Uploaded April 2026 | Updated November 2026, 5 days ago
Reinforcement learning is becoming central to agentic systems, but moving from RL for LLMs to RL for agents introduces a new set of challenges: environments, rollouts, tool use, inference bottlenecks, reward design, and evaluating multi-step behavior in the real world.
In this live Hugging Face workshop, we bring together researchers and builders working on the frontier of RL for agents. The session will feature short talks followed by a discussion on what is working today, where open methods still fall short, and what comes next.
Speakers include:
- Lewis Tunstall, Hugging Face
- Will Brown, Prime Intellect
- Ofir Press, Princeton University
- Alex Zhang, MIT CSAIL
additional guests TBA
Topics include:
- training agents with open source tools
- scaling RL for language agents
- multi-step verification and reward design
- benchmarking agent capability beyond static tasks
- recursive reasoning and new agent architectures
Reinforcement learning is becoming central to agentic systems, but moving from RL for LLMs to RL for agents introduces a new set of challenges: environments, rollouts, tool use, inference bottlenecks, reward design, and evaluating multi-step behavior in the real world.
In this live Hugging Face workshop, we bring together researchers and builders working on the frontier of RL for agents. The session will feature short talks followed by a discussion on what is working today, where open methods still fall short, and what comes next.
Speakers include:
- Lewis Tunstall, Hugging Face
- Will Brown, Prime Intellect
- Ofir Press, Princeton University
- Alex Zhang, MIT CSAIL
additional guests TBA
Topics include:
- training agents with open source tools
- scaling RL for language agents
- multi-step verification and reward design
- benchmarking agent capability beyond static tasks
- recursive reasoning and new agent architectures






![[Monday evening short video] Summary of two new amazing LLM benchmarking papers: GAIA and GPQA
Sharing a summary of two amazing LLM benchmarking papers published just last week:
- GAIA: the General AI Assistant benchmark (disclaimer: Im a co-author with amazing co-authors)
- GPQA: the Graduate Level Google Proof QA benchmark (disclaimer: the authors are also awesome)
When two teams (covering a diverse range of actors like Anthropic, Cohere, New-York University, Hugging Face, Meta AI) independently come up with benchmarks that share so many aspects (while being really different in goals and approaches), you know the future of LLM benchmarking is basically changing under your eyes.
Both are super difficult with ~30% GPT4 success rate, both are small (450 questions) and carefully hand-crafted question by question, with a single gold answer and a strong interest for the reasoning it-self rather than memorization capabilities. Both very challenging test bed for coming models capabilities
And above all: Im super excited that these are open-source benchmark giving us common ground for comparison of the coming frontier models. On to the future (of open evaluation)!
Papers and more information:
- GAIA: https://huggingface.co/papers/2311.12983
- GPQA: https://huggingface.co/papers/2311.12022 [Monday evening short video] Summary of two new amazing LLM benchmarking papers: GAIA and GPQA](https://i.ytimg.com/vi/e4jOGywUCc4/mqdefault.jpg)


![[Whisper Fine Tuning Event] Talks: Whisper & VoxPopuli
[Whisper Fine Tuning Event] Talks: Whisper & VoxPopuli [Whisper Fine Tuning Event] Talks: Whisper & VoxPopuli](https://i.ytimg.com/vi/fZMiD8sDzzg/mqdefault.jpg)
