Research map / Architectures and efficiency / Neural architecture search
LiteTransformerSearch: Training-free Neural Architecture Search for Efficient Language Models
Mojan Javaheripi, Gustavo Henrique de Rosa, Subhabrata Mukherjee, Shital Shah, Tomasz Lukasz Religa, Caio Cesar Teodoro Mendes, Sebastien Bubeck, Farinaz Koushanfar, Debadeepta Dey
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 |
|---|---|---|---|
| use decoder parameter count instead of partial training (500 steps) as ranking proxy | LTS NAS · 1200 Transformer-XL models, WT103 | SRC: 0.92 -> 0.98 | Improved |
| use decoder parameter count instead of partial training (5,000 steps) | LTS NAS · 1200 Transformer-XL models, WT103 | SRC: 0.96 -> 0.98 | Improved |
| use decoder parameter count instead of total parameter count as proxy | parameter-count proxy · Random architectures, WT103 | SRC | Improved |
| use decoder parameter count instead of total parameter count on homogeneous models | parameter-count proxy · 100 homogeneous Transformer-XL models | SRC: 0.95 | Improved |
| change adaptive input embedding factor k in {1,2,4} (ablation) | Transformer · WikiText-103 | validation perplexity | No clear effect |
| vary width-to-depth aspect ratio in [1,630] at fixed decoder params (ablation) | GPT-2 backbone · WikiText-103 | validation perplexity: <12.8% | No clear effect |
| vary aspect ratio in [8,323] at fixed decoder params (ablation) | Transformer-XL backbone · WikiText-103 | validation perplexity: <7% | No clear effect |
| vary topology at fixed decoder params (ablation) | GPT-2 backbone · WikiText-103 | latency / peak memory: 2.8x latency, 5.5x memory | Hurt |
| per-architecture learning rate sweep vs fixed 0.01 (ablation) | GPT-2, 100 models · WikiText-103 | SRC with decoder params: 0.96 -> 0.93 | No clear effect |
| layer-scaled baseline replaced by LTS heterogeneous search | Transformer-XL · LM1B on TITAN Xp | latency at same perplexity: 30.5 | Improved |
| LTS search vs layer scaling for GPT-2 | GPT-2 · LM1B on ARM CPU | latency | No clear effect |
| relu_log_det proxy on homogeneous models | low-cost proxy · 100 homogeneous models | SRC: 0.42 | Hurt |
Closest papers by meaning
- HW-GPT-Bench: Hardware-Aware Architecture Benchmark for Language Models (NeurIPS 2024 · similarity 0.88)
- Search for Efficient Large Language Models (NeurIPS 2024 · similarity 0.88)
- Memory-Augmenting Decoder-Only Language Models through Encoders (Student Abstract) (AAAI 2024 · similarity 0.87)
- Numerical Pruning for Efficient Autoregressive Models (AAAI 2025 · similarity 0.87)
- Trainable Transformer in Transformer (ICML 2024 · similarity 0.86)
- MatFormer: Nested Transformer for Elastic Inference (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.
