Notice changes, not just scenes.
Track what moved, unlocked, disappeared, or became reachable after each intervention.
Case study · Adaptive agents
Exploration in unfamiliar worlds.
An intelligent system cannot always wait for a complete map. We study how an agent can gather the right evidence, discover controllable state, and choose useful actions while its model of the world is still changing.
- Status
- Active research direction
- Central question
- How should an agent decide what to learn before it knows what matters?
- Evidence standard
- Interactive outcomes, documented evaluation context, and explicit failure analysis.
01 · Question
What makes an action informative?
In a new environment, progress depends on more than selecting the action that looks best now. An agent may need to probe an object, revisit a location, or test a rule simply to reduce uncertainty. The research problem is to balance purposeful progress with experiments that make later decisions better.
We are especially interested in environments where goals must be inferred from interaction, state is only partially visible, and an early assumption can send a long plan in the wrong direction.
02 · Approach
Treat exploration as part of reasoning.
Keep hypotheses revisable.
Represent possible rules and uncertainty instead of collapsing too early onto one explanation.
Use plans as probes.
Prefer actions that can advance the goal while also distinguishing between competing world models.
Let surprise change the route.
When an outcome breaks the plan, update memory and re-plan from the new evidence.
03 · Evidence
A public signal, kept in proportion.
Kaggle public score
>1.85
Work by our team includes an ARC‑AGI‑3 competition submission with a Kaggle public score above 1.85.
This is a public-leaderboard result in an interactive-reasoning competition. Public values can change with submissions, infrastructure, rules, and organizer processes. It is not a final private score, placement, medal, award, peer-reviewed result, or evidence that artificial general intelligence has been achieved.
Sources verified August 18, 2026: Kaggle competition · ARC Prize context
04 · Limitations
Interactive success is not general reliability.
A benchmark samples a bounded set of environments and evaluators. A strategy that works there may not transfer to physical systems, scientific practice, open-ended software tasks, or human institutions. Scores can also hide brittle behavior, inefficient search, and failures on rare states.
Interpretation boundary. “Toward general AI” names a research direction involving exploration, memory, world modeling, goal acquisition, and planning. It is not a certification that general intelligence has been reached.