Where NetHack agents stand
Progress in other games has come much faster than in NetHack. In the last year language models have reached human level at Pokémon battles, Baba Is You, and several games in the BALROG benchmark suite for LLM game agents, and much of that progress came from the harness around the model, meaning its memory, tools and skills.
On BALROG's NetHack leaderboard most frontier models score under 4% on its progress metric, and the best entry, GPT-6 Astra, reaches 13% at about $100 a game. Astra also produced the first recorded ascension by a language model this September, on a public server and on its third attempt, in a harness it wrote itself, with a person giving suggestions and stopping and resuming sessions. That shows a model can win the game. But it was one win in three games, with no fixed budget, spread over twelve days, so it doesn't tell us how often a model wins under fixed conditions.
BALROG, the standard evaluation, runs every model in one fixed harness that shows the last 16 observations and offers no skills. Meanwhile, harnesses like Claude Code, Codex and Prime Agent give a model persistent memory, reusable skills and subagents. Continual Harness lets an agent rewrite its own memory and tools between games, and GEPA improves prompts by reflecting on the agent's own traces. If any of this is real progress on long-horizon play, NetHack is where we would expect it to show, but BALROG's leaderboard can't tell us whether it does.
In this post we take on three problems with how NetHack agents are evaluated, and then test whether search helps:
- The interface. BALROG's harness gives a model raw keys and a short history, so its leaderboard measures that harness as much as the model. We built a sandboxed harness around NetCore, a small set of skills that hand control back whenever a decision is needed, and ran five current models in it. They score 1.3 to 1.7% on a balanced progress metric, within noise of each other, and they still dive deeper than their experience supports and die. Because we haven't run the same model in both harnesses, we can't yet say how much BALROG's harness holds models back.
- The engine. NLE holds one game per process, cannot save or restore a game, and exposes nothing to adjust, so we rebuilt it; that work is in part 1.
- The metric. BALROG's score rewards depth on its own, so diving looks like progress, and we therefore report a balanced version, BALmin, next to it.
- Prime-Agent-Explore. The score itself matters less to us than whether an agent can turn more compute into more progress, since good play is exactly the data NetHack has always lacked. Prime-Agent-Explore resumes play from checkpoints that an orchestrator chooses, and with 30 attempts it takes GLM-5.2 from 1.32% BALmin on a single life to 10.99%, while the best of three independent lives reaches only 2.45%. With 15 attempts, taking away the orchestrator's choice of checkpoint and its plan drops the score from 9.53% to 2.45%, so most of the gain comes from where it resumes and what it tells the next attempt. The comparison we still owe is against the best of 30 independent lives.
Setup
NetHack (1987) is one of the original roguelikes: a player descends a procedurally generated dungeon with a single life to find the Amulet of Yendor and climb back out, and the whole world is drawn as an 80×21 grid of characters.
The game is hard for people too: of the 46,796 complete human games we decoded from the NetHack Learning Dataset, 1.8% end in an ascension, and winning games run for tens of thousands of moves. Most of a run is spent deciding when to go deeper, because each staircase down brings tougher monsters, and a player who descends faster than they gain experience dies.
Formally, NetHack is a partially observable Markov decision process, in which the agent sees part of the hidden game state (the screen, its inventory, the message line), picks an action, and the game then moves on at random. In our evaluations an agent plays one game and is scored by BALmin before it dies or its budget of model calls runs out; each action is a tool call that can expand into many keystrokes.
Because so little progress has been made on the full game, our default setup takes away what is hardest about perceiving NetHack so that an evaluation can focus on the long-horizon decisions we care most about. Fog of war is off, so the agent sees the whole floor, and the full map, the inventory and the visible monsters are always one call away through an observe skill that spends no game time. Fog makes NetHack a map-reading problem as much as a planning one, and agents so far have mostly been tested on navigation and survival. With the map reading removed, what is left is closer to the decisions we want to measure, such as when to take a staircase down and when to fight. We report fog-of-war results separately where we have them.
Everything in this post runs on NetHack+, a rebuilt NetHack engine that holds thousands of games in one process and forks a game in about 20 microseconds by copying its memory, against 0.74 seconds to replay the same game from its seed after 40,000 steps. That fork is what makes Prime-Agent-Explore cheap, because it can resume play from any saved checkpoint instead of replaying the game. How we rebuilt the engine is in part 1.
Each model plays in Prime Agent, a general agent harness, and reaches the game only through NetCore, our set of eight skills: explore the floor, walk to a square, fight, kick, apply an item, rest, pray, and press a raw key, plus an observe call that returns the full map and inventory without spending game time. One skill call can run for up to 100 game turns, but it hands control back whenever a decision is needed (a level change, a nearby hostile, a sharp drop in HP or food about to spoil), and it never answers a game prompt for the model. A game ends when the hero dies or its budget of [TODO: N] skill calls runs out. In the current system every call is recorded as the keystrokes the engine consumed, and the score comes from replaying that record from the seed in a fresh sandbox, not from anything the agent or the live game reports.
One caveat is that the runs in this post date from July to September 2026 and used an earlier version of this setup, on the hand-migrated engine. In that version the harness reached a game server on the host, and scores were read from the deepest level and experience level the game reported rather than from a keystroke replay.
How models play
We ran GPT-5.6 Sol, GPT-5.6 Luna, Gemini-3.7 Flash, Qwen-3.8 Max and GLM-5.2 in Prime Agent with NetCore skills, on five fixed seeds with three runs each, with full vision. Every number below comes from those 15 runs per condition unless noted. That is not many runs, and NetHack is noisy (in an earlier harness, two configurations that were byte-for-byte identical scored 2.31 and 4.48 over five games each), so differences of a few tenths of a percent should be read as noise.
Measuring progress without rewarding depth alone
NetHack's real reward, ascending, almost never happens, so every NetHack benchmark needs a progress metric. BALROG's converts the deepest dungeon level and the highest experience level an agent reached into two human-derived probabilities of eventually winning, then reports the higher one, which we call BALmax. The problem with this metric is that an agent can raise it just by taking stairs, and leaderboard runs do exactly that, diving and dying far deeper than their experience supports.
We report BALmin alongside it, which takes the lower of the two probabilities. Under BALmin, once depth outruns experience, more depth earns nothing and the only way to raise the score is to get deeper and stronger together, which is how successful human games progress.
How far agents get compared with people
The comparison we care about most is against human players, so Figure 1 plots, for each agent, the median BALmin of the runs still alive against game turns, next to the median of 433 winning human games. While they are alive, single lives keep roughly the pace of winning players, and they end at 2.0 to 2.9%, about where a winning player is around turn 2,000 (2.4%). That is as far as they go, because fewer than two runs of any model are still alive by turn 2,200, whereas a winning game runs for tens of thousands of turns. Prime-Agent-Explore is the only agent that keeps going, to about 16,600 turns. It keeps pace with winning players for the first 2,000 turns or so and then falls behind, reaching 6.0% at 10,000 turns against their 11.7%.
| Agent | n | BALmin mean | median | BALmax mean | Dlvl | XL | median run beats % of human games |
|---|---|---|---|---|---|---|---|
| Humans (NLD) | 46,796 | 5.41 | 2.12 | 8.05 | 6.48 | 5.04 | – |
| GPT-6 Astra, BALROG harness, fog of war (current best) | 5 | 4.03 | 3.69 | 13.24 | 9.80 | 5.20 | – |
| Gemini-3 Pro, BALROG harness, fog of war (best when we ran these) | 5 | 1.84 | 1.54 | 6.77 | 5.60 | 3.40 | 30.9 |
| GLM-5.2 | 15 | 1.32 | 1.85 | 9.96 | 8.00 | 2.33 | 41.6 |
| GPT-5.6 Sol | 15 | 1.65 | 2.08 | 5.28 | 6.47 | 2.67 | 45.3 |
| GPT-5.6 Luna | 15 | 1.50 | 1.85 | 3.79 | 5.53 | 2.27 | 41.6 |
| Gemini-3.7 Flash | 15 | 1.47 | 1.85 | 9.80 | 8.33 | 2.20 | 41.6 |
| Qwen-3.8 Max | 15 | 1.74 | 2.12 | 4.56 | 4.73 | 3.40 | 52.4 |
The five models are within noise of each other, since the standard errors on these means are 0.2 to 0.3, about the size of the gaps between them. A modern harness doesn't move single-life play much: the best single-life mean here is 1.74, while GPT-6 Astra in BALROG's own harness averages 4.03 with fog on. That run used a different model and only five games, so the two aren't directly comparable.
Why they die early: going deeper than their experience supports
What every model shares is the dive, in which they reach dungeon levels 5 to 8 at experience level 2 or 3, whereas human games at the same depth have about twice the experience (Figure 2). Their traces suggest that they know the game, since they name monsters correctly and understand the goal. The deaths we read come from a few irreversible decisions, like fighting a losing battle or taking the stairs on low health, although we haven't yet tallied death causes across all 75 runs.
Every one of the 75 single lives ended in death rather than by running out of budget, but the means hide how differently the models get there. The sketches below combine each run's depth and experience path with an excerpt from the model's own trace. (GPT-5.6's reasoning is only available as the provider's summaries, and Gemini-3.7 Flash records none, so for Gemini we quote the comments it wrote in its tool calls.)
Gemini-3.7 Flash dives fastest
Gemini-3.7 Flash reached the deepest floors on average (8.3) at the lowest experience level (2.2), its depth outran its experience in 14 of 15 runs, and its median run died on game turn 425. The following is a typical ending, on dungeon level 9 at experience level 3 (tool output abridged):
# Move to (25, 15) stairs down
obs = await nethack.np_move_to(x=25, y=15)
[INTERRUPTED: HP 12/37 -- below half during this skill. Something is hitting you.
ADJACENT: W=a(ant/insect)[hostile]]
# Attack soldier ant at (15, 12)
obs = await nethack.np_melee_attack(x=15, y=12)
Qwen-3.8 Max plays patiently and knows the rules
Qwen-3.8 Max went the least deep (4.7) but gained the most experience (3.4), dove in only 8 of 15 runs, and lived longest, with a median of 1,005 turns and one run past 4,000. That is why it tops BALmin and sits near the bottom on BALmax. Its reasoning also shows real game knowledge: surrounded on dungeon level 4 at HP 6/25, it works through prayer correctly (low HP only counts as an emergency your god will fix below a seventh of your maximum, or below 6) and then retreats up the stairs:
"Prayer: HP 6/25 is roughly 1/4, not below 1/7 (≈3.6). Prayer has low probability of healing, and there's also a cooldown. The situation is dire, but not fatal enough for a guaranteed divine response."
GLM-5.2 varies widely between runs
Four of GLM-5.2's 15 runs reached dungeon level 12 or 13, while six never got past level 5 or experience level 1. This spread is why it jumps from last to second when we take each seed's best life. The turn shown below comes from one of its quiet dives.
GPT-5.6 Sol and Luna take the stairs
Both GPT-5.6 models dove in 13 of 15 runs, with Sol going a little deeper and living a little longer (median 548 turns against Luna's 436), and their reasoning summaries state the strategy openly. Here is Sol on dungeon level 11 at experience level 3, at turn 489:
"I'm thinking this auto descends quite well! It's been effective so far."
Luna's summary reads as follows, on dungeon level 7 at experience level 2 with HP 6/23, turn 239:
"I need to descend since my HP is low, so I'll take the stairs."
Taking the stairs to escape a fight is a real NetHack tactic, but doing it at experience level 2 on dungeon level 7 makes the next floor worse, and this is the pattern that BALmax scores as progress.
HP: 14/16 AC: 6 Dlvl: 3 Turn: 368 XP: 1 $: 14 Pos: (23,17)
model call: np_explore_level()
reply:
[Executing skill 'explore_level'. Skill has been running for 100 timesteps without interruption. Rethinking. GAME: You hear a door open.]
=== STATUS ===
HP: 16/16 AC: 6 Dlvl: 4 Turn: 515 XP: 1 $: 14 Pos: (3,6)
F) beside it and a bat (B) in another room, the agent asks NetCore to explore the floor. The skill runs its full 100 steps without an interrupt, and 147 game turns later the hero is on dungeon level 4, still at experience level 1. Nothing went wrong in this call, and the problem is that each call like it is a small, reasonable step deeper.We also looked at where the calls go. For GLM-5.2, a life averages about 110 tool calls and the median life lasts about 640 game turns. Of those calls, 38% are np_move_to, 14% raw key presses, 12% each for looking at the map and exploring a floor, 7% melee and 6% resting. NetCore's most common reason to hand control back was an item or monster coming into view, then a path that couldn't be found, then a level change.
One life, or the best of three
A mean over single lives (pass@1) hides how good a model's best lives are. Taking each seed's best of its three runs (a pass@3-style score) reorders the field: GLM-5.2 moves from last to second, and Gemini-3.7 Flash, whose best runs top out lowest, falls to last. With 15 runs per model neither ranking is settled, so this should be read as a warning about single-life means rather than as a ranking. A model whose good runs are far better than its average gains the most from any method that keeps its best attempt, and search is one of those methods.
What a dollar buys
We can now put a price on every model in our harness, using the paper's token totals for GLM-5.2 and their own traces for the other four. A life costs about $0.06 with GPT-5.6 Luna, $0.78 with GLM-5.2, $1.38 with Gemini-3.7 Flash, $1.40 with GPT-5.6 Sol and $4.02 with Qwen-3.8 Max, a 60× spread for scores that sit within noise of each other. These trace-based prices are not estimates in the loose sense: summed over the grid's 60 lives plus DeepSeek V4 Flash's five, they come to $119.52, exactly what the account was billed for those 65 rollouts. Tripling the budget to keep each seed's best life helps every model a little, but only search moves the score far, at about 125 times the cost of a life. The cheapest point on the frontier is worth noting, because Luna's best of three lives costs about $0.19 per seed and scores 2.20%, which is better than any model's single life. The models we left out also say something about cost. Prime recorded zero cached tokens for Anthropic's models, and because a NetHack rollout resends its whole conversation every call, one Claude Fable 5 rollout billed $120.04 and a 15-run Claude Sonnet 5 cell would have cost about $462. Kimi K3 hit the two-hour timeout while still alive, and DeepSeek V4 Flash lost two of five runs to simply stopping, so neither finished the grid. The right panel puts our search result next to BALROG's public leaderboard, which publishes token counts for nine models under fog of war.
Fewer, stronger skills (not a matched comparison)
We also wanted to know how much the action interface matters, but we never ran a matched sweep, so the comparison crosses experiments run months apart with different harness builds and budgets, and only measures BALmax. With that caveat, GLM-5.2 in Prime Agent got almost nowhere pressing BALROG's 106 raw commands (BALmax 1.08%, deepest level 2.2 over 5 games), reached level 5 with NetPlay's 31 skills (3.75%), and reached level 8 with NetCore's eight (9.96%, the runs in the table above).
Which change fixes which failure
Because the engine exposes its settings, some hypotheses about why agents die take one changed value to test. We grouped interventions by the capability they probe: planning (choosing goals beyond the next action), control (reliably doing what you meant) and perception (reading the state correctly). All of them use GLM-5.2 on the same 5 seeds × 3 runs. With 15 runs per arm, only the revive probe is clearly outside the noise on score, while the others are directional, and they are worth reporting because the behavior each one targeted measurably changed.
- Warnings change behavior but rarely outcomes. When we interrupt a descent that a successful human wouldn't make at that experience level, or an attack below a third of maximum HP, the agent backs off 44% of the time. However, only 2% of the time does it actually gain a level before trying again.
- Locked doors may matter. Generating every ordinary door unlocked cut kicks from 4.9% to 3.9% of tool calls, and BALmin rose from 1.32 to 1.76, about one standard error. We haven't found the mechanism, since a one-point drop in kicks is too small to explain the score on its own.
- A stale map costs fights. Our harness only redraws the map when asked in order to save tokens, and as a result 9.7% of attacks hit tiles where a monster used to be. Sending the ASCII map every turn cut that to 2.4%, while a JSON map cut it to zero but helped less overall.
- Undo is the strongest single change. Reviving the agent on death with a short rollback ended runs at an average depth of 10.3 and doubled BALmin to 2.64. Revived agents remember how they died, so they can learn where the monsters are, and because that changes the game rather than the harness, we treat it as a probe. Prime-Agent-Explore keeps the idea of going back without that memory.
So far we have tested two settings on agents. Turning fog of war back on barely changes GLM-5.2's single life (1.36 vs 1.32 BALmin), so full vision isn't what holds single lives back. Removing locked doors cuts kicks per run from 18 to 3 in a mechanism check. Most of the other 15 settings haven't been tried on agents yet.
Can agents improve their own harness?
We also let an orchestrator rewrite the agent's harness between games, in a GEPA-style train, reflect and evaluate loop on top of Continual Harness. Five subagents play, an orchestrator reads their traces and edits a shared store, and the frozen store is scored on held-out seeds.
| What the orchestrator may edit | BALmin | BALmax |
|---|---|---|
| Nothing (baseline) | 1.32 | 9.96 |
| Memory entries | 1.55 | 8.01 |
| Memory, after reading two NetHack wiki strategy pages | 1.96 | 3.73 |
| Memory, wiki and NetCore's skills (1 training seed, 1 subagent, n=5) | 1.26 | 2.00 |
Memory edits look helpful, though both memory rows are within noise of each other, and most edits correct facts and thresholds ("never descend unless XL increased since last descent"). Letting the orchestrator also write skills did worse, in a smaller run with one training seed. In that run it built ambitious long-horizon tools, like one meant to take the agent through a whole dungeon level, that agents abandoned after they failed, which left the orchestrator almost nothing to learn from.
An exploration benchmark
Exploration is what agents most obviously lack, and it is also one of the hardest things to measure. This matters because once a capability has a benchmark, people curate data for it and train on it, and it improves. Exploration has had few benchmarks of its own, partly because a single-life score mixes how well a model plays with how well it uses a second chance. A forkable engine lets us measure the two separately, and because new seeds cost nothing, the test set can be swapped out whenever a model might have trained on the old one.
We score each model twice on the same seeds: once playing a single life (the standard score), and once inside a fixed exploration algorithm with a fixed budget of attempts (the explore score). The difference is what the model gains from being allowed to search:
exploration score = explore score − standard score
As defined, that gap mixes three things, namely keeping the best of many attempts, a much larger budget (30 attempts, about 125 times the cost of one life), and the search itself. A cleaner version would subtract the best of 30 independent lives instead of one, but we haven't run 30 lives per seed; for scale, GLM-5.2's best of three lives per seed averages 2.45%.
[TODO] We expect models to rank differently with search than without, but we can't show it on NetHack yet.
Prime-Agent-Explore
Inspired by Go-Explore, we save a checkpoint every 20 model calls during play, whenever the dungeon level or experience level changes, and whenever the agent calls a save skill. Between attempts, an orchestrator reads a summary of the last attempt and the list of every checkpoint (depth, experience, health, hunger, score), picks one to resume from, and writes a short plan for the next attempt. Checkpoints are files written with NetHack's own level and save formats plus the state the refactor made visible, so an attempt resumes in a fresh process. The in-memory fork is meant for algorithms that branch thousands of times, whereas this one branches only a few dozen.
Search makes it easy to cheat without meaning to, so we limit the information passed between attempts. A resumed agent gets the conversation that led up to its checkpoint, a summary of the last attempt and the orchestrator's plan. It sees nothing from attempts that branched elsewhere, meaning no other traces, death causes or item positions. We also don't restore the random number generator, so that the orchestrator can't memorize what a seed will do. (In the released version, a resumed agent starts with fresh memory instead.)
As an example of the orchestrator's choices, in one run attempt 9 reached dungeon level 10 at experience level 7 and died there under-leveled. The orchestrator then turned down the deepest checkpoint and went back to one on level 8 with full health, telling the next agent to stay on that floor and fight until it reached experience level 8 or 9.
| Model | Standard (1 life) | Explore (30 attempts) | Exploration score |
|---|---|---|---|
| GLM-5.2, full vision | 1.32 | 10.99 | +9.66 |
| GLM-5.2, fog of war | 1.36 | 6.03 | +4.67 |
All values are BALmin, in %. Standard scores are means over 5 seeds × 3 single lives. Explore scores average each seed's runs first (the full-vision runs include seed 1 three times; the unweighted mean of all 7 runs is 11.38). The other four models are next, and their runs need about $250–500 each at GLM's prices.
Even with those caveats, the effect is large: BALmin rises about 8× over one life (10.99 vs 1.32 with seeds weighted equally; 8.6× on the unweighted mean the paper reports), and BALmax rises 3.5×. The seven runs average dungeon level 18.4 at experience level 8.6, and the median run beats 92% of human games. Against the median of 433 winning human games (Figure 1), it keeps pace for the first 2,000 turns or so and then falls behind, so that at 10,000 turns it is at 6.0% BALmin while winning players are at 11.7%. With fog of war on, which is the setting BALROG uses, it reaches 6.03% BALmin and 30.93% BALmax, against 4.03% and 13.24% for GPT-6 Astra on a single life in BALROG's harness. Our runs cost about $51 per seed against Astra's $101, though our price includes a cache discount the leaderboard numbers don't.
Both halves of the orchestrator matter: with 15 attempts (5 runs each), removing the plan drops BALmin from 9.53 to 4.21, always resuming from the latest checkpoint instead of choosing one drops it to 2.81, and removing both leaves 2.45.
Jun 2020 In June 2020, the NetHack Learning Environment wrapped NetHack 3.6 as a gym environment that ran at about 14,400 steps per second, which made it one of the fastest rich environments of its time.
Sep 2021 MiniHack used NetHack's level language to build small custom tasks. Agents have since largely solved them, without that transferring to the full game.
Dec 2021 At the NeurIPS 2021 NetHack Challenge, after 631 submissions, the hand-written bot AutoAscend won with a median score of 5,300, nearly three times the best neural entry, and no agent came close to finishing the game.
Nov 2022 The NetHack Learning Dataset released 10 billion transitions from 1.5 million human games and 3 billion from AutoAscend, which made imitation learning possible at scale.
Jun 2023 Imitation learning with hierarchy and more compute improved neural agents (Piterbarg et al., Tuyls et al.), but symbolic bots still led by more than four times in median score.
Sep 2023 Motif used Llama 2's preferences over game messages as a reward and beat an RL agent trained on the score itself, an early sign that a language model's knowledge of the game could help.
Mar 2024 NetPlay gave GPT-4 a library of interruptible NetHack skills. It played sensibly but averaged a score of 285, against AutoAscend's 11,342.
Nov 2024 BALROG made NetHack one game in a standard LLM benchmark. The best model at launch reached 1.6% on its progress metric, and the leaderboard has since climbed to 13.2% (GPT-6 Astra, September 2026).
Aug 2025 Scalable Option Learning trained hierarchical RL agents on 30 billion frames of NetHack, at about 35 times the throughput of earlier hierarchical methods.
Jan 2026 Glyphbox let a model write Python against the game, and GPT-5.2's best run reached dungeon level 10.
Apr 2026 An ICLR 2026 blog post argued that NLE's interface hides menus, inventory and attributes that an agent needs.
Sep 2026 In September 2026, an RL agent trained with PufferLib reportedly passed AutoAscend's challenge score, five years after the challenge.
Sep 2026 A week later, GPT-6 Astra ascended on a public server after 37,140 turns, on its third attempt, in a harness it built for itself with a human's help.
Oct 2026 This post: an engine that holds thousands of games per process and forks them in microseconds, a sandboxed harness with interruptible skills, and a search algorithm that turns more compute into more progress.
On the method side, Prime-Agent-Explore descends from Go-Explore and its LLM successor Intelligent Go-Explore, and our harness experiments build on Continual Harness, GEPA and Prime Agent. The engine plugs into PufferLib and Prime Intellect's Verifiers.
What we learned
- Report BALmin alongside BALmax. Under BALmax diving looks like progress, and four of our five interventions would have looked like regressions.
- Single-life means should not be trusted at n=15. The five frontier models are within noise of each other, and taking each seed's best of three lives reorders them.
- Fix what the interface hides before blaming the model. Statues drawn as monsters, corpses that don't show their age, and prompts the harness answered silently each wrecked whole runs, while the five models we tried are within noise of each other.
- Search is where NetHack's headroom is. The same model with 30 checkpointed attempts went from dungeon level 8 to 18. However, in three simpler games the best of 10 fresh lives did about as well as checkpointed attempts. Whether search beats plain retries in NetHack, where the horizon is far longer, is our next experiment, which compares it against the best of 30 independent lives per seed.
Open questions
- How do the other frontier models rank by exploration score, and how much of it survives subtracting the best of 30 independent lives? These are the runs we plan to do next.
- Can Prime-Agent-Explore traces serve as imitation-learning data? Past imitation learning on NetHack was limited by the quality of AutoAscend's play, so Prime-Agent-Explore traces may be a better source.
- Given human games as an answer key, can an agent turn them into skills and memories that win, without memorizing the exact moves?
- Can the 100B-token Dungeons and Data corpus of human play warm up a model before RL in this environment, and can LLM-judged rewards in the style of Motif shape how it plays?
- Does a multi-game environment group (NetHack curricula alongside Crafter and Baba Is AI) measure generalist agents better than any one game?
Try it
Play NetHack+ in your browser at liu-jonathan.com/nethack. To run the evaluation yourself, see NetHack-hub.