A Field Guide to Representation Hypotheses

We want to get a bird's eye view of different hypotheses on the structure of representations in neural networks. Do representations converge to a universal representation ? Or is there a plurality of end states to which they converge ? Is the claim for a subclass of models or shared across many model classes. The Platonic Representation Hypothesis claims a single universal representation, the Umwelt Representation Hypothesis argues for a plurality. Explore in two maps, eleven hypotheses about representations, 2015–2026. The first map plots the number of end-states vs scope, separating universalists from pluralists. A second map plots granularity (global to local) vs strength (described, empirical to proven) of claim. Select a point for the claim and source.

MAP 1 · ENDPOINTS vs SCOPE

What do representations converge to — and how many winners are there?

MAP 2 · GRANULARITY vs STRENGTH

Is it observed, exploited, or proven — and at what level of description?

READING THE AXES

An endpoint is a representational destination at the limit of training and scale — one attractor means all good systems end up in the same representation space; plurality means several stable spaces persist.

Map 1 · Endpoints vs Scope

Y — Number of endpoints
How many final representation spaces the hypothesis allows. Bottom: one global attractor (PRH, Strong PRH). Top: durable plurality (URH's ecological clusters, Superposition's per-model interference; CKA sits high because as a metric it registers dissimilarity as real structure).
X — Scope of claim
Who the hypothesis is about. Left: inside one network (Manifold, Superposition, LRH). Middle: across trained models (Convergent, Anna Karenina, Perfect PRH). Right: all representation-learners including brains (PRH, Universality, URH).
The corners
PRH bottom-right: everything converges to one. URH top-right: everything is constrained but splits into clusters.

Map 2 · Claim strength × Granularity

Y — Granularity
Level of description. Bottom: global geometry — distances, subspaces, linear maps (PRH, Manifold). Top: identifiable mechanisms — curve detectors, induction heads (Circuits; Superposition and Linear bridge partway up).
X — Strength of claim
Epistemic status. Left: empirical observation (Convergent 2015, CKA). Middle: constructive demonstration (vec2vec built a translator) or falsifiable test (Universality's fMRI predictions). Right: mathematical proof (Perfect PRH) or calibrated formal claim (Aristotelian).
The tension
The strongest claims (right edge) are all geometric. The 2026 critiques sit between the rows, arguing geometry-level agreement may only be the linear part.