Research map / Large language models / Language, safety and interpretability
The Lattice Representation Hypothesis and Concept Algebra in LLMs
Bo Xiong
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| estimate attribute directions with regularized LDA instead of class-mean difference or random (ablation) | LLM embeddings (LLaMA3.1-8B, Gemma-7B, Mistral-7B) · WordNet domains formal context recovery | F1: Gemma WN-Animal 50.1 (Mean) -> 83.2 | Improved |
| use projection-based soft inclusion with LDA directions for subsumption (ablation) | LLM embeddings · WordNet partial order inference | F1: LLaMA WN-Animal 66.7 (Mean) -> 77.1 | Improved |
| use lattice concept algebra (meet/join) instead of mean-embedding similarity (ablation) | LLM embeddings · WordNet meet/join ranking | MRR | Improved |
| move from physical to abstract domains | lattice geometry · WordNet Event and Cognition vs Animal/Plant/Food | F1: LLaMA formal context F1 82.5 (Animal) -> 71.5 (Event) | Hurt |
| scale LLaMA-3 from 3B to 70B (ablation) | lattice geometry · WordNet domains | Macro-F1 | Improved |
| Linear versus Mean on Mistral WN-Plant partial order (ablation) | soft inclusion · WN-Plant partial order | F1: Mean 60.5 vs Linear 57.1 (Mistral) | Mixed |
Closest papers by meaning
- From Tokens to Lattices: Emergent Lattice Structures in Language Models (ICLR 2025 · similarity 0.91)
- The Geometry of Categorical and Hierarchical Concepts in Large Language Models (ICLR 2025 · similarity 0.89)
- The Linear Representation Hypothesis and the Geometry of Large Language Models (ICML 2024 · similarity 0.87)
- Latent Concept Disentanglement in Transformer-based Language Models (ICLR 2026 · similarity 0.86)
- On the Origins of Linear Representations in Large Language Models (ICML 2024 · similarity 0.85)
- Geometric Latent Reasoning Induces Shorter Generations in LLMs (NeurIPS 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
