Improving Collaborative Storytelling with a Multi-Agent Framework Based on Large Language Models

Technical Summary: Improving Collaborative Storytelling with a Multi-Agent Framework Based on Large Language Models

Problem Statement
The paper addresses the challenge of generating high-quality, appropriate, and coherent narratives for young children (ages 3–6) within a collaborative storytelling context. While existing AI co-creation research predominantly targets adult users in digital environments, this study focuses on a physical, screenless board game called YOLI. In this setting, children select physical tiles representing story elements (protagonist, location, mood, object, activity), which are then used by an AI to generate a story.

A critical technical hurdle identified is the issue of faithfulness hallucinations in Large Language Models (LLMs). In this context, hallucinations refer to deviations from user instructions or input contexts, where the generated story fails to accurately incorporate the specific elements selected by the children. Such failures can lead to user frustration and disengagement. The paper posits that standard single-pass generation often lacks the necessary refinement to ensure strict adherence to these constraints while maintaining narrative quality.

Methodology
The authors propose a multi-agent framework centered on an iterative Writer–Editor loop. This system operates entirely at inference time without requiring additional training or fine-tuning.

1. The Game Loop: Children select tiles on the YOLI board. These selections serve as input constraints for the story generation system.
2. The Writer Agent: An LLM (specifically Gemma 2 2B in the study) acts as the "Writer." It receives a prompt containing the selected story elements and generates a narrative.
3. The Editor Agent: A separate LLM acts as the "Editor." It evaluates the Writer's output against a standardized rubric focusing on Relevance (alignment with input tiles). The Editor assigns a numerical score (0–100%) and provides a rationale/critique.
4. Iterative Refinement:
   • Loop 1: The Writer receives a zero-shot prompt to generate an initial story.
   • Loop 2+: The Writer receives a one-shot prompt that includes the original story, the Editor's critique, and the score. The Writer attempts to generate an improved version.
   • This process repeats for up to five iterations, with the Editor re-evaluating the new output in each cycle.

Simulation Study Design
To evaluate the framework, the authors conducted a simulation study using a dataset of approximately 1,000 tile tuples (combinations of story elements).

• Writer: Fixed as Gemma 2 (2B parameters).
• Editors: Varied across different model families and parameter sizes: Gemma 2 (2B, 9B, 27B), Llama 3.1 (8B), and Mistral (7B).
• Metric: The primary metric was the "perceived quality" score assigned by the Editor LLM, tracking the trajectory of improvement across iterations.

Key Results

1. Consistent Quality Improvement: Across all Editor models tested, the average perceived quality score increased with each successive Writer–Editor iteration. This confirms that the iterative critique process effectively guides the Writer to refine its output.
2. Diminishing Returns: The most significant improvement occurred in the transition from the first to the second loop (moving from zero-shot to one-shot prompting with feedback). Subsequent iterations (loops 3–5) yielded progressively smaller gains.
3. Convergence Speed: The study suggests that the optimal stopping point for the iterative process lies within five iterations. Survival analysis indicated that for Editors with 7+ billion parameters, there is a 90% likelihood of reaching their maximum score within five loops.
4. Model Capacity Correlation: Larger Editor models (e.g., Gemma 2 27B) demonstrated a higher probability of reaching their peak score earlier in the process compared to smaller models (e.g., Gemma 2 2B), suggesting that more sophisticated evaluators can guide refinement more efficiently. However, even smaller models showed consistent improvement trends.
5. Self-Preference Bias: While the Gemma 2 2B model acting as both Writer and Editor achieved the highest absolute scores, the authors note this may indicate a self-preference bias. However, manual inspection suggested that the use of a guided rubric mitigated this issue, and larger models (like Llama 3.1 8B) still achieved comparable final scores (89% vs. 91%).

Significance and Claims
The paper claims that the proposed Writer–Editor framework offers a lightweight, practical alternative to training-intensive alignment methods like Reinforcement Learning from Human Feedback (RLHF) or Reinforcement Learning from AI Feedback (RLAIF).

• No Training Required: The system relies solely on inference-time interactions, eliminating the need for expensive, labor-intensive human preference datasets or reward model training.
• Scalability for Child-Centered Design: The framework effectively balances the open-endedness required for creative play with the strict constraints necessary for child-appropriate content, specifically addressing the problem of faithfulness hallucinations.
• Efficiency: The finding that high-quality outputs can be achieved in a small number of iterations (2–3) makes the system viable for real-time interactive storytelling, where latency and responsiveness are critical.

Limitations
The authors modestly acknowledge that the quality scores are synthetic, derived solely from LLM-based Editors rather than external human evaluation. While a preliminary qualitative assessment by professional writers and a YOLI designer suggested alignment with human judgment, the paper emphasizes that future work must validate these scores against assessments from early-childhood educators and direct ratings from children to ensure the improvements align with genuine human-centered engagement.