Calypso: LLMs as Dungeon Masters' Assistants [pdf]

> It was amazingly satisfying to all my players, and it's also the exact sort of thing that I don't think the current batch of SOTA LLMs (even with the temperature cranked to 1) would ever think of in a MILLION years.

That actually seems like it'd be a pretty easy one for an LLM to get to. I tried

    You are a highly intelligent and creative D&D campaign planner. Imagine a creature that is immortal with mechanics like The Picture of Dorian Gray - its soul is protected in a painting and that gives its body immortality. Please brainstorm some interesting ways to defeat this antagonist that might connect to other aspects of a campaign, suggesting new ideas.

and on the 2nd run on a small local LLM got "*Reflections and Mirrors*: The creature's soul is reflected in mirrors and reflective surfaces, making them vulnerable to attacks. The party must use mirrors to their advantage, luring the creature into a trap or using them to distract it while they attack." and "* *The Corruption of Beauty*: The immortal creature is a symbol of beauty and corruption. To defeat it, the party must find a way to strip it of its beauty and corruption, using it against itself. This could involve using a magical artifact that amplifies the creature's own corruption, or finding a way to expose its true nature and render it powerless.".

It seems quite reasonable that a large modern LLM could come up with exactly that idea in far less than a million years. Art, mirrors and defacing art are all pretty standard sounding themes for handling that sort of monster.

I disagree - a generic suggestion to use a "reflection to defeat Dorian" isn't at all the same thing as standing up two mirrors to create a hall of mirror effect on a magic mirror that "ages" the world around it. The key difference is in the level of specificity.

Furthermore, the true test of the LLM would have been to devise both deciding to use a "Dorian Gray" type creature along with an appropriate mechanism for its demise.

IMHO, you've guided it by posing that this monster exists within the world. I'd like to see an entire campaign devised by an LLM - no guidance necessary. I don't know what small model you are using (mistral, vicuna, etc) but try asking it to create a list of monster based encounter/puzzles.

This is why I think the current batch of LLMs aren't really capable of being more than an assistant at best for writing. Interestingly I actually think AI dungeon released back in 2019 was far more capable in this regard and IIRC used a significantly older model - GPT 2.0.

I added "Be really specific" to the prompt and on the second go got:

11. *The Art of Decay*: The Portrait's soul is tied to the painting, but what if the players could use the painting's power to accelerate the decay of The Portrait's physical form? This could involve using a powerful magical effect or finding a way to imbue the painting with a magical property that would accelerate the decay process.

in the brainstorm. I don't think it'd take a million years; it's got the concepts and it is trying to put them together with only 5 attempts so far; and that is only 2Gb of weights on a relatively cheap GPU.

As for coming up with an entire campaign that copys an idea from Wilde, well it depends and I'd agree that is a task that would want a human in the loop. It does seem difficult to do in one prompt - but I think a million years is wildly overestimating the novelty of the campaign you're describing. It is actually damn hard to come up with good, novel ideas and the odds are the LLMs have literally seen variants of the Dorian Grey idea 100 times.

It's just describing the fact that the players need to find a way to age the portrait. At the risk of being a bit glib, I mean, duh. And its solution is the rather pedestrian "powerful magical effect or finding a magical way" jazz hands pile of claptrap.

I will admit that a million years was a bit of hyperbole.

I’m only barely familiar with the actual underlying story—is aging the portrait actually obvious? I thought the solution was to destroy it. Isn’t the whole point of the figure that it ages (instead of Dorian)?

I would disagree, but on,y considering I think this is not the right prompt to test against. Hence not the right question.

While you're asking to find creative ways to get rid of the player, I think what LLMs are unable to do (at this point?) is to come up with the idea of an aging mirror, let it sit for a while and only then get back to it when they met its attached character.

The dungeon master did not follow a track of events but rather picked interesting somewhat random contents and moments from the campaign and picked them up to create new story lines.

That doesn't seem like something an LLM would do easily.

The fact is that a small llm can come up with scenarios that would be kind of like it, and you are dismissing that because you want humans to be the only things capable of originality. In fact you lifted an entire aspect of the campaign from a classic novel, and the mirrors being magic concept is pretty much ingrained in cultural tropes, so I'm not sure why you think that system which has as its basis the entirety of human written output which was designed to predict more output couldn't come up with it.

" and you are dismissing that because you want humans to be the only things capable of originality."

I have not seen convincing cases of LLM's being capable of orginality. But they do have a lot of originality in their training data.

But then, human artists have a lot of training and influences too, so it's not so easy for humans to be truly original either (if that is a thing).

And then, if you are too original, no one will like it either ..

I don't think there is something "truly original" but in this concrete case, I suspect something very similar was in the training data.

I'm not dismissing LLMs, but there's a reason that the current crop of LLMs aren't writing hit scripts for TV shows, books, etc. The suggestion to use a "magical effect" is the equivalent of asking an LLM, "How can I break into this password protected computer?", and it responds, "You should use an exploit.".

Its not something an LLM would do with simple chatGPT type prompts, but it's something I can imagine building an agentic system to do. It's not trivial but seems feasible with current day LLMs.

If you design the system to have this exact quality (among many others), where clues are dropped earlier in the quest line for later quests. It's a matter of breaking up the prompts and iteratively refining the outputs.