Research

How concepts form inside a model.

Eigan's work starts from three papers on concept geometry in transformer language models: where a concept is assembled across the model's depth, how its direction moves while that happens, and whether the structures found are causal rather than descriptive. They are the instruments our products are built on. Code and data for all three are open.

Foundations

The starting toolkit

Published in 2026 by James R. A. Henry as independent research (the Rosetta Program). Please cite the current versions below: the arXiv v2 revisions correct and replace v1.

Foundation · arXiv:2605.24856 · v2, Aug 2026

The Concept Allocation Zone: Tracking How Concepts Form Across Transformer Depth

James Henry · Independent Researcher

The standard way to read a concept out of a model is to probe at the single layer where it separates best. This paper treats concept formation as a process instead: a concept becomes separable across one or more contiguous regions of the residual stream, its Concept Allocation Zones. It defines three layer-wise metrics (Separation, Concept Coherence and Concept Velocity) and a scored detector that finds the zones automatically.

Evaluated across 35 models, 8 architectural families and 7 concepts.
A concept's separation curve is frequently multimodal: it is assembled in more than one place.
Scored detection surfaces faint “gentle” allocation zones that standard peak detection misses.

Abstract PDF

Cite (BibTeX)

@misc{henry2026caz,
  title         = {The Concept Allocation Zone: Tracking How Concepts
                   Form Across Transformer Depth},
  author        = {Henry, James},
  year          = {2026},
  eprint        = {2605.24856},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2605.24856}
}

Foundation · arXiv:2605.25848 · v2, Aug 2026

Geometric Evolution Maps: Extracting Stable Concept Probes from Transformer Residual Streams

James Henry · Independent Researcher

A concept probe is only as reliable as the layer it is taken from. Inside an allocation zone, a concept's direction rotates substantially before it settles. A Geometric Evolution Map records that trajectory, takes the settled direction at the end of the zone, and evaluates it at the handoff layer just beyond. A concept usually spans several zones; the set of its maps is its atlas, and the atlas is what the method delivers for downstream use.

Rotation during assembly is substantial and consistent across 29 base models and 17 concept types.
The extracted direction is causal: ablating it suppresses far more separation than ablating a random direction.
No single site carries the signal alone. Probing after the rotation settles is more precise, and a depth-matched control shows that advantage comes from depth rather than the handoff boundary itself.

Abstract PDF

Cite (BibTeX)

@misc{henry2026gem,
  title         = {Geometric Evolution Maps: Extracting Stable Concept
                   Probes from Transformer Residual Streams},
  author        = {Henry, James},
  year          = {2026},
  eprint        = {2605.25848},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2605.25848}
}

Foundation · doi:10.5281/zenodo.22016917 · v1, Aug 2026

Concept Encoding Strategies Across 28 Transformers: A Concept Allocation Zone and Geometric Evolution Map Evaluation

James Henry · Independent Researcher

The validation study for both companion papers, testing CAZ and GEM unchanged across 28 base models from 8 families and 17 concepts. Its central result is that the structures they detect are causal, not descriptive artifacts: ablating the identified concept directions suppresses the concept in the residual stream, and activation patching at the same sites restores it downstream.

The effect is direction-specific: random-direction controls at the same layers suppress almost nothing.
It holds across architectures and across the full range of detection confidence, including the faint zones a fixed threshold would miss.
A behavioral pilot finds the same direction-specific effect on next-token predictions in 27 of 28 models.
No single fixed layer, short of the model's last, reliably captures the causal structure. A concept is a trajectory across depth, not a point.

Paper (Zenodo)

Cite (BibTeX)

@misc{henry2026cazvalidation,
  title     = {Concept Encoding Strategies Across 28 Transformers:
               A Concept Allocation Zone and Geometric Evolution
               Map Evaluation},
  author    = {Henry, James},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.22016917},
  url       = {https://doi.org/10.5281/zenodo.22016917}
}

How they fit

Where, which direction, and does it matter

01 · CAZ

Where a concept forms

Find the regions of depth where the model organises its geometry to make a concept separable, including the faint ones.

02 · GEM

Which direction to read

Follow the concept's direction as it rotates through each region and take it once it settles, instead of at an arbitrary layer.

03 · Validation

Whether it is causal

Remove the direction and the concept goes with it; restore it and the concept comes back. Random directions do neither.

Code & data

Open, so the results can be checked

rosetta_tools Python library implementing CAZ detection, GEM extraction, ablation and patching. GitHub DOI

Rosetta_Analysis Analysis, figure and reproduction scripts for the papers. GitHub DOI

Rosetta Concept Pairs Contrastive text pairs for each concept, generated by 14 models from four AI labs. GitHub DOI

Rosetta Activations Pre-extracted residual-stream activations and per-model analysis outputs. Hugging Face DOI