Skip to content
TheChatArchiveThe Chat Archive
vicuna-33b· LMSYSHugging Face Arena Human Preference 55K

Lets discuss something. I've been trying to use chatbot LLMs to write a story in the…

From a Chatbot Arena battleThe same person sent this prompt to two anonymous models at once. Compare it with Metallama-2-13b-chat’s answer.
  1. User

    Lets discuss something. I've been trying to use chatbot LLMs to write a story in the format of a novel, creating first an overall scheme of the novel listing style, plot, chapter structure and more. Several problems have raised:

    -The context window is limited, so the LLM loses track of what the story is about after a few chapters. I have tried to mitigate this by reprompting the foundational document of the story and keeping track of important events and details in a list of notes that also gets reprompted, but doing this lowers the overall quality of the writing for some reason. A balance between continuity and quality is hard to find.

    Even when structuring in chapters, because the LLM can only type 2000 characters at a time it tends to write conclusions at the end of every single chapter that spoil the whole story and sound too grandiose, instead of patiently waiting for the conclusions and focusing only in the scope of the chapter itself. It doesn't "lose" itself in the prose and instead tries to rush things and explain instead of actually writing things that happen (like for example, explaining dialogues instead of creating them). Because it wants to wrap the story every chapter subconsciously due to it's fine-tuning as a chatbot. -Tendency to, every time I give instructions or feedback, to implement them in the text in an explicit manner instead of implicitly, for example if I say "the overall arch should be like a hero'es journey" it would just implement it directly in the story during the narrator voice saying something like "and the character follower the tapestry of his heroes journey". As I say, using instructions too explicitly instead of implicitly.

    And more. Some patches can be used to mitigate some of these problems, but I have reached the conclusion that instead of doing that, I should find a type of story that the LLM can write well (perhaps some unconventional or unexplored one) instead of trying to recreate a narration in the form of a regular human novel. Do you have any suggestions?

  2. Assistant

    It seems like you're facing some challenges in using LLMs to create a novel-like narrative. I agree that it might be more fruitful to explore alternative storytelling formats that align better with the capabilities and limitations of LLMs. Here are a few suggestions:

    1. Epistolary novel: An epistolary novel is told through a series of letters, diary entries, or other documents. This format allows you to create a story with a more fragmented narrative, which might work better with the context window limitations of LLMs. You can also use the documents to keep track of important events and details, making it easier to maintain continuity.
    2. Interactive storytelling: Instead of a traditional novel, you could create an interactive story where the reader makes choices that influence the narrative. LLMs can be used to generate branching narratives, and the interactive format would allow you to work around the context window limitation by focusing on smaller portions of the story at a time.
    3. Serialized short stories: Instead of writing a novel, you could create a series of loosely connected short stories. Each story could focus on a specific event or theme within the larger narrative, making it easier for LLMs to maintain continuity. You can then compile the short stories into a larger work, if desired.
    4. Experimental or avant-garde narratives: LLMs might be well-suited for creating experimental or avant-garde narratives that play with structure, format, and storytelling techniques. For example, you could create a story that is intended to be read non-linearly, with each chapter or section functioning as a self-contained vignette or fragment.
    5. Collaborative storytelling: You could use LLMs to generate the initial drafts of chapters or scenes and then collaborate with human writers to refine and expand upon the generated content. This approach can help to mitigate some of the limitations of LLM-generated text while still leveraging the power of the AI.

    By exploring these alternative storytelling formats, you might be able to create a more successful and engaging narrative using LLMs.

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.