LLMs Get Creative: Collective Critics for Storytelling

📄 Full Paper

💬 Ask

Large language models (LLMs) are revolutionizing how we write, but getting them to generate truly creative long-form stories is a challenge. Most approaches focus on narrative coherence, but lack the spark of creativity and expressiveness that makes a story truly captivating. A new paper published by Minwook Bae and Hyounghun Kim introduces a novel framework called CRITICS, which aims to address this gap by incorporating collaborative critique into the story generation process.

Think of CRITICS as a group of LLM “critics” who work together to refine a story plan and then the story itself. The framework has two main stages:

1. CRPLAN (Plan Refinement): This stage involves a group of LLM critics, each with a specific persona and expertise, who provide suggestions for improving the story plan. For example, one critic might focus on making the plot more original by suggesting unexpected twists, while another might suggest changing the setting to create a more unique and engaging atmosphere. A “leader” critic then evaluates these suggestions and selects the best one, which is then applied to refine the story plan. This process is repeated iteratively to create a highly refined plan.

2. CRTEXT (Story Generation): Once the plan is finalized, another group of LLM critics, each with expertise in specific criteria such as “Image” or “Voice,” provide suggestions for refining the story text. For example, one critic might suggest using more vivid imagery or descriptive language, while another might suggest using a more unique and engaging voice. Again, a leader critic evaluates these suggestions and selects the best one, which is then applied to refine the story text.

The researchers demonstrate that CRITICS can significantly enhance story creativity and reader engagement, while also maintaining narrative coherence. They conducted human evaluations of stories generated by CRITICS and compared them to those generated by a state-of-the-art baseline system. The results show that CRITICS significantly outperforms the baseline in terms of creativity and interestingness. The authors also note that CRITICS allows for active participation from human writers, enabling an interactive human-machine collaboration in story writing.

Here are some concrete examples to illustrate the paper’s main ideas:

• Enhancing a story’s setting: In the example of Aimee Kincaid’s loneliness, CRPLAN helps to make the story more engaging by suggesting that Aimee’s loneliness should be personified as a physical entity. This idea, while seemingly simple, is a powerful one because it adds a layer of depth and complexity to the story, making it more emotionally resonant.

• Improving a story’s language: In the example of a story about a young man named John, CRTEXT helps to make the story more expressive by suggesting using more vivid imagery or descriptive language. For example, the phrase “He leaned against the moss-covered trunk of a nearby tree, peering at Jonathan with curious eyes,” could be improved by using a more unusual verb such as “arched,” making the sentence more evocative.

The researchers believe that CRITICS is a significant step forward in the development of LLM storytelling. By incorporating collaborative critique, CRITICS can help LLMs to generate truly creative and engaging long-form stories that are more than just coherent narratives. The paper opens up exciting possibilities for the future of LLM storytelling, and it will be interesting to see how CRITICS is further developed and applied in future research.

Chat about this paper

To chat about this paper, you'll need a free Gemini API key from Google AI Studio.

Your API key will be stored securely in your browser's local storage.