Research map / Optimization / Optimizers
A Physics-Inspired Optimizer: Velocity Regularized Adam
Pranav Vaidhyanathan, Lucas Schorling, Natalia Ares, Michael A Osborne
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 AdamW with VRAdam (velocity-regularized learning rate) | Transformer · WikiText-2 | validation loss: 6.47 -> 5.99 | Improved |
| replace AdamW with VRAdam | CNN · CIFAR-10 | validation loss: 0.522 -> 0.476 | Improved |
| replace AdamW with VRAdam | GFlowNet · GridWorld flow matching | validation loss: 2.41 -> 1.25 | Improved |
| replace AdamW with VRAdam | GPT-2 124M · FineWebEdu-10B | validation loss: 3.511 -> 3.447 | Improved |
| replace AdamW with VRAdam | GPT-2 124M · FineWebEdu-10B | training time (s): 48549.56 -> 48522.40 | No clear effect |
| replace AdamW with VRAdam in 4-bit QLoRA | LLaMA-2-7B · OASST2 | perplexity: 3.84 -> 3.55 | Improved |
| replace AdamW with VRAdam in full fine-tuning | GPT-2 Large · GSM8K | perplexity: 4.12 -> 3.53 | Improved |
| replace AdamW with VRAdam in full fine-tuning | GPT-2 Large · GSM8K | exact-match accuracy: 35% -> 42% | Improved |
| use higher velocity-penalty power than 2 in the learning-rate gate (ablation) | VRAdam · CIFAR-10 test, 10 epochs | test loss: 0.932 (power 2) vs 1.155, 0.974, 1.004, 1.036 (powers 3-6) | Hurt |
| use SGD+Nesterov or RMSProp instead of VRAdam | Transformer · WikiText-2 | loss: NaN | Hurt |
Closest papers by meaning
- Surge Phenomenon in Optimal Learning Rate and Batch Size Scaling (NeurIPS 2024 · similarity 0.87)
- Adam with model exponential moving average is effective for nonconvex optimization (NeurIPS 2024 · similarity 0.87)
- A Comprehensive Framework for Analyzing the Convergence of Adam: Bridging the Gap with SGD (ICML 2025 · similarity 0.87)
- Enhancing Optimizer Stability: Momentum Adaptation of The NGN Step-size (NeurIPS 2025 · similarity 0.87)
- The Hidden Ratio in Adam: Stable Structure, Compression, and Sign Dynamics (NeurIPS 2026 · similarity 0.87)
- Scaling Exponents Across Parameterizations and Optimizers (ICML 2024 · similarity 0.87)
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.
