Research

The institute works in six areas. Two cover how adaptive systems are built, three cover fields where the institute applies them, and one covers how they are evaluated.

Systems

How adaptive systems are built

Adaptive and agentic AI systems

AI systems that change their behavior in response to feedback, and agents that plan and carry out tasks over many steps. This area studies how such systems are designed and how they are kept under human direction.

Questions

How should an agent that works over many steps report what it has done, so that a person can check it?
Which decisions should an agent stop and hand to a person?
How does an agent's behavior change when its tools, instructions or working environment change?

Visual communication and multimodal AI

AI that works with images, video, audio and text together. This area studies how such systems explain information visually and how well people understand what they produce.

Questions

When does a diagram, an animation or a narrated video explain a subject better than text?
How can a system check that a picture it produced says what the source material says?
How well do people understand and remember what these systems produce?

Fields

Where the institute applies them

AI for learning and education

AI used in teaching and learning, such as tutoring software that adjusts to what a student already knows. This area studies how to build these tools and how to measure whether students learn more with them.

Questions

How should tutoring software decide what a student works on next?
How can learning be measured at the level of understanding, beyond recall?
What do teachers need to see in order to trust and correct an AI tutor?

AI for defense and national security

AI that supports analysts and operators in defense work. This area studies how such systems perform when information is incomplete or an adversary is trying to mislead them.

Questions

How does a system's performance change when information is missing, late or false?
How should a system show an analyst what it knows, what it inferred and what it assumed?
How can shared definitions of entities and relations let systems from different sources work together?

Information security and AI cybersecurity

AI used to defend networks and software, and the security of AI systems themselves. This area studies how AI agents find and fix vulnerabilities, and how models and agents are attacked and protected.

Questions

How reliably can AI agents find and fix vulnerabilities in real code?
How are models and agents attacked through their inputs, tools and memory?
Which controls keep an agent's actions inside the limits its operator set?

Evaluation

How they are tested

Evaluation and assurance of AI systems

Methods for testing what an AI system does before and after it is put to use. This area studies how to gather evidence that a system will behave as intended, and how to report that evidence to the people who decide whether to field it.

Questions

What evidence shows that a system will behave as intended once it is in use?
How can each finding about an agent be traced to the record of what the agent did?
How should test results be reported to the people who decide whether to field a system?

Contact

Program managers, contracting officers and research partners can write to the institute through the contact form.

Write to the institute