Research map / Large language models / Language, safety and interpretability
Meet in the Middle: A New Pre-training Paradigm
Anh Tuan Nguyen, Nikos Karampatziakis, Weizhu Chen
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 |
|---|---|---|---|
| replace FIM with MIM pre-training (2.7B) | 2.7B code model · HumanEval | pass@1 (%): 28.5 -> 30.7 | Improved |
| replace FIM with MIM pre-training (2.7B) | 2.7B code model · MBPP | pass@1 (%): 38.2 -> 42.2 | Improved |
| replace FIM with MIM pre-training (1.3B) | 1.3B code model · MBPP | pass@1 (%): 25.9 -> 26.8 | Improved |
| replace FIM with MIM pre-training (2.7B) | 2.7B code model · HumanEval Infilling | pass@1 (%): 22.8 -> 26.3 | Improved |
| replace FIM with MIM pre-training (2.7B) | 2.7B code model · MBXP | pass@1 (%): 30.4 -> 35.7 | Improved |
| replace FIM with MIM pre-training (2.7B) | 2.7B code model · HumanEval-X C++ | pass@1 (%): 24.5 -> 27.4 | Improved |
| replace FIM with MIM pre-training (2.7B) | 2.7B code model · APPS Introductory | pass@1 (%): 12.4 -> 14.3 | Improved |
| replace FIM with MIM pre-training (2.7B) | 2.7B language model · OpenWebText | perplexity: 11.92 -> 9.54 | Improved |
| replace FIM with MIM pre-training (2.7B) | 2.7B language model · The Pile | perplexity: 10.92 -> 9.24 | Improved |
| remove bidirectional context (lambda 0.3 -> 0) (ablation) | MIM-2.7B · HumanEval Infilling | pass@1 (%): 26.3 -> 24.7 | Hurt |
| remove bidirectional context (lambda 0.3 -> 0) (ablation) | MIM-350M · MBXP | pass@1 (%): 16.4 -> 14.1 | Hurt |
| remove token-level agreement regularizer (ablation) | MIM-2.7B · HumanEval Infilling | pass@1 (%): 26.3 -> 23.2 | Hurt |
Closest papers by meaning
- Evaluation of LLMs on Syntax-Aware Code Fill-in-the-Middle Tasks (ICML 2024 · similarity 0.87)
- Not All Tokens Are What You Need for Pretraining (NeurIPS 2024 · similarity 0.86)
- Clean First, Align Later: Benchmarking Preference Data Cleaning for Reliable LLM Alignment (NeurIPS 2025 · similarity 0.86)
- Dataset Decomposition: Faster LLM Training with Variable Sequence Length Curriculum (NeurIPS 2024 · similarity 0.86)
- DataMan: Data Manager for Pre-training Large Language Models (ICLR 2025 · similarity 0.86)
- Align-Pro: A Principled Approach to Prompt Optimization for LLM Alignment (AAAI 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.
