[2E6-GS-10n] AI application

5:30 PM - 5:45 PM JST(8:30 AM - 8:45 AM UTC)

[2E6-GS-10n-01] Boltzmann-GPT: Persona-Based Conditional Review Generation via World Models

〇Junichiro Niimi¹ (1. Meijo University)

Keywords:

large language model,world model,representation learning,energy-based model,marketing

While large language models (LLMs) are utilized across diverse tasks, discussions continue over whether they explicitly understand the structure of the world. This study proposes an architecture that separates the world model for unsupervised learning of domain structure from the language model for text generation. We employ a Deep Boltzmann Machine (DBM) as the world model to learn co-occurrence structures among variables such as consumer purchase behavior and product ratings. The belief vector obtained from the DBM's hidden layers is transformed into soft prompts via an adapter and fed into a frozen GPT-2, such that the world model determines what to say while the language model determines how to say it. Experiments using Amazon review dataset demonstrate that i.) the world model alone comprehends market structures (e.g., brand-price dynamics), and ii.) the proposed model coherently reflects customer ratings and preferences in generated content, whereas the baseline model exhibited information gaps and unintended content. These results suggest that even small-scale language models can achieve controllable generation through appropriate integration with world models.