Research map / Learning theory and algorithms / Prediction and decision losses
Black-Box Assisted Regression: Phase Transitions and Minimax Optimality
Yan Zhou
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 holdout safe selection (fallback to black-box) to the residual estimator (ablation) | Residual estimator · 1D synthetic regression (n=20, delta=0.0) | MSE: 0.0335 -> 0.0027 | Improved |
| add holdout safe selection to the residual estimator in high dimensions with a poor prior (ablation) | Residual estimator · High-dimensional regression (d=20, n=100, delta=1.0) | MSE: 1.026 -> 1.007 | Improved |
| learn a nonlinear residual instead of a weighted ensemble of black-box and scratch (ablation) | Safe Residual vs WGT · 1D synthetic regression (n=20, delta=0.8) | MSE: 0.1452 -> 0.0971 | Improved |
| learn a residual around the black-box instead of training from scratch (ablation) | Safe Residual vs Scratch · 1D synthetic regression (n=50, delta=0.0) | MSE: 0.0325 -> 0.0006 | Improved |
| use residual learning when the black-box is badly biased | Residual estimator · 1D synthetic regression (n=20, delta=1.5) | MSE: 2.2496 -> 0.1461 | Improved |
| use zero-initialized safe residual instead of validation-tuned weighted ensemble | CLIP on CIFAR-100 · CIFAR-100 (n=2000 labeled) | accuracy (%): 63.9 -> 66.7 | Improved |
| use zero-initialized safe residual instead of validation-tuned weighted ensemble | CLIP on CIFAR-100 · CIFAR-100 (n=200 labeled) | accuracy (%): 54.5 -> 60.4 | Improved |
| use safe residual instead of weighted ensemble in ultra-low-sample NLP regime | Qwen3-8B on AG News · AG News (n=64 labeled) | accuracy (%): 54.7 -> 74.7 | Improved |
| use safe residual instead of weighted ensemble with more labeled data | Qwen3-8B on AG News · AG News (n=256 labeled) | accuracy (%): 75.6 -> 77.0 | Improved |
| concatenate black-box output to features instead of residual learning | CLIP on CIFAR-100 · CIFAR-100 (n=100 labeled) | accuracy (%): 0.4 collapse versus 59.8 for safe residual | Hurt |
| add zero-initialization of residual head (ablation) | Safe Residual neural head · CIFAR-100 and AG News | accuracy / early-training variance | Improved |
| vary validation fraction rho from 0.15 to 0.3 (ablation) | Safe Residual · CIFAR-100 | accuracy | No clear effect |
Closest papers by meaning
- On Volume Minimization in Conformal Regression (ICML 2025 · similarity 0.85)
- Posterior Inference with Diffusion Models for High-dimensional Black-box Optimization (ICML 2025 · similarity 0.85)
- Posterior Inference in Latent Space for Scalable Constrained Black-box Optimization (NeurIPS 2026 · similarity 0.85)
- Optimal transport-based conformal prediction (ICML 2025 · similarity 0.85)
- Learning with Synthetic Data via SGD in High-Dimensional Linear Regression (NeurIPS 2026 · similarity 0.85)
- Batch and match: black-box variational inference with a score-based divergence (ICML 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.
