BART

Meta AI (Facebook AI Research)

Structures found in this family (3)

Linear Direction, Linear Subspace, Platonic Representation Hypothesis

Hypotheses argued for by this family (1)

Platonic Representation Hypothesis

By model (2)

BART-base · 139M

Linear Direction, Linear Subspace, Platonic Representation Hypothesis

MolGen

Linear Direction

Observations (5)

KnowledgeEditor gates the fine-tuning gradient with a hypernetwork outer product

Splicing probed entity-state representations causally shifts generation

Transformer-Patcher corrects mistakes by adding one key-value neuron per patch

Othello-move representations from seven architecturally distinct real models (GPT-2, BART, T5, Flan-T5-XL, LLaMA-2-7B, Mistral-7B, Qwen2.5-7B) Procrustes-align into a shared geometric space at cosine similarity up to 97.2% (unsupervised), with PCA trajectories converging across models and nearest-neighbor tile embeddings recovering the board's actual physical adjacency structure

SLIM SAE features linearly correlate with molecular properties and steer editing

Papers

Editing Factual Knowledge in Language Models (2021), Implicit Representations of Meaning in Neural Language Models (2021), Transformer-Patcher: One Mistake Worth One Neuron (2023), Revisiting the Othello World Model Hypothesis (2025), SLIM: Sparse Latent Steering for Interpretable and Property-Directed LLM-Based Molecular Editing (2026)