Research map / Architectures and efficiency / Attention and transformers
Softplus Attention with Re-weighting Boosts Length Extrapolation in Large Language Models
Bo Gao, Michael W. Spratling, Letizia Gionfrida
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 |
|---|---|---|---|
| replace Softmax with LSSA (Softplus, l1-norm, entropy scale) | GPT-2 small · FineWeb-10B val, 8K | validation loss: 6.2823 -> 5.9403 | Improved |
| add re-weighting p=15 to LSSA (ablation) | LSSA · FineWeb-10B val, 8K | validation loss: 5.9403 -> 3.3171 | Improved |
| add re-weighting p=15 to Softmax (ablation) | Softmax · FineWeb-10B val, 1K | validation loss: 3.1911 -> 5.3878 | Hurt |
| add re-weighting p=3 to Softmax (ablation) | Softmax · FineWeb-10B val, 8K | validation loss: 6.2823 -> 6.2339 | Improved |
| add re-weighting p=15 to Sigmoid (b=-log N, l1-norm) (ablation) | Sigmoid l1 · FineWeb-10B val, 8K | validation loss: 6.1846 -> 3.8587 | Improved |
| add l1-normalisation to Sigmoid (b=-log N) (ablation) | Sigmoid · FineWeb-10B val, 8K | validation loss: 9.5939 -> 6.1846 | Improved |
| use b=-log N instead of -log L in Sigmoid attention (ablation) | Sigmoid · FineWeb-10B val, 8K | validation loss: 14.4995 -> 9.5939 | Improved |
| use re-weighting p=3 instead of p=15 in LSSAR (ablation) | LSSAR · FineWeb-10B val, 8K | validation loss: 3.3171 -> 6.3007 | Hurt |
| replace Softmax with LSSAR (p=15) in passkey retrieval | GPT-2 small · passkey retrieval, 1K | accuracy: 64% -> 86% | Improved |
| replace Softmax with LSSAR (p=15) in passkey retrieval beyond 1.5K | GPT-2 small · passkey retrieval, 1.5K | accuracy: 0% -> 45% | Improved |
| replace Softmax with LSSAR (p=15) on ARC-E | GPT-2 small · ARC-E | normalised accuracy: 39.77 -> 40.57 | Improved |
| replace Softmax with LSSAR (p=15) on ARC-C | GPT-2 small · ARC-C | normalised accuracy: 23.72 -> 22.61 | Hurt |
Closest papers by meaning
- Unveiling and Harnessing Hidden Attention Sinks: Enhancing Large Language Models without Training through Attention Calibration (ICML 2024 · similarity 0.89)
- When Attention Sink Emerges in Language Models: An Empirical View (ICLR 2025 · similarity 0.89)
- Selective Attention: Enhancing Transformer through Principled Context Control (NeurIPS 2024 · similarity 0.89)
- Scaling Stick-Breaking Attention: An Efficient Implementation and In-depth Study (ICLR 2025 · similarity 0.88)
- From Attention to Activation: Unraveling the Enigmas of Large Language Models (ICLR 2025 · similarity 0.88)
- Beyond Linear Approximations: A Novel Pruning Approach for Attention Matrix (ICLR 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.
