Advanced AI Research
Advancing Native AI Intelligence through world understanding, uncertainty-aware reasoning, and safe real-world action.
Himitsu Lab's research focus begins with Native AI Intelligence—an approach where reasoning, autonomy, and safety emerge from coherent internal processes rather than external orchestration. Our work contributes to the design of advanced Artificial Intelligence systems and future AI platforms capable of reasoning, learning, and acting safely under uncertainty.
This direction represents a significant foundational step toward the future, it emphasizes decision-complete systems that can understand context, evaluate uncertainty, act with intention, and learn from outcomes. This includes AI safety research focused on decision timing, uncertainty handling, escalation pathways, and safe plan-to-act execution. Only a small number of research groups globally pursue this depth of long-horizon work, and our contribution is shaped through sustained research and experimental system development.
Five Core Research Pillars
These five research pillars together form the foundation of next-generation Artificial Intelligence technologies and solutions designed for real-world environments.
World Model+ Decision Architecture
How systems understand the world and decide when to act.
World Model+ is our decision-complete architectural framework for Native AI Intelligence. It builds on classical 'World Model' and extend it with uncertainty-aware reasoning, decision timing, escalation pathways, safe plan-to-act execution, and continuous learning.
Why it matters:
Real-world intelligence requires more than prediction. Systems must know:
- when to act,
- when not to act,
- and what to do when they do not know what to do.
Key Themes:
- hierarchical world models
- evidence-driven decisions
- safe commitment & deferral
Cognitive & Reasoning Models
How AI interprets context and makes explicit, uncertainty-aware decisions.
We design reasoning systems that understand meaning, assess evidence, and adapt under ambiguity. This enables clarity in environments where pattern recognition alone is insufficient.
Key Themes:
- explicit inference
- uncertainty-aware reasoning
- metacognitive control
Swarm & Distributed Intelligence
How multiple agents coordinate under uncertainty.
We develop distributed systems where agents share beliefs, align uncertainty, and collaborate safely—even with partial information. This strengthens resilience across failures or adversarial events.
Key Themes:
- multi-agent coordination
- uncertainty alignment
- resilient collective action
Quantum Security & Long-Horizon Risk
How intelligent systems remain trustworthy over time.
We study risks that evolve across long horizons—delayed feedback, adversarial adaptation, and quantum-era threats. Our work includes post-quantum security foundations embedded directly into AI systems.
Key Themes:
- post-quantum security
- adversarial risk
- long-term safety
Physical & Embodied AI
How AI operates safely in the physical world.
We integrate cognition into robots, drones, and autonomous devices that must act under noise, uncertainty, and real-time constraints. These systems know when to act, pause, escalate, or refine understanding.
Key Themes:
- embodied decision-making
- safe action under uncertainty
- human-aware autonomy