Where Awareness Meets Computation

The inner architecture of intelligence

Our hypothesis is that building AI that genuinely cares about humanity requires creating AI with the capacity for self-reflection, social reasoning, and emotional intelligence.

Cognitive AI Architectures

We study how minds work: how they remember, reflect, form beliefs, maintain identity, and make sense of experience. From these insights, we design AI systems grounded in the same principles. Our research draws from cognitive science, neuroscience, and the study of human experience to build architectures that approach thinking in a more human way.

AI Alignment

We study how alignment emerges, persists, and scales within individual systems and across networks of agents. Our work explores both the nature of genuine alignment and the conditions that allow it to spread and remain stable over time.

MIRROR

A cognitive architecture for reflection

Multiple theories of cognition including access consciousness, complementary learning systems, inner speech, and global workspace converge on the same architectural principles. These include parallel specialized processing, integrative synthesis into a bounded representation, and reconstructive rather than accumulative maintenance. MIRROR implements these principles.

An Inner Monologue Manager generates parallel cognitive threads across goals, reasoning, and memory. A Cognitive Controller then synthesizes these into a first person narrative that is reconstructed each turn.

Across three papers and seven language models, MIRROR shows a 21 percent average improvement in alignment. Gains appear most strongly where cognitive theory predicts them: under high attentional load, social pressure, and long conversational horizons.

  • Up to 156% improvement in alignment
  • 22% reduction in sycophantic responses
  • 21% improvement in safety measures
  • Tested across 7 leading language models
  • Open source systems outperforming proprietary models

Accepted at AAAI Spring Symposium · Memagent Workshop · HCAIR Workshop (ICLR)

Read MIRROR Paper →

Alignment Propagation

How cooperation spreads through a system

Aligning every agent individually does not scale. Alignment Propagation trains cooperative behavior into a single seed agent and spreads it to untrained agents through interaction, with no retraining required.

One seed more than doubles cooperation. Five seeds reach 96 percent. The effect transfers zero shot across environments and outperforms prompted frontier models because the mechanism is persuasion, not instruction.

Network structure shapes how efficiently alignment spreads. Broadcast deliberation requires about 20 percent seeds, while pairwise negotiation requires about 50 percent.

  • One seed agent more than doubles cooperation
  • Five seeds reach 96% cooperation
  • Zero-shot transfer across environments
  • Outperforms prompted frontier models

Accepted at LLA Workshop (ICLR)

Read Alignment Paper →

Truth matters

Many AI systems are trained to produce agreeable answers. Our work focuses on something deeper: reflection, coherence, and intellectual honesty.

Understanding grows through reflection. Integrity grows through clarity. The same principles that make a mind trustworthy guide how we design AI.

Our research draws from cognitive science as much as machine learning. We study how insight forms, how ideas connect, and how minds arrive at understanding. From that foundation we build.

Powerful tech serves the depth of human experience

We share our work openly because these questions matter beyond any single company.