Research map / Optimization / Optimizers
Mechanic: A Learning Rate Tuner
Ashok Cutkosky, Aaron Defazio, Harsh Mehta
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 |
|---|---|---|---|
| pre-train with Mechanic instead of tuned AdamW | BERT-B · MLM | MLM accuracy: 71.5 -> 71.7 | Improved |
| pre-train with Mechanic instead of tuned Lion | BERT-B · MLM | MLM accuracy: 71.8 -> 72.0 | Improved |
| pre-train with Mechanic instead of tuned AdamW | BERT-L · MLM | MLM accuracy: 75.4 -> 75.3 | No clear effect |
| pre-train with Mechanic instead of tuned Lion | BERT-L · MLM | MLM accuracy: 75.7 -> 75.5 | No clear effect |
| pre-train with Mechanic instead of tuned Lion, then finetune with Lion | BERT-B · GLUE QNLI | accuracy: 86.8 -> 89.0 | Improved |
| pre-train with Mechanic instead of tuned optimizer on JFT-300M | ViT-B/16 with Lion · JFT-300M | pre-training accuracy: 47.0 -> 49.6 | Improved |
| pre-train with Mechanic instead of tuned optimizer on JFT-300M | ViT-L/16 with Lion · JFT-300M | pre-training accuracy: 52.0 -> 55.4 | Improved |
| pre-train with Mechanic instead of tuned AdamW on JFT-300M | ViT-L/16 · JFT-300M | pre-training accuracy: 54.4 -> 54.4 | No clear effect |
| use Mechanic on small dataset from scratch | ResNet-18 Momentum · CIFAR-100 | accuracy: 77.6 -> 75.3 | Hurt |
| raise weight-decay-like term lambda from 0.01 to 0.1 (ablation) | Mechanic ResNet-18 · CIFAR-100 | accuracy: 75.3 -> 76.6 | Improved |
| raise lambda from 0.01 to 0.1 (ablation) | Mechanic ResNet-18 · CIFAR-10 | accuracy: 94.1 -> 95.3 | Improved |
| use Mechanic instead of tuned Momentum | WRN-40-10 · CIFAR-100 | accuracy: 79.9 -> 79.6 | No clear effect |
Closest papers by meaning
- The Surprising Agreement Between Convex Optimization Theory and Learning-Rate Scheduling for Large Model Training (ICML 2025 · similarity 0.86)
- Investigation into Training Dynamics of Learned Optimizers (Student Abstract) (AAAI 2024 · similarity 0.85)
- The Road Less Scheduled (NeurIPS 2024 · similarity 0.85)
- Stepping on the Edge: Curvature Aware Learning Rate Tuners (NeurIPS 2024 · similarity 0.85)
- Prodigy: An Expeditiously Adaptive Parameter-Free Learner (ICML 2024 · similarity 0.84)
- Surge Phenomenon in Optimal Learning Rate and Batch Size Scaling (NeurIPS 2024 · similarity 0.84)
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.
