@ What it is?
NetHackers is an open effort to build the first program that can reliably win NetHack 3.6.6. Write the bot by hand, evolve it with coding agents, improve the thing that improves it, or recurse until something interesting happens. Whatever path you take, the evaluation is the same: the symbolic bot that comes out of the loop. The ultimate goal is an ascension; until then, the board measures how far each bot gets.
NetHack (1987) is among the oldest unsolved challenges in games. Winning — an ascension — means descending some fifty procedurally generated levels, seizing the Amulet of Yendor, and escaping through five final planes: tens of thousands of turns under permadeath, randomized item identity, and a knowledge burden its own developers say takes years to master.
The entire history of autonomous wins is three games, in 2015, by one hand-coded bot, on a version whose winning exploit was patched out that same year. On modern NetHack, no program has ever won.
| Line of attack | Best known result | Asc. |
|---|---|---|
| Humans, all logged games | — | 0.4% |
| Humans, experts | 15.9% win rate, 47–71% in form; record streak 61 | reliable |
| Symbolic — BotHack (2015) | won on v3.4.3 via a pudding-farming exploit, removed in 3.6 | 3* |
| Symbolic — AutoAscend (2021) | 7.8% progression — the floor this board measures; Medusa in 0 of 109,545 games | 0 |
| Learned — imitation (2021–) | scaling flattens below the very bot it imitates | 0 |
| Learned — RL (2026) | 11.78% progression score-trained, 16.98% depth-trained — first to beat AutoAscend since 2021 | 0 |
| LLM direct play (2026) | 13.2% progression in BALROG | 0 |
Every machine line: zero ascensions on NetHack 3.6.6, the NLE research standard. Progression is one 0–1 metric throughout, but each line is measured on its own seeds and episode counts, so read it as indicative, not head-to-head. *BotHack's three wins were on the older 3.4.3, via an exploit the developers then removed.
One number did just move. In September 2026 an RL agent beat AutoAscend for the first time since 2021 — roughly doubling its challenge median, with depth-trained policies reaching the Castle, deeper than AutoAscend has ever gone. On progression that is 16.98% against AutoAscend's 7.8%.
In the last couple of years, coding agents that write and refine programs in a loop have started cracking problems that resisted everything else. They push abstract-reasoning puzzles that stalled LLMs for years, beat the human winners of SAT-solver competitions, and turn up new, provably-correct algorithms. The pattern is consistent: a model that cannot reliably do the task itself can write a program that does.
NetHack is where that pattern has not yet held. The researchers behind the NetHack Learning Environment put it forward as a grand challenge for AI, and it is still unsolved years later; NLE co-author Tim Rocktäschel marks the anniversary each year with "AI still can't learn to play NetHack." It is the hardest game in the BALROG suite, and the one frontier models are worst at by a wide margin. Whether the approach cracking everything else can crack this one, nobody knows.
That is what NetHackers is for: an open attempt to find out together. What we score is the program a coding agent writes, and every result compounds on the last instead of restarting with each paper.
The unit of evaluation is the program — a deterministic bot, cheap to run and exactly replayable. Objectives grid over the 73 starting identities and the milestone ladder, so specialists and generalists all have somewhere to land. A thin hub keeps each objective's best elites; anyone can pull one, improve it, and register the result — so one contributor's improvement becomes everyone's parent. The hub never runs your search and assigns no work; how you make bots is entirely up to you.
! Why you should care
There are at least three ways to fall into this dungeon.
If Recursive, Ricursive, Discovery Loop, AIDE², and the Darwin Gödel Machine all appeared in your timeline before breakfast; if every benchmark bump is “it’s happening” and every plateau means “add another outer loop”; if your honest answer to “what improves the improver?” is “another improver” — welcome. Build the seed, mutate the harness, fork the fork. The intelligence explosion can start with not dying to a grid bug.
Winning NetHack is the headline; the research problem is generalization. A bot must turn wiki knowledge into action, decompose a tens-of-thousands-of-steps objective, discover and compose reusable skills, and recover when unfamiliar seeds or stochastic events break its plan. Held-out evaluation tests whether those skills transfer rather than whether one trajectory was memorized — and the result is a symbolic program you can inspect.
These are not game-only problems. ASPIRE applies a similar loop to robotics, repairing code-as-policy programs after failed rollouts and saving skills for new tasks; Code as Policies composes perception, control, and tools into executable robot behavior. NetHack is a cheap, fast arena for studying long-horizon planning, skill composition, tool creation, and generalization to unfamiliar seeds — without a robot lab. » how held-out evaluation works
You don't need to be good at NetHack, or an ML researcher. A coding agent and a laptop will do. Point it at a bot, watch it evolve and climb the board, and go for something no machine has managed in nearly four decades: get a program to win. It runs locally, it is genuinely addictive (a slot machine of stupid deaths and small breakthroughs), and every win you register becomes someone else's starting point — your name on the frontier. » start solving
< The Frontier
Mean progression (0–100%) across all 73 identities, and how far each sits above the AutoAscend floor. Click any identity for its directly comparable leaderboard, sources, episodes, dates, and current holder.
& Hackers moving the frontier
Recognition is attached to concrete records, roles, and source commits — not collapsed into one global score.
Current identity leaders above AutoAscend. Click a row for the hacker's complete contribution history.
* The Oracle
NetHack's Oracle sells prophecy for gold; ours is free — and it's you. Two questions on whether, and how, a program finally wins. Cast your prophecy to see where the crowd — and your own tribe — lands.
% FAQ
Same dungeon, three different jobs.
| Competition | Your job | Evaluation |
|---|---|---|
| NetHackers | Build an autonomous symbolic bot that plays NLE's NetHack 3.6.6. | An ongoing, open leaderboard for progression and ascension across fixed identities, with every bot available as a starting point for the next improvement. |
| NeurIPS NetHack Challenge 2021 | Build an agent to play the full game through NLE 3.6.6. | A fixed 2021 event: ascensions first, then median in-game score over random characters. |
| Mazes of Menace | Port NetHack 5.0 from C and Lua to readable ES6 JavaScript. | Bit-exact screen and PRNG parity on public and held-out sessions, followed by a generalization phase. |
They ask opposite questions about the same gap.
ARC-AGI limits what the solver is told: infer a transformation from a few examples, then apply it to a new input. It measures how efficiently a system adapts when task evidence is deliberately scarce.
NetHackers makes the inverse choice. NetHack's rules are public — you can read the implementation, consult decades of wiki knowledge, study every bot that came before, and run the game as often as you like. Knowledge from earlier attempts is something to keep and exploit, not something to withhold. The question is how much competence a process can build out of all that.
Having the rules still leaves the decisions. A running program sees only what it has explored. A simulator does not hand you a tractable strategy — you still have to choose what to consider and what to ignore. And choices interact across tens of thousands of turns, where progress now can consume what you needed later.
The two are complementary. A system can recognize a new pattern from three examples and still fail to build and maintain a large, reliable controller; a strong NetHack specialist would say little about adapting to unrelated tasks.
Because it is a strong starting point, not a finished answer. AutoAscend won the 2021 challenge and already contains serious symbolic machinery for exploration, combat, inventory, altars, and Sokoban. It also contains plenty of strange, sometimes plainly dumb behavior: every role begins by farming dungeon level 1 until experience level 8; Monks are hard-coded never to choose a melee weapon or body-armor suit; autopickup is replaced by a brittle hand-written item-priority system; bag use is disabled; shopping is unimplemented; and the post-Mines plan still turns into TODOs. That is exactly what we want from a seed: enough competence to reach interesting states, and enough legible mistakes for humans and coding agents to start fixing immediately.
Because volunteer computing solves the opposite problem. BOINC ships code the project wrote and signed to machines it does not trust, then validates results by having two of them agree. Both halves invert here. The untrusted thing is the code — a stranger's bot, or a coding agent running with its permission prompts switched off — and a volunteer's side of that bargain is that they never have to read it, because the project signed it and stands behind it. We cannot make that promise about code we neither wrote nor reviewed, and routing it in as an input file to a signed wrapper only means the signature stops covering the part that actually runs. Agreement would not buy us much either: a bot overfitted to the public seeds is not a disagreement between hosts, it is a number every replica reproduces perfectly. The only check that catches it is re-running on dungeons the author has never seen — and a volunteer machine cannot hold a secret. Cycles were never the scarce thing anyway; a full private pass is roughly a thousand episodes, hours on one box. What is scarce is good programs, and the agent tokens to find them. So the hub hands out no work at all: you search on your own machine, and we keep the link.