Emerging research · in development
AI models change. How do we know what changed?
Genome is a Cosen Lab research initiative that aims to make a model’s behavioral evolution understandable, beyond its architecture and version number.
A model is more
than its parameters.
Models are evaluated, adapted, fine-tuned and exposed to changing contexts. Their observable behavior can evolve even when the version label says very little.
- 01
Model
Architecture and parameters define the technical artifact.
- 02
Context
Conditions of use shape what is observed.
- 03
Change
Adaptation and fine-tuning can alter its behavior.
- 04
The question
What changed, and where did the change come from?
Giving models
a behavioral identity.
Not a personality, nor a humanized profile: a research representation of what is observed, comparable across versions and conditions.
Behavior
What remains recognizable over time.
Variation
How what we observe shifts as the context changes.
Drift
When behavior moves away from a reference.
Lineage
Which versions and derivations a model descends from.
A version number tells you which model you have. It doesn’t tell you how its behavior got there.
A model should have a history.
Not just a version number.
Select a version: the label barely changes, while the behavioral identity can shift significantly. Genome works to make that difference visible and tie it to the model’s history.
What the label says
v1.0
Baseline model
A version number tells you which model you have. It doesn’t tell you how its behavior has changed.
What Genome aims to make visible
Behavioral identity · compared with the reference
ReferenceAbstract representation · not real data or methods
Lineage · conceptual diagram
Versions are nodes. Derivations are relationships.
Change is more useful when its origin stays visible.
Four questions.
One evolving model.
Those who adopt, govern or certify AI systems need answers that are hard to obtain today.
- What changed?Understand behavioral variations from one model state to the next.
- Where did it come from?Preserve the lineage that links versions and derivations.
- Does it still behave as expected?Compare what is observed today against an earlier reference.
- Can we explain how it evolved?Connect change, evaluation and lineage into a coherent history.
Genome is
under construction.
This page describes a research direction, not a finished system or validated performance. Methods and results will be shared as the work matures.
- 01
Concept
Defined
- 02
Active research
Underway
- 03
Experimentation
Next phase
- 04
Pilot partnerships
By invitation
Confidential details
To protect work in progress, methods and architecture are not public. We are glad to discuss them with research partners and interested organizations under a non-disclosure agreement.
Models evolve. Their capabilities change. Their behavior can drift. A version number tells you where a model stands. Genome explores how to understand the way it got there.
Understand the model. Understand its evolution.
Behavioral identity for AI that changes. Emerging research from Cosen Lab.