The NetHack Learning Environment

Pleased to announce the NetHack Learning Environment has a new home, and a new set of maintainers! Find it at github.com/heiner/nle. A huge thanks to @stephen_oman and @MartinKlissarov for helping with the project! Happy hacking!

Excuse me?

@Ishmokin This is an excellent idea.

@NetHack_LE Is there a way to spectate games? Would be awesome if we could watch like how we do at nethack servers like alt.org and hardfought.org

1 43.6 Grok-4-Wiz-AI-Cha died in The Dungeons of Doom on level 1. Killed by a housecat.

💯 Who knew that the International Math Olympiad (IMO) is much easier than @NetHack_LE for AI.

Happy "@NetHack_LE is still completely unsolved" day for those of you who are celebrating it. We released The NetHack Learning Environment (arxiv.org/abs/2006.13760) on this day five years ago. Current frontier models achieve only ~1.7% progression (see balrogai.com). For a recent blog post on what makes it so hard for AI, check out @HenaffMikael's analysis: mikaelhenaff.substack.com/p/first-nethac…

@HenaffMikael @dgant @_rockt arguably

@dgant @_rockt @NetHack_LE Probably, since you're presumably been able to make progress in the real world which is more complex than NetHack.

Great post by @HenaffMikael (after ascending, what an achievement!) on what makes @NetHack_LE so extremely difficult for AI (even LLMs: balrogai.com). "While NetHack is complex in comparison to other RL benchmarks, it still contains only a tiny fraction of the complexity of the real world (its source code is 4.2MB, which provides an upper bound on its Kolmogorov complexity). As long as we can’t reliably solve this game for which we can easily collect lifetimes worth of data, have access to detailed textual resources (and even the underlying source code), and large-scale datasets of human gameplay, I think AGI remains a ways off."

A couple bits of news: 1. Happy to share my first (human) NetHack ascension-next step is RL agents :) 2. I wrote a post discussing some @NetHack_LE challenges & how they map to open problems in RL & agentic AI. Still the best RL benchmark imo. mikaelhenaff.substack.com/p/first-nethac…

New home! github.com/NetHack-LE/nle

NetHack-LE accepts your allegiance.

Happy to announce the latest release of @NetHack_LE (version 1.2.0). You can now use the seed function to make the dungeon layout reproducible across training episodes. The in-level interaction and combat is still randomly determined and doesn't impact lower level layouts.

⚔️ MiniHack Updates! ⚔️ 1️⃣ MiniHack 1.0.0 is here! Following popular demand, it now supports the new Gymnasium API and is built on NLE 1.1.0. Huge thanks to @Stephen_Oman (maintainer of @NetHack_LE ) for his outstanding contribution! 🙌

Can AI agents adapt zero-shot, to complex multi-step language instructions in open-ended environments? We present MaestroMotif, a method for AI-assisted skill design that produces highly capable and steerable hierarchical agents. To the best of our knowledge, it is the first method that, without expert labeled datasets, solves compositional tasks requiring hundreds of steps for completion. All the modules within MaestroMotif are learned from interaction: from the highest level of planning to the lowest-level of sensorimotor control. On the open-ended domain of NetHack, it surpasses existing approaches, including those that are fine-tuned specifically for each task. At the heart of MaestroMotif is the idea that decomposing a task into subtasks significantly helps decision making. MaestroMotif leverages an agent designer's intuition about a domain to identify important skills and describe them in natural language. These short descriptions then get converted into adaptable hierarchical agents through AI feedback and in-context learning. Our paper was recently published at ICLR 2025 and we open-source the whole project including the code, prompts and pre-trained models. Paper: arxiv.org/abs/2412.08542 Code: github.com/mklissa/maestr… NotebookLM Podcast: bit.ly/4jLi6mo This work was done with the amazing @HenaffMikael, @robertarail, @shagunsodhani, Pascal Vincent, @yayitsamyzhang, @pierrelux, Doina Precup, with equal supervision by @MarlosCMachado and @proceduralia. Take a look at the following thread:

@demishassabis @karpathy cc @_rockt had that idea early!

@karpathy Cool idea!

I quite like the idea using games to evaluate LLMs against each other, instead of fixed evals. Playing against another intelligent entity self-balances and adapts difficulty, so each eval (/environment) is leveraged a lot more. There's some early attempts around. Exciting area.

💯 For me this is NetHack (see @NetHack_LE and balrogai.com). I am still holding my breath.

In the meantime, @NetHack_LE and balrogai.com...

cc @_rockt

Yearly reminder

@SteveStricklan6 ARC will be solved before Nethack

@jsuarez @_rockt @PaglieriDavide Seems slightly dubious in that you can run 1 NetHack / core at 1000k+ SPS total, per machine.

@_rockt @PaglieriDavide As fast as nethack is, the main difficulty for RL is that it is too slow. I think we would have seen much more progress if it trained at 500k sps like our new envs

💯‼️ That's why @PaglieriDavide created balrogai.com