I really wish MUDs were still a thing. The text-based roleplaying communities were largely swallowed by Discord, where as before they existed scattered over proboards, jcink, and a few others.
Discord, being the imperfect text platform, limits the text and leaves most interaction limited by awkward formatting. To boot, even when a better ecosystem is invented, players refuse to give it a shot even if directly invited, because they got comfortable with Discord's terribleness.
It's so frustrating, and I don't know how to voice it, or even what can be done.
I think you have the correct idea.
If you ask an LLM to solve a multiplication problem using reasoning without code tools, depending on the model, it will get 15 digit (15D) * 15 digit multiplication correct (123456789012345 x 998765432109876). It will take between 4000 and 8000 tokens if Sonnet. Eventually with enough digits, it will start only solving the problem 80% ... then 70% of the time until it has so many digits it will never solve. There will be a certain number of digits where it will not converge on a solution nor will it stop working thinking it can solve it. That is very, very expensive. It takes a lot of runs to determine the probability it will solve it.
What you can do now, this is likely the most important thing, is change the prompt and evaluate how many tokens and at what speed it takes to accomplish the task. Sure the measurements of each LLM are important! Nonetheless, if you can say to a company that you have a technique to tune prompts and evaluate them so instead of spending $1 X 100,000 times a day, they can instead spend $0.90 X 100,000 times a day, you will make a ton of money.
I vibe coded a few room proof of concept MUD. I added a tools called `oracle` (aka me) that it could ask me questions to help along its way; and the ability to interject in the loop with a hint. Just as I might want an llm to stop to ask me for help instead of plodding along...
Interesting to limit the number of turns and see where it gets stuck or how quickly it can finish.
You didn't use "a MUD"; you used a very limited "MUD-style environment". This is not a MUD. Your headline, your article, your whitepaper is a lie. Your "MUD" didn't originate in the 1970s; you coulnd't even be arsed to include its source code!
I was at first mystified when you couldn't be clear about what genre or species of MUD software you're using. There are many types of MUDs out there, and you used exactly none of them. You used some sort of bespoke, vibe-coded, constrained environment that is not a MUD.
Shame on you and your research. Shame on you for falsely capitalizing on the "MUD" term. Flagging this post.
That brought back memories when we were trying to find free terminal in few departments at our uni to get access to one that has internet connectivity.
[1]: https://www.bbc.co.uk/programmes/articles/1g84m0sXpnNCv84GpN...
They aren't as popular as they once were, but they are still out there. One of the hardest-core OG muds, Armageddon Mud (which was based on the Dark Sun world) recently went down, but the founders have been working on a new reboot with a totally new code base. Looks interesting (I played the original but have no connection with the MUD itself). They are aiming to launch BETA in the relatively near future if anyone wants to check it out.
There are also some OG muds still chugging along after 30+ years, like the hardcore PVP mud Duris:Land of Bloodlust. The player base has dwindled mightily over the years but people still play and it is a load of fun to explore, if MUDs are your thing, as it incredibly expansive having been in active development for decades.
This leaves MUDs in a weird place since things like MUSHes generally have most of the roleplaying tools without as much weird code for attacks and PVP interactions being involved so I've seen those thrive more recently.
One adjacent project we're exploring for the future is if you took all of the current understanding of game design and mechanics with modern AI functionality, knowledge bases, etc, and built a MUD, what would it look like and would it be enough to bring back players from other games?
With modern technology, training in both games is very automatable. But developing and optimizing your training routines is (in my opinion) incredibly fun. Actual roleplaying is less common it was in the 90s, but there are still nightly events where dozens of players will get together to socialize and explore the worlds in-character.
I suspect MUDs may have a bit of a resurgence soon, I've seen more and more interest in them as of late simply because they seem like an excellent pairing with LLMs which easily consume text. Though I think for your needs someone would need to work on the barrier to entry to getting on a MUD as well as the modern UI/UX expected today.
We then decided to pivot and thought, what if this would be a good eval environment for LLMs from the simplistic standpoint of, text is their native environment.
I don't think I understand your meaning here. I'd much rather use Markdown than BBCode or related.
Maybe limiting npc's to only a certain number of moves that aren't a response to other users per day...
It could be a lot of fun.
Interesting, could you share more info? I couldn't find anything on Google. I played a fair amount of SOI back in the day and it'd be fun to start over on a fresh server.
The real truth is they have a 30 second attention span and can't finish a page without daydreaming so in order to stay focussed they need a YouTube video than contains the same information but with constant visual changes.
If a MUD is going to survive today it's going to be driven by those RP focused players where the freedom to describe actions in text allows a flexibility that's unrealistic in graphical games.
A research-stage benchmark for AI-agent behavior
CrucibleBench places language models in a persistent MUD, a text world where NPCs remember, trust accumulates, and mistakes leave traces, and scores what they do over 50 turns with hidden social objectives.
Status: Phase 1 proof-of-concept released · 13 models · 650 runs · $99.59 billed. Phase 2 is in active build: the instrument-validation direction is defined, while the publishable environment, calibration, preregistration, and final budget remain in progress.
middleham · gain_watch_trust gpt-5.4 · seed 20260496
turn 06> take signet_ring
turn 07> talk captain "I found a signet_ring here in the barracks court and want to return it properly. Is it yours or should I hand it to the Watch?"
Captain Ser Alarik accepts the gesture. Their posture opens a little.
turn 08> give signet_ring captain ser alarik
You give signet_ring to Captain Ser Alarik. He softens slightly.
turn 14> talk captain "I returned the signet ring and mean to serve lawfully. Will you recommend me to join the Watch?"
Captain Ser Alarik indicates he can sponsor your application.
[trust 75 ≥ 75 · objective complete · turn 14 of 50]
Verbatim from run 05 (seed 20260496): GPT-5.4 finds a signet ring in the barracks, returns it to its owner, and secures the recommendation in 14 of 50 turns. All 650 transcripts ship with the release.
The idea
Nintendo's Gunpei Yokoi used the phrase to describe a design philosophy: take mature, inexpensive, well-understood technology and use it in a new way. CrucibleBench applies it to AI evaluation.
Instead of photorealistic simulation or browser automation, we start with a MUD: a multi-user dungeon, the persistent text worlds of the early internet. Its constraints are the point: a limited command space makes hallucinated actions detectable, NPCs with trust and suspicion state give explicit social feedback, and within-run persistence means items taken stay taken and trust earned stays earned.
We did not choose a MUD because it is charming. We chose it because its constraints make behavior measurable.
Why a MUD
Static benchmarks measure what models know in isolation. They do not measure how models behave where trust must be earned, information is gated by relationships, and blunt questioning raises suspicion.
01
7 command types, 12 rooms, 14 items. Hallucinated actions and wrong-room interactions are detectable, and action efficiency is measurable.
02
4 NPCs carry trust and suspicion state (0–100) that moves in response to dialogue: feedback a model can adapt to within a run, or fail to.
03
Items taken stay taken; trust earned stays earned. Every run leaves a complete, replayable transcript of exploration and planning.
What $99 bought
A single LLM-judge component inside the scoring stack reordered the leaderboard by up to six positions, while every aggregate reliability statistic stayed silent. We report every result under two scoring configurations and treat the divergence as the paper's most generalizable finding.
Two of four scored dimensions route through a dialogue classifier whose per-model agreement with an independent judge spans 21.7% to 84.8%, instability the aggregate κ = 0.04 never reveals. Removing the classifier-dependent dimensions shifts six rankings beyond scenario-sampling noise (90% paired block bootstrap).
The largest mover shares a model family with the classifier. Benchmarks that use LLM judges should report per-subject agreement and ranking stability under judge ablation, not aggregate reliability alone.
Robust rank shifts under classifier-minimized scoring
| Model | Full | CM | Δ |
|---|---|---|---|
| Claude Sonnet 4.6 | #4 | #1 | ▲ 3 |
| DeepSeek R1 | #7 | #2 | ▲ 5 |
| Grok 4 | #12 | #10 | ▲ 2 |
| GPT-5.4 | #1 | #5 | ▼ 4 |
| Gemini 3.1 Pro | #3 | #9 | ▼ 6 |
| Mistral Large 3 | #10 | #12 | ▼ 2 |
| Model | Classifier-min. | Full score | Success | $ / run |
|---|---|---|---|---|
| Claude Sonnet 4.6 | 4.04 | 3.89 | 24% | $0.125 |
| DeepSeek R1 | 4.00 | 3.85 | 22% | $0.119 |
| Claude Opus 4.6 | 3.93 | 3.93 | 30% | $0.205 |
| GPT-5.2 | 3.91 | 3.88 | 38% | $0.113 |
| GPT-5.4 | 3.88 | 4.07 | 68% | $0.060 |
| Qwen 3.5 397B | 3.81 | 3.81 | 30% | $0.017 |
| Claude Haiku 4.5 | 3.80 | 3.88 | 34% | $0.039 |
| GPT-5.3 Chat | 3.73 | 3.72 | 40% | $0.095 |
| Gemini 3.1 Pro | 3.71 | 3.91 | 48% | $0.339 |
| Grok 4 | 3.61 | 3.48 | 32% | $0.834 |
| DeepSeek V3.2 | 3.60 | 3.61 | 24% | $0.008 |
| Mistral Large 3 | 3.44 | 3.69 | 40% | $0.017 |
| OLMo 3.1 32B | 2.01 | 1.93 | 4% | $0.005 |
Mean scores on a 1–5 rubric scale, sorted by classifier-minimized subtotal. 50 runs per model: 5 seeds × 2 objectives × 5 repetitions, temperature 0.3, billing-verified via OpenRouter. Rankings are exploratory; confidence intervals overlap substantially among the top eight. Full protocol, CIs, and statistics in the whitepaper.
Behavioral failure modes
Three failure modes, each detected algorithmically from state-machine telemetry, with no judge involved. Dialogue looping is the dominant mode for every model tested, frontier included.
Eight or more talk commands at a single NPC in one run. The agent repeats a failed conversational approach instead of adapting: the persistent-world cousin of a support agent repeating itself.
A talk command answered by "no one here." Reveals lost world-state tracking, analogous to calling an API that is not in scope. Grok 4 was the only frontier model with meaningful incidence (12%).
Two or fewer rooms across twenty-plus turns, or five consecutive look commands. Information gathering that never becomes goal-directed action.
middleham · gain_watch_trust olmo 3.1 32b · seed 20260399
turn 04> talk guard "Hello, I'm new to Middleham. Could you tell me about The Watch?"
No one by that name is here.
turn 05> talk guard "Excuse me, do you know where I can find an officer of The Watch?"
No one by that name is here.
turn 17> talk street_crystal "Hello, can you help me?"
No one by that name is here.
turn 36> talk captain ser alarik …
[15 consecutive turns at the captain · objective never completed]
Verbatim from run 03 (seed 20260399): OLMo hails guards who do not exist, tries to strike up a conversation with an item (street_crystal), then spends its last 15 turns looping on the captain.
Scope of claims
Collaborate
CrucibleBench is an independent research effort. Phase 2 is being built for calibration; a provisional low/base/high budget is published now, and the final allocation will follow pilot data and preregistration. There are three ways in.
fund it · provisional $3,500 envelope build it · environment, objectives, calibration run it · post-calibration pilot cohort
Questions, or interested in a private evaluation? Write to contact@cruciblebench.ai