Research map / Large language models / Language, safety and interpretability
LLM Pretraining with Continuous Concepts
Jihoon Tack, Jack Lanchantin, Jane Yu, Andrew Cohen, Ilia Kulikov, Janice Lan, Shibo Hao, Yuandong Tian, Jason E Weston, Xian Li
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 |
|---|---|---|---|
| select concepts by attribution score instead of activation value (ablation) | CoCoMix 69M · OpenWebText (20B tokens) | sample efficiency: 17.5% better sample efficiency | Improved |
| predict SAE concepts instead of predicting the full hidden state directly (ablation) | CoCoMix 69M · OpenWebText (20B tokens) | perplexity | Improved |
| add concept prediction loss alone (no insertion) (ablation) | CoCoMix 69M · OpenWebText (20B tokens) | perplexity | Improved |
| add concept insertion on top of concept prediction (ablation) | CoCoMix 69M · OpenWebText (20B tokens) | perplexity | Improved |
| add concept insertion alone without prediction (ablation) | CoCoMix 69M · OpenWebText (20B tokens) | perplexity | No clear effect |
| insert concept vector as separate token instead of adding it to the hidden state (intervention) (ablation) | CoCoMix 69M · OpenWebText (20B tokens) | perplexity | Improved |
| use continuous concepts instead of learnable pause tokens (ablation) | CoCoMix 69M · OpenWebText (20B tokens) | perplexity | Improved |
| add CoCoMix concept prediction and mixing to next-token prediction | 1.38B model, 200B tokens · OpenWebText validation perplexity | tokens to reach equal perplexity: 21.5% fewer training tokens | Improved |
| use CoCoMix instead of knowledge distillation with a weak 124M teacher | 386M model · Downstream average perplexity (20B tokens) | perplexity: 23.4 (KD) -> 21.4 | Improved |
| use knowledge distillation instead of next-token prediction | 386M model · Downstream average perplexity (20B tokens) | perplexity: 24.2 -> 23.4 | No clear effect |
| add CoCoMix to next-token prediction | 69M model · Downstream average accuracy (20B tokens) | accuracy (%): 42.7 -> 42.9 | No clear effect |
| add CoCoMix to next-token prediction | 1.38B model · Downstream average accuracy (20B tokens) | accuracy (%): 48.7 -> 49.7 | Improved |
Closest papers by meaning
- Beyond Next Token Prediction: Patch-Level Training for Large Language Models (ICLR 2025 · similarity 0.88)
- Token Assorted: Mixing Latent and Text Tokens for Improved Language Model Reasoning (ICML 2025 · similarity 0.87)
- Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization (ICLR 2025 · similarity 0.87)
- Better & Faster Large Language Models via Multi-token Prediction (ICML 2024 · similarity 0.87)
- On the Generalization Ability of Next-Token-Prediction Pretraining (ICML 2025 · similarity 0.86)
- Not All Tokens Are What You Need for Pretraining (NeurIPS 2024 · 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.
