GPT-OSS

OpenAI

Structures found in this family (6)

Linear Direction, Linear Separability, Polytope (Simplex), Linear Subspace, Dimensional collapse, Anisotropy

By model (1)

GPT-OSS-20B · 20B

Linear Direction, Linear Separability, Polytope (Simplex), Linear Subspace, Dimensional collapse, Anisotropy

Observations (10)

Persona-vector taxonomy sorts traits as natural, steerable, or intractable

Probes decode a grid cognitive map that reasoning then reorganizes

Concept manifolds transiently untangle right before a reasoning step

Fourier periodicity in number embeddings is universal; linear separability is not

Language identity is linearly separable from layer 1; alignment tracks pretraining mix

Temporal-probe decodability predicts reasoning in high-resource languages

Massive activations provably cause compression valleys synced with attention sinks

Insecure-code SAE directions sit closer to toxic directions than secure-code ones

Sycophantic agreement, genuine agreement, and praise occupy distinct steerable directions

MoE routing reflects hidden-state geometry, not domain expertise

Papers

What Models Express, Suppress, and Resist: Auditing Open-Weight LLMs with Persona Vectors (2026), A Behavioural and Representational Evaluation of Goal-Directedness in Language Model Agents (2026), Emergent Manifold Separability during Reasoning in Large Language Models (2026), Convergent Evolution: How Different Language Models Learn Similar Number Representations (2026), How Language Directions Align with Token Geometry in Multilingual LLMs (2026), What Really Controls Temporal Reasoning in LLMs: Tokenisation or Representation of Time? (2026), Attention Sinks and Compression Valleys in LLMs Are Two Sides of the Same Coin (2025), Understanding Emergent Misalignment via Feature Superposition Geometry (2026), Sycophancy Is Not One Thing: Causal Separation of Sycophantic Behaviors in LLMs (2025), The Myth of Expert Specialization in MoEs: Why Routing Reflects Geometry, Not Necessarily Domain Expertise (2026)