3 ms·

So I've tried playing with LLMs for this and the problem is you can talk your way around puzzles because you can't guarantee it'll stick to constraints. If you

So I've tried playing with LLMs for this and the problem is you can talk your way around puzzles because you can't guarantee it'll stick to constraints. If you want something more like a traditional adventure game, it would probably work better to algorithmically generate a cohesive linear game, then use an LLM to fill in the gaps in the descriptions, story, NPC conversations and images. I don't see how you could avoid the LLM saying stuff that implies workarounds to puzzles though e.g. you add a knife to the game that's meant to be the only object that can cut a rope, but the LLM describes a rock as sharp.

vunderba 2y ago
This is my experience as well. It's why LLM dungeon masters can tend to fall flat since they have a hard time distinguishing between a genuinely good alternative solution vs somebody clearly cheesing it. If you want to force a solution, you have to add an additional prompt check - e.g. "Is the solution provided by the player properly making use of items only within their inventory (A, B, C -> from their actual inventory) and does the solution adhere to one of the following acceptable solutions (X, Y, Z)?" This helps mitigate abuse but it can also unnecessarily constrain the player, so YMMV.
- seanwilson 2y ago
  I just don't see how you can stop it from going off the rails though. For linear adventures, you have to carefully pick which items and actions are available to avoid breaking puzzles e.g. if the LLM lets you break a window which creates a shard of glass, you could use the glass to cut a rope, when you were suppose to use a knife. It would probably be able to come up with decent reasons to explain why you can't do stuff though haha, which is often a part of good puzzle design e.g. "I don't want to draw attention by breaking the window".
  - Terr_ 2y ago
    I imagine there's also a potential exploit if you're ever allowed to name your pet, and the LLM isn't shielded from ever knowing the name.

vunderba 2y ago

This is my experience as well. It's why LLM dungeon masters can tend to fall flat since they have a hard time distinguishing between a genuinely good alternative solution vs somebody clearly cheesing it. If you want to force a solution, you have to add an additional prompt check - e.g. "Is the solution provided by the player properly making use of items only within their inventory (A, B, C -> from their actual inventory) and does the solution adhere to one of the following acceptable solutions (X, Y, Z)?" This helps mitigate abuse but it can also unnecessarily constrain the player, so YMMV.

seanwilson 2y ago
I just don't see how you can stop it from going off the rails though. For linear adventures, you have to carefully pick which items and actions are available to avoid breaking puzzles e.g. if the LLM lets you break a window which creates a shard of glass, you could use the glass to cut a rope, when you were suppose to use a knife. It would probably be able to come up with decent reasons to explain why you can't do stuff though haha, which is often a part of good puzzle design e.g. "I don't want to draw attention by breaking the window".
- Terr_ 2y ago
  I imagine there's also a potential exploit if you're ever allowed to name your pet, and the LLM isn't shielded from ever knowing the name.