MAGNET and ATLAS framework reduces hallucinations in long-form story generation

Researchers introduce MAGNET, a multi-agent goal-driven narrative engine that generates stories using persona-grounded character agents, alongside ATLAS, a graph-based pipeline for detecting hallucinations. The system tracks a shared world state and evolving story goals to maintain narrative consistency in long-form narratives.

• MAGNET utilizes persona-grounded character agents to propose actions based on a shared world state.
• ATLAS compares scene-level world representations across the generated story to identify inconsistencies.
• At 100 pages, the framework reduced annotations by 41% and hallucinations by 50% compared to a single model baseline.
• The approach also outperformed IBSEN, reducing annotations by 34% and hallucinations by 45%.

The results suggest that explicit world-state tracking and goal-driven multi-agent generation provide a foundation for controllable and structurally coherent long-form narrative generation.