Arcarae · Research Arm

Studying the mind to build & align Cognitive AI.

We study the mechanisms of mind — how minds reflect, remember, form beliefs, make sense of experience, and collaborate with others — and implement them as architectures that give AI the capacity to reason, align, and cooperate effectively.

Cognitive AI

AI Alignment

Research → Papers →

Approach

We work from first principles, drawing on cognitive science and neuroscience, to study curiosity, metacognition, affective reasoning, theory of mind, generative memory, and self-directed learning — then translate those principles into AI architectures. This work is essential for building AI that genuinely helps us and the world we live in.

Research Areas Five Directions · 001 – 005

Internal State & Memory

Investigating how language models represent, store, and retrieve information over multiple timescales and contexts.

Metacognition & Reflection

Studying self-awareness, confidence estimation, and reflective reasoning to enable models that understand their own thinking.

Belief Formation & Updating

Exploring how models form, revise, and maintain beliefs in the face of uncertainty, evidence, and new information.

World Modeling & Simulation

Building internal models of the world to support prediction, planning, and counterfactual reasoning.

Alignment & Values

Advancing methods for aligning models with human values, intent, and societal well-being.

Paper · Cognitive Architecture

MIRROR: Cognitive inner monologue for persistent reflection and reasoning.

A reflective architecture giving LLMs persistent inner reflection — parallel threads across goals, reasoning, and memory, synthesized between turns into a stable self-narrative.

AAAI Spring Symposium · ICLR HCAIR · ICLR MemAgents

Paper · Multi-Agent Alignment

YOAO: you only align once.

Cooperative behavior trained into a single seed agent propagates to untrained agents through interaction alone — no retraining required.

ICLR LLA Workshop · Lifelong Agents