Research map / Architectures and efficiency / RNNs and normalization
Disentangling Geometry, Performance, and Training in Language Models
Atharva Kulkarni, Jacob Mitchell Springer, Arjun Subramonian, Swabha Swayamdipta
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 |
|---|---|---|---|
| higher effective rank of the unembedding matrix | OLMo-style models (4M-75M) · Pile-10k ID loss | loss | Mixed |
| use effective rank instead of ID loss to predict OOD loss | OLMo-style models · Paloma Dolma-100 programming subset | OOD loss | No clear effect |
| rank collapse in small models under post-training quantization (GPTQ) | OLMo-4M, OLMo-28M · post-quantization ID loss | loss | Hurt |
| train OLMo-14M matching Pythia-14M config (sequential instead of parallel attention) (ablation) | OLMo-14M · late-stage pretraining saturation | ID loss | Improved |
| reduce token budget to 8B and batch size to 32 for OLMo-14M (ablation) | OLMo-14M · saturation | ID loss | No clear effect |
| increase batch size (ablation) | OLMo-style pretraining · Pile ID loss | effective rank and loss | Mixed |
| increase weight decay (0.01 to 0.5) (ablation) | OLMo-style pretraining · Pile ID loss | effective rank and loss | Mixed |
| scale learning rate by 0.1, 10, 100 (ablation) | OLMo-style pretraining · Pile ID loss | loss | Mixed |
| anneal learning rate to 10% instead of 0-1% of peak (ablation) | OLMo-style pretraining · Pile ID loss | effective rank and loss | Hurt |
| train on more tokens (8B to 128B) (ablation) | OLMo-style pretraining · Pile ID loss | loss and effective rank | Improved |
| fine-tune models with varied initial effective rank on StarCoder-Python | OLMo-style models · fine-tuning and forgetting on Pile-10k | loss | No clear effect |
| control for training hyperparameters (partial Spearman) when relating geometry to loss (ablation) | geometric metrics · Pile-10k ID loss | Spearman correlation: -0.699 -> -0.130 (effective rank) | Hurt |
Closest papers by meaning
- Tracing the Representation Geometry of Language Models from Pretraining to Post-training (NeurIPS 2025 · similarity 0.89)
- Emergence of a High-Dimensional Abstraction Phase in Language Transformers (ICLR 2025 · similarity 0.89)
- Diff-eRank: A Novel Rank-Based Metric for Evaluating Large Language Models (NeurIPS 2024 · similarity 0.88)
- Less is More: Local Intrinsic Dimensions of Contextual Language Models (NeurIPS 2025 · similarity 0.88)
- Understanding Emergent Abilities of Language Models from the Loss Perspective (NeurIPS 2024 · similarity 0.88)
- The Curse of Depth in Large Language Models (NeurIPS 2025 · similarity 0.88)
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.
