Research map / Architectures and efficiency / RNNs and normalization
Stabilizing Native Low-Rank LLM Pretraining
Paul Janson, Edouard Oyallon, Eugene Belilovsky
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 spectral renormalization on top of orthogonalization (ablation) | Spectron · FineWeb val (Factorized Llama-94M) | perplexity: 20.95 -> 20.26 | Improved |
| add orthogonalization on top of spectral renormalization (ablation) | Spectron · FineWeb val (Factorized Llama-94M) | perplexity: 24.01 -> 20.26 | Improved |
| add either spectral control component to naive factorized training (ablation) | naive low-rank training · FineWeb val (Factorized Llama-94M) | perplexity: 1042.02 -> 24.01 | Improved |
| use rank ratio 0.4 instead of 0.25 (ablation) | Spectron · FineWeb val (Factorized Llama-94M) | perplexity: 20.26 -> 20.00 | No clear effect |
| reduce rank ratio to 0.125 (ablation) | Spectron · FineWeb val (Factorized Llama-94M) | perplexity: 20.26 -> 42.65 | Hurt |
| use Spectron instead of self-guided training with dense auxiliary weights | Factorized Llama · FineWeb val (Factorized Llama-454M) | perplexity: 13.70 -> 12.11 | Improved |
| use Spectron instead of naive AdamW | Factorized Llama · FineWeb val (Factorized Llama-94M) | perplexity: 26.43 -> 21.86 | Improved |
| train factorized model with Spectron at higher rank ratio (ablation) | Factorized Llama-94M · FineWeb val loss (rank ratio 0.4) | validation loss: 3.070 (dense) -> 3.028 | Improved |
| train Spectron factorized model at rank ratio 0.25 versus dense (ablation) | Factorized Llama-94M · FineWeb val loss (rank ratio 0.25) | validation loss: 3.070 (dense) -> 3.084 | No clear effect |
| combine Spectron with CoLA architecture | CoLA · FineWeb val (Factorized Llama-94M) | perplexity: 24.21 -> 19.95 | Improved |
| use Spectron instead of LoRO | low-rank training · FineWeb val (Factorized Llama-94M) | perplexity: 24.56 -> 21.86 | Improved |
| use Muon with self-guided training instead of AdamW | Self-guided training · FineWeb val (Factorized Llama-94M) | validation loss | Hurt |
Closest papers by meaning
- Fira: Can We Achieve Full-rank Training of LLMs Under Low-rank Constraint? (NeurIPS 2025 · similarity 0.91)
- SLTrain: a sparse plus low rank approach for parameter and memory efficient pretraining (NeurIPS 2024 · similarity 0.90)
- From Low Rank Gradient Subspace Stabilization to Low-Rank Weights: Observations, Theories, and Applications (ICML 2025 · similarity 0.90)
- Building on Efficient Foundations: Effective Training of LLMs with Structured Feedforward Layers (NeurIPS 2024 · similarity 0.90)
- AdaRankGrad: Adaptive Gradient Rank and Moments for Memory-Efficient LLMs Training and Fine-Tuning (ICLR 2025 · similarity 0.89)
- Dynamic Low-Rank Sparse Adaptation for Large Language Models (ICLR 2025 · similarity 0.89)
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.
