Individual neurons rarely encode single concepts

summary

In standard transformers, most neurons respond to multiple unrelated concepts (polysemanticity), and single-concept neurons are rare exceptions rather than the norm.

Multiple studies show that probing individual neurons yields mixed selectivity. However, recent work on larger models suggests some neurons do become more monosemantic at scale, creating a genuine dispute about whether this is a fundamental property or a scale-dependent phenomenon.

trust profile

dimensions

evid

72%

repl

55%

cons

48%

meth

71%

cred

82%

scop

45%

brdg

22%

cont

58%

derived scores

supp

60%

fron

44%

stab

50%

claim_support_vector v1.0 · 2026-03-09 19:58 UTC

evidence 3

↑ supporting 2

supports · artifact

Towards Monosemanticity (Anthropic, 2023)

Monosemanticity paper shows many neurons are polysemantic in standard-trained models.

Towards Monosemanticity (Anthropic, 2023) · 84% — Polysemanticity observations

supports · artifact

Toy Models of Superposition (Elhage et al., 2022)

Toy models paper confirms polysemanticity is the default in models with more features than dimensions.

↓ disputing 1

disputes · artifact

Scaling Monosemanticity (Anthropic, 2024)

The scaling paper shows that some neurons become more monosemantic at larger scales, partially disputing the claim that monosemanticity is always rare.

Scaling Monosemanticity (Anthropic, 2024) · 78% — Scale-dependent monosemanticity observed

scope

holds_in Small toy models (<10M parameters) model_scale

neighborhood 1

holds_in → Small toy models (<10M parameters) context

attestations

Curator (Human) verifies 0.75

Claim wording acceptable but note the contested status due to scale-dependence.

domains

Mechanistic Interpretability 100%

view status

Strict Empirical included

computation trace

show raw trace data

{
  "note": "Contested: scale-dependent monosemanticity partially disputes the claim",
  "inputs": {
    "artifacts": 3,
    "attestations": 1,
    "disputing_edges": 1,
    "supporting_edges": 2
  }
}