Show HN: Experiments Mapping the "Primitive Layer" in Language Models

2 points

3 months ago

I spent months running experiments on what language models do when they encounter inputs outside their training distribution — random phonemes, invented morphemes, Wierzbicka's semantic primitives.

The finding that surprised me: Language model behavior follows a reproducible taxonomy (Synthesis, Collapse, Overflow, Metacognition, Linguistic Drift). It's not random noise — it's classifiable.

The finding that matters for interpretability: Structure is a more reliable control variable than content. Telling a model how to structure reasoning produces consistent outputs. Telling it what to reason about doesn't.

The finding that might matter most: Wierzbicka's semantic primitives (WANT, KNOW, FEEL, TIME, etc.) appear as measurable activation patterns in small language models across four different architectures — Qwen, Gemma, LLaMA, and SmolLM2.

18 experiments. 4 architectures. Cross-validated. Real data.

Full paper, experiment code, and primitive vocabulary JSON: https://github.com/dchisholm125/graph-oriented-generation

The primitive layer is waiting to be mapped.