REXZEN · RESEARCH DIVISION

2πarc, the research division of Rexzen

Adaptive world models

What should a model do when conditions change? Learn a new correction, try something it learned before, or look for a different explanation? We test these ideas in small experiments.

The 2πarc character study.
World models, memory and evidence before lasting change.

THE QUESTION

Memory and adaptation

2πarc is Rexzen’s research division. We study how models predict what happens next and learn when conditions change.

When earlier conditions return, can a model use what it learned before to predict more accurately? We start with controlled simulations and compare the alternatives.

Expanding ripples on a real water surface.
RESEARCH PRACTICE

The Rex-WM Micro prototype

Write down the prediction before checking the result. Allow for noisy measurements. Compare methods fairly and explain what the test cannot tell us.

Inside the research directory
  1. PREDICTION

    Start with what happens.

    Compare a model’s prediction with the next observation. Study the difference.

  2. ADAPTATION

    Learn from the change.

    Study when a model should learn a correction and when earlier learning may help.

  3. EVIDENCE

    Compare the explanations.

    Consider noisy observations, changing conditions, recent-history inference and the cost of each method.

  4. REPRODUCIBILITY

    Keep the limits visible.

    Release code and inspected examples with their scope. Implementation checks are not independent scientific replications.

EOS · RESEARCH IN PROGRESS

Evidence-driven updates

Search Rexzen

15 pages found