Research map / Learning theory and algorithms / Combinatorial and search optimization
RoMA: Robust Model Adaptation for Offline Model-based Optimization
Sihyun Yu, Sungsoo Ahn, Le Song, Jinwoo Shin
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 |
|---|---|---|---|
| Add gradient-norm regularization only (drop consistency term) vs both terms (ablation) | RoMA · HopperController | 100th percentile score: 2183.4 -> 2466.5 | Improved |
| Add output-consistency term to gradient-norm term (ablation) | RoMA · HopperController | 100th percentile score: 2183.4 -> 2466.5 | Improved |
| Add gradient-norm term to consistency term (ablation) | RoMA · HopperController | 100th percentile score: 2081.9 -> 2466.5 | Improved |
| Replace naive pre-training with robust pre-training strategy (ablation) | RoMA · HopperController | 100th percentile score | Improved |
| Add model adaptation on top of robust pre-training (ablation) | RoMA · HopperController | 100th percentile score | Improved |
| Use larger iteration count T=500 vs 300 (ablation) | RoMA · all 6 tasks | 100th percentile score | No clear effect |
| Reduce alpha from alpha0 to 0.5*alpha0 (ablation) | RoMA · HopperController | 100th percentile score: 2466.5 -> 2420.5 | No clear effect |
| Reduce alpha to 0.1*alpha0 (ablation) | RoMA · HopperController | 100th percentile score: 2466.5 -> 2412.3 | No clear effect |
| Use smaller perturbation 0.2*eps0 (ablation) | RoMA · HopperController | 100th percentile score: 2466.5 -> 2303.6 | Hurt |
| Reduce both alpha (0.1x) and eps (0.2x) (ablation) | RoMA · HopperController | 100th percentile score: 2466.5 -> 2088.1 | Hurt |
| Use alpha=0.1*alpha0 instead of alpha0 (ablation) | RoMA · Molecule | 100th percentile score: 6.890 -> 6.941 | Improved |
| Use eps=0.2*eps0 instead of eps0 (ablation) | RoMA · Superconductor | 100th percentile score: 103.9 -> 106.3 | Improved |
Closest papers by meaning
- Robust Fine-Tuning from Non-Robust Pretrained Models: Mitigating Suboptimal Transfer With Epsilon-Scheduling (ICLR 2026 · similarity 0.86)
- Accelerating Adversarially Robust Model Selection for Deep Neural Networks via Racing (AAAI 2024 · similarity 0.84)
- Robustifying and Boosting Training-Free Neural Architecture Search (ICLR 2024 · similarity 0.84)
- Robust NAS under adversarial training: benchmark, theory, and beyond (ICLR 2024 · similarity 0.84)
- RoMa: Robust Dense Feature Matching (CVPR 2024 · similarity 0.84)
- Come Together, But Not Right Now: A Progressive Strategy to Boost Low-Rank Adaptation (ICML 2025 · 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.
