Translating stories to games with language models

10 Jan, 2026

First of all, happy new year to all!

I read this short paper by Daskalakis, et al.,¹ on translating stories to game-theory representations, e.g., game trees, payoff functions, etc., and then using game theory solvers to learn more about the strategic interactions of the story, e.g., by finding equilibria. The paper applies this to Romeo & Juliet as an example, generating different game paths (some of which lead to outcomes that are different from the common story). This way, the authors can explore aspects of the story from a quantitative perspective. This conversion of a story to a game-theoretic representation is also a theme in other works now, e.g., by Deng, Wang, and Savani².

The screenshot above is from the paper by Daskalakis, et al.¹ The authors even provide a colab for exploring the games they generated!

Having an LLM produce options for the Council of Elrond

I played around with having GPT-5.2 generate an extended form for the “Council of Elrond” scenario from the Lord of the Rings. Specifically, I wanted to see what kind of options the LM would cook up.

First try:

- **Destroy the Ring in Mount Doom (stealthy Fellowship):** **40 / 100**
    
- **Use the Ring against Sauron (a “good” wielder):** **20 / 100** _(beats Sauron as-is, but tends to “fail the future” by becoming the next Dark Power)_
    
- **Hide the Ring indefinitely:** **10 / 100**
    
- **Destroy by other means / elsewhere:** **5 / 100**
    
- **Open war now while securing the Ring:** **15 / 100**

But I wasn’t happy with the scores as they seemed to be biased towards the known story, so I asked GPT to re-estimate with reasoning. The answer isn’t bad:

- **Destroy the Ring in Mount Doom (stealthy Fellowship):** **25 / 100**
    
    _Hard mission, tiny margin for error, many points of failure (capture, betrayal, travel, surveillance)._
    
- **Use the Ring against Sauron (a “good” wielder):** **35 / 100**
    
    _As a pure “defeat Sauron” criterion, this could look tempting: power vs power. The Council fears the cost, not the feasibility._
    
- **Hide the Ring indefinitely:** **20 / 100**
    
    _Not a “win,” but the Council might believe in long games: keep it from him, let him exhaust himself elsewhere, wait for a shift._
    
- **Destroy by other means / elsewhere:** **15 / 100**
    
    _Unclear lore, but the Council might not assign it near-zero: they’d consider experimentation, Elven craft, Dwarven secrets, etc._
    
- **Open war now while securing the Ring:** **30 / 100**
    
    _They’ve seen alliances win wars before. It’s grim, but “win by arms + distraction” might feel comparably plausible to “one tiny quest.”_

Pretty fun! This can be further extended by looking at the possible effects of each action, etc.

Could they destroy the ring by other means?

Technically the "destroy by other means part" is a bit fuzzy -- Elrond did say something like "the ring cannot be destroyed by any art we here possess", so it does not seem like an easy destruction option is on the table. When pressed on that, the LM rationalized by saying "well, I meant to find arts they don't possess".