Research map / Large language models / Language, safety and interpretability
Large language models implicitly learn to straighten neural sentence trajectories to construct a predictive representation of natural language.
Eghbal A. Hosseini, Evelina Fedorenko
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 |
|---|---|---|---|
| train the model (trained vs untrained GPT2-XL) (ablation) | GPT2-XL · UDsubset8408 | curvature across layers | Improved |
| increase model size (82M to 1,558M parameters) | GPT2 family · UDsubset8408 | curvature reduction | Improved |
| increase training corpus from 1M to 100M words | GPT2 (12 layers) · UDsubset8408 | curvature reduction | Improved |
| increase training corpus from 100M to 1B words | GPT2 (12 layers) · UDsubset8408 | curvature reduction | No clear effect |
| use model-generated continuation instead of ground-truth sentence | GPT2-XL · UDsubset8408 10-token sentences | curvature drop | Improved |
| use beam search or sampling instead of greedy generation (ablation) | GPT2-XL · UDsubset8408 subset (n=1573) | curvature drop | No clear effect |
| switch corpus to OpenWebText (ablation) | GPT2-XL · OpenWebText (8408 sentences) | curvature across layers: ~120 degrees | No clear effect |
| use other architectures (RWKV, pythia-1.4b, OPT-1.3B) (ablation) | trained models · UDsubset8408 | curvature vs surprisal | No clear effect |
| use unigram frequency control instead of 3-gram or PCFG surprisal (ablation) | GPT2-XL · UDsubset8408 middle layers | curvature-surprisal relationship | Hurt |
Closest papers by meaning
- Lines of Thought in Large Language Models (ICLR 2025 · similarity 0.88)
- Flexible Language Modeling in Continuous Space with Transformer-based Autoregressive Flows (NeurIPS 2025 · similarity 0.87)
- Arrows of Time for Large Language Models (ICML 2024 · similarity 0.87)
- Language Models Are Implicitly Continuous (ICLR 2025 · similarity 0.87)
- Language Model Behavioral Phases are Consistent Across Architecture, Training Data, and Scale (NeurIPS 2025 · similarity 0.86)
- Emergence of a High-Dimensional Abstraction Phase in Language Transformers (ICLR 2025 · similarity 0.86)
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.
