Reproduce “Practical Adversarial Multivalid Conformal Prediction” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Osbert Bastani, Varun Gupta, Christopher Jung, Georgy Noarov, Ramya Ramalingam, Aaron Roth
The MVP algorithm enables online conformal prediction with strong, fine-grained empirical coverage guarantees across arbitrary, intersecting groups and thresholds, operating robustly in adversarial and non-exchangeable data settings. It achieves nearly statistically optimal convergence rates of O(1/√n) for group-specific coverage while maintaining computational efficiency.
Key insights
- A practical, computationally lightweight algorithm can achieve robust threshold-calibrated and group-conditional coverage guarantees even against an adaptive adversary without exchangeability assumptions.
- The proposed method provides nearly statistically optimal convergence rates of O(1/√n) for fine-grained coverage, significantly improving upon previous theoretical bounds.
- Enforcing a mild smoothness assumption on the distribution of conformal scores allows for strong theoretical guarantees without compromising the practical applicability of the approach.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.