Research Philosophy
Understanding intelligence
Building systems that know when - and when not - to act.
Predictive capability alone is insufficient for real-world operation. AI systems must reason explicitly about uncertainty, evaluate risk, and commit to action only when appropriate.
Himitsu Lab studies the mechanisms that enable decision completeness: the ability to act, wait, or seek human input based on evidence and operational constraints.
Why This Research Matters?
The Core Problem
AI systems often fail when:
- information is incomplete or delayed
- timing is critical
- actions are costly or irreversible
- ambiguity creates risk
- human escalation is necessary
These failures arise not from prediction errors, but from decision errors.
Our Mission
We investigate how AI systems can:
- build and update internal models of the world
- represent and quantify uncertainty
- choose whether and when to act
- remain safe under ambiguity and time pressure
Our goal is Native AI Intelligence that behaves responsibly in complex environments—not by producing answers, but by exercising judgement.
Our View of Intelligence
World Understanding
Systems must maintain internal models that reflect structure, constraints, and change—not just surface-level predictions.
Uncertainty-Aware Reasoning
Evidence quality determines conclusion quality. When signals are weak or contradictory, systems must lower confidence, defer, or escalate.
Safe Real-World Action
Intelligence requires understanding when not to act—especially when actions carry irreversible or costly consequences.