KAWAKAMI, HARUYA

Research

I work on dual-head time-series generation models, in which a language decoder and an action decoder attend to each other and emit text and drawing operations on an external workspace in parallel, step by step. Much as people write out intermediate calculations or sketch a diagram to keep track of a story, externalizing the model’s reasoning into a structured visual space is intended to improve long-range and spatial consistency, and to make interactive collaboration between people and AI easier.

Awards

• JSAI SIG Research Award 2025 (SIG-AGI-032-05, “A Dual-Head Time-Series Generation Model for Simultaneous Language and Action Production”)