Research map / Large language models / Language, safety and interpretability
Late-to-Early Training: LET LLMs Learn Earlier, So Faster and Better
Ji Zhao, Shitong Shao, Yufei Gu, Xun Zhou, liang xiang, Zeke Xie
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 |
|---|---|---|---|
| Align early layers of target LLM to a small pretrained model's late layer (LET) vs standard training | 1.4B LLM · 9 downstream tasks, one-shot | average accuracy: 41.6 -> 43.6 | Improved |
| Align early layers to a small pretrained model (LET) vs standard training | 7B LLM · 9 downstream tasks, one-shot | average accuracy: 43.3 -> 45.5 | Improved |
| Use reverse knowledge distillation from small teacher (RKD) | LLM pretraining · 9 downstream tasks, one-shot | average accuracy: 41.6 -> 41.4 (1.4B); 43.3 -> 42.2 (7B) | Hurt |
| Use LET instead of SALT | LLM pretraining · 9 downstream tasks, one-shot | average accuracy: 42.9 -> 43.6 (1.4B); 44.7 -> 45.5 (7B) | Improved |
| Train LET for only 67% of steps | 1.4B LLM · 9 downstream tasks, one-shot | average accuracy: 41.6 -> 42.5 | Improved |
| Align middle layer of T instead of final layer (ablation) | LET · average downstream accuracy | average accuracy | Hurt |
| Align T's last layer to early layer of M (L2E) instead of middle or last layer of M (ablation) | LET · downstream accuracy and perplexity | average accuracy | Improved |
| Set lambda above 0.1 (ablation) | LET · average downstream accuracy | average accuracy | Hurt |
| Use lambda=0.01 instead of 0.1 (ablation) | LET · average downstream accuracy | average accuracy | Hurt |
| Replace cosine alignment loss with logsum loss (ablation) | LET · 9 downstream tasks | average accuracy: 43.6 -> 43.7 | Improved |
| Replace cosine alignment with CCA (ablation) | LET · 9 downstream tasks | average accuracy: 43.6 -> 42.7 | Hurt |
| Tokenizer mismatch between T and M (ablation) | LET · 9 downstream tasks | average accuracy: 43.6 -> 42.3 | Hurt |
Closest papers by meaning
- Unveiling the Secret Recipe: A Guide For Supervised Fine-Tuning Small LLMs (ICLR 2025 · similarity 0.92)
- Benchmarking Optimizers for Large Language Model Pretraining (NeurIPS 2026 · similarity 0.91)
- EE-LLM: Large-Scale Training and Inference of Early-Exit Large Language Models with 3D Parallelism (ICML 2024 · similarity 0.91)
- Sheared LLaMA: Accelerating Language Model Pre-training via Structured Pruning (ICLR 2024 · similarity 0.90)
- Fixing It in Post: A Comparative Study of LLM Post-Training Data Quality and Model Performance (NeurIPS 2025 · similarity 0.90)
- The Journey Matters: Average Parameter Count over Pre-training Unifies Sparse and Dense Scaling Laws (ICLR 2025 · similarity 0.90)
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.
