Research map / Architectures and efficiency / RNNs and normalization
NerVE: Nonlinear Eigenspectrum Dynamics in LLM Feed-Forward Networks
Nandan Kumar Jha, Brandon Reagen
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 |
|---|---|---|---|
| remove LayerNorm (GELU) (ablation) | GPT-2 125M GELU · GPT-2 125M (CodeParrot) | perplexity: 2.714 -> 3.223 | Hurt |
| use ReLU instead of GELU in norm-free model (ablation) | norm-free GPT-2 125M · GPT-2 125M (CodeParrot) | perplexity: 3.223 -> 2.988 | Improved |
| use learnable-slope Leaky ReLU instead of GELU in norm-free model (ablation) | norm-free GPT-2 125M · GPT-2 125M (CodeParrot) | perplexity: 3.223 -> 3.081 | Improved |
| use ReLU instead of GELU with LayerNorm (ablation) | GPT-2 125M · GPT-2 125M (CodeParrot) | perplexity: 2.714 -> 2.774 | Hurt |
| add spectral normalization to FFN in norm-free Leaky ReLU (ablation) | norm-free GPT-2 125M · GPT-2 125M (CodeParrot) | perplexity: 3.081 -> 3.000 | Improved |
| add weight normalization to FFN in norm-free Leaky ReLU (ablation) | norm-free GPT-2 125M · GPT-2 125M (CodeParrot) | perplexity: 3.081 -> 3.041 | Improved |
| add hyperspherical normalization to FFN in norm-free Leaky ReLU (ablation) | norm-free GPT-2 125M · GPT-2 125M (CodeParrot) | perplexity: 3.081 -> 3.122 | Hurt |
| use RoPE instead of NoPE (ablation) | GPT-2 OpenWebText · OpenWebText (26B tokens) | evaluation perplexity: 16.78 -> 15.20 | Improved |
| use PreLN instead of PostLN (ablation) | GPT-2 125M · GPT-2 125M (CodeParrot) | perplexity: 2.830 -> 2.714 | Improved |
| use PreLN instead of PostLN (ablation) | LLaMA-70M · C4 | perplexity: 33.6 -> 34.2 | Hurt |
| use PreLN instead of PostLN (ablation) | LLaMA-130M · C4 | perplexity: 26.7 -> 26.4 | Improved |
| use MixLN instead of PreLN (ablation) | LLaMA-250M · C4 | perplexity: 24.5 -> 24.2 | Improved |
Closest papers by meaning
- Dynamics of the Transformer Residual Stream: Coupling Spectral Geometry to Network Topology (NeurIPS 2026 · similarity 0.89)
- Tracing the Representation Geometry of Language Models from Pretraining to Post-training (NeurIPS 2025 · similarity 0.86)
- Building on Efficient Foundations: Effective Training of LLMs with Structured Feedforward Layers (NeurIPS 2024 · similarity 0.86)
- Recurrent Self-Attention Dynamics: An Energy-Agnostic Perspective from Jacobians (NeurIPS 2025 · similarity 0.85)
- Scaling Laws and Tradeoffs in Recurrent Networks of Expressive Neurons (NeurIPS 2026 · similarity 0.85)
- Scaling Law with Learning Rate Annealing (NeurIPS 2025 · similarity 0.85)
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.
