async RL plumbing (we get it, you have models to train).
15 sources(click to expand)
Edward Beeching - Personal Websiteedbeeching on GitHub (90 repos, Godot RL Agents 1,400 stars)edbeeching on Hugging Face (381 models, 103 datasets)Ed Beeching - Research Scientist - Hugging Face | LinkedInEdward Beeching on Google ScholarDBLP: Edward Beeching publicationsZephyr: Direct Distillation of LM Alignment (arXiv)Open LLM Leaderboard on Hugging FaceThe Alignment Handbook on GitHubTRL: Transformer Reinforcement Learning on GitHubSmolLM3 blog post on Hugging FaceKeep the Tokens Flowing: Lessons from 16 Open-Source RL LibrariesGodot Reinforcement Learning Agents (arXiv)Jack of All Trades, Master of Some (arXiv)Optimizing Test-Time Compute via Meta Reinforcement Fine-Tuning (arXiv / ICML 2025)
Research generated March 19, 2026
Researchers & Thinkers/Edward Beeching
All Profiles