MATH · IN · MODELS

OLMo

Allen Institute for AI

Structures found in this family (8)

Linear Direction, Linear Separability, Tree Metric Embedding, Linear Subspace, Intrinsic-dimension profile across depth, Dimensional collapse, Platonic Representation Hypothesis, Jacobian non-normality (Schur decomposition of per-layer operators)

By model (5)

OLMo-7B · 7B
Linear Direction, Linear Separability, Tree Metric Embedding, Linear Subspace, Dimensional collapse, Platonic Representation Hypothesis
OLMo-1B · 1B
Linear Direction, Intrinsic-dimension profile across depth, Dimensional collapse, Platonic Representation Hypothesis, Linear Subspace
OLMo 3 7B · 7B
Linear Subspace, Linear Direction, Linear Separability, Jacobian non-normality (Schur decomposition of per-layer operators)
OLMo-7B-0724-hf · 7B
Linear Direction
OLMo-7B-SFT · 7B
Linear Direction

Observations (13)

A PCA gender direction carries first-name discriminative signal
Diff-in-means reward-hacking directions forecast and causally control misalignment
A polar probe splits in-context relational graphs into existence and type readouts
DAS finds number- and unit-specific quantity-comparison subspaces
Independently-trained embedding models share linear structure (CKA 0.60-0.88)
Temporal-probe decodability predicts reasoning in high-resource languages
Editing MLP output-vector directions suppresses memorized sequences
A participation-ratio spectral signal finds attention circuits label-free across scale
Pretraining passes through three effective-rank phases
Refusal is mediated by a single direction across models
Cross-model and cross-modal representational alignment increases with scale and competence
Transformer Jacobians grow non-normal with depth, collapsing effective rank
Scale determines whether late-layer geometry stays organized by the readout

Papers

On the Mutual Influence of Gender and Occupation in LLM Representations (2025), Proxy Reward Internalization and Mechanistic Exploitation: A Learned Precursor to Reward Hacking and Its Generalization (2026), Polar probe linearly decodes semantic structures from LLMs (2026), Language Models Compare Quantities Using Number-specific and Unit-specific Heuristics (2026), Characterizing Linear Alignment Across Language Models (2026), What Really Controls Temporal Reasoning in LLMs: Tokenisation or Representation of Time? (2026), Output Vector Editing for Memorization Mitigation in Large Language Models (2026), Spectral Probe-Circuits: A Three-Step Recipe for Identifying Attention-Head Circuits in Pretrained Transformers (2026), Tracing the Representation Geometry of Language Models from Pretraining to Post-training (2025), Refusal in Language Models Is Mediated by a Single Direction (2024), Representation Engineering: A Top-Down Approach to AI Transparency (2023), Programming Refusal with Conditional Activation Steering (2024), Refusal Direction is Universal Across Safety-Aligned Languages (2025), The Platonic Representation Hypothesis (2024), Dynamics of the Transformer Residual Stream: Coupling Spectral Geometry to Network Topology (2026), Scale Determines Whether Language Models Organize Representation Geometry for Prediction (2026)