Abstract
This paper maps 29 language models into a high-dimensional behavioral space using deterministic feature extraction. The main idea is that model families can be studied as positions and clusters, not just winners and losers on a scoreboard.
What it establishes
- Behavioral signatures can be represented geometrically.
- Model family relationships become visible through distance and clustering.
- Stability questions become easier to reason about when behavior is treated as trajectory as well as position.
Why it matters
The geography framing gives the research library a more durable language for model comparison. It makes room for movement, neighborhoods, and separation instead of collapsing everything into a single scalar rank.