Zork Reimagined: Interactive Fiction in the Age of AI

Aleksey Tikhonov, Inworld.AIFollow

Submission Type

Paper

Start Date/Time (EDT)

18-7-2024 3:30 PM

End Date/Time (EDT)

18-7-2024 4:30 PM

Location

Hypertexts & Fictions

Abstract

Interactive fiction built on classic engines is commendable but faces known limitations: parser issues, a large amount of manual work, and default fallback responses in story branches not anticipated by the authors. On the other hand, contemporary neural network-based games, such as AI Dungeon, NovelAI, KoboldAI, and others using LLM-powered engines, offer unique adventures in open-ended worlds but often suffer from hallucinations and a lack of coherence, as well as missing a macro story structure.

Our project proposes using an LLM-wrapped interactive fiction, which aims to merge the strengths of both approaches. This means creating games with an open interface grounded in a formal story structure, formalized in languages like Inform 7, for example. This approach allows for mitigating parser errors, reducing/removing syntax requirements for the command interface, and supporting reactions to user actions that were not foreseen by the game's author, as well as organizing narrative personalization -- for instance, it can automatically style texts in a particular literary genre/form or alter modality (on an ideological level similar to that proposed by the Curveship project), and create a unique gaming experience for each player without compromising the story's structure.

As a proof of concept, we propose "ReZork" — a tribute to the classic game Zork where the original game is transformed into one of several selected literary styles (classical, poetic, Lovecraftian, etc.) while preserving the original storyline and overall game logic. We also augment the narrative with auto-generated illustrations using the Dalle-2 model. Currently, there is an early prototype, plus a demonstration video of the gameplay (link is omitted for anonymization).

We also endeavor to make it available as part of the experimental track (not guaranteed yet).

Bio

Aleksey Tikhonov is a scientist, researcher, and artist from Berlin, Germany, currently working as a Head of Research for Inworld.AI.

Aleksey does a lot of experiments in different media, including:

  • Paper "Unrolling Virtual Worlds for Immersive Experiences" at Machine Learning for Creativity and Design NeurIPS 2023 Workshop.
  • Spotlight Award at Machine Learning for Creativity and Design NeurIPS 2021 Workshop for CLIP-guided text2pixelart generation project.
  • ASCIIvyaz (2021) is an exploration into automating the creation of ligatures in the style of Slavic calligraphy, known as Vyaz. This project embarked on converting traditional Vyaz into ASCII art, leveraging a blend of algorithms for assembling ligatures in various sequences to achieve optimal aesthetic resemblance at the intersection of technology and traditional art forms.
  • Paranoid Transformer - the project on fiction generation, started as a NaNoGenMo (2019) entry; received New Sight award from Dead Alive Magazine, later published as a book in hardcover; the project was presented on NeurIPS ML for Creativity Workshop Gallery (2020), followed by the paper on Future Internet (2020).
  • Various projects on music generation, including Neuro-Scriabin — generation of music for a live orchestra in the style of Scriabin, was performed live on the opening of Yet Another Conference (2017); Neural music box — RPy-based infinite music generator, based on variational recurrent network, presented as a paper on CMMR (2017); Drum patterns from latent space — generation and interactive exploration of the drum patterns latent space (2019).
  • Various projects on stylized poetry generation, including Neural Defense and Neurona — neuro-tributes to punk groups of the 90s, were presented on NeurIPS ML for Creativity Workshop Gallery and published as EPs online.

This document is currently not available here.

Share

COinS
 
Jul 18th, 3:30 PM Jul 18th, 4:30 PM

Zork Reimagined: Interactive Fiction in the Age of AI

Hypertexts & Fictions

Interactive fiction built on classic engines is commendable but faces known limitations: parser issues, a large amount of manual work, and default fallback responses in story branches not anticipated by the authors. On the other hand, contemporary neural network-based games, such as AI Dungeon, NovelAI, KoboldAI, and others using LLM-powered engines, offer unique adventures in open-ended worlds but often suffer from hallucinations and a lack of coherence, as well as missing a macro story structure.

Our project proposes using an LLM-wrapped interactive fiction, which aims to merge the strengths of both approaches. This means creating games with an open interface grounded in a formal story structure, formalized in languages like Inform 7, for example. This approach allows for mitigating parser errors, reducing/removing syntax requirements for the command interface, and supporting reactions to user actions that were not foreseen by the game's author, as well as organizing narrative personalization -- for instance, it can automatically style texts in a particular literary genre/form or alter modality (on an ideological level similar to that proposed by the Curveship project), and create a unique gaming experience for each player without compromising the story's structure.

As a proof of concept, we propose "ReZork" — a tribute to the classic game Zork where the original game is transformed into one of several selected literary styles (classical, poetic, Lovecraftian, etc.) while preserving the original storyline and overall game logic. We also augment the narrative with auto-generated illustrations using the Dalle-2 model. Currently, there is an early prototype, plus a demonstration video of the gameplay (link is omitted for anonymization).

We also endeavor to make it available as part of the experimental track (not guaranteed yet).