Research map / Architectures and efficiency / RNNs and normalization
When Does Sparsity Mitigate the Curse of Depth in LLMs
Dilxat Muhtar, Xinyuan Song, Sebastian Pokutta, Max Zimmer, Nico Pelleriti, Thomas Hofmann, Shiwei Liu
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 |
|---|---|---|---|
| add moderate weight decay (0 to 0.1) (ablation) | 1.2B Pre-LN LLM · validation perplexity | perplexity: 15.63 -> 14.83 | Improved |
| raise weight decay to 3.0 (ablation) | 1.2B Pre-LN LLM · validation perplexity | perplexity: 14.83 -> 773.42 | Hurt |
| increase training sequence length from 256 to 2048 (ablation) | 1.2B Pre-LN LLM · validation perplexity | perplexity: 18.51 -> 14.51 | Improved |
| increase training sequence length to 8192 (ablation) | 1.2B Pre-LN LLM · validation perplexity | perplexity: 14.52 -> 16.30 | Hurt |
| use multi-query attention (G=16) instead of MHA (ablation) | 1.2B Pre-LN LLM · validation perplexity | perplexity: 14.52 -> 14.47 | Improved |
| replace dense FFN with fine-grained MoE (1B active) (ablation) | Dense-1B · validation perplexity | perplexity: 14.52 -> 13.82 | Improved |
| replace dense FFN with fine-grained MoE (400M active) (ablation) | Dense-400M · validation perplexity | perplexity: 17.56 -> 15.89 | Improved |
| naively increase depth from L=16 to L=32 at constant parameters | 1.2B Pre-LN LLM · ARC-C and HellaSwag average | accuracy and usefulness score: usefulness 0.75 -> 0.53; accuracy drops by 0.5 | Hurt |
| add GQA (G=2) and MoE on top of depth L=32 | L=32 1.2B model · ARC-C and HellaSwag average | average accuracy: to 44.1 (usefulness 0.75) | Improved |
| add weight decay 1.0 on L=32 model (ablation) | L=32 1.2B model · ARC-C and HellaSwag average | accuracy and usefulness | Hurt |
| scaled initialization instead of normal initialization (ablation) | L=16 model, T=1024 · loss | loss: 2.70 -> 2.71 | No clear effect |
| add longer sequence length T=4096 on top of scaled initialization (ablation) | L=16 model · loss | loss: 2.71 -> 2.64 | Improved |
Closest papers by meaning
- The Curse of Depth in Large Language Models (NeurIPS 2025 · similarity 0.93)
- Determining Layer-wise Sparsity for Large Language Models Through a Theoretical Perspective (ICML 2025 · similarity 0.91)
- Adaptive Layer Sparsity for Large Language Models via Activation Correlation Assessment (NeurIPS 2024 · similarity 0.91)
- Do Language Models Use Their Depth Efficiently? (NeurIPS 2025 · similarity 0.90)
- Sparsing Law: Towards Large Language Models with Greater Activation Sparsity (ICML 2025 · similarity 0.90)
- Outlier Weighed Layerwise Sparsity (OWL): A Missing Secret Sauce for Pruning LLMs to High Sparsity (ICML 2024 · 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.
