Inria and UC Berkeley's MVCS Learns Tighter AI Uncertainty Regions by Minimizing Volume
A new optimization framework learns the tightest possible multivariate prediction regions by jointly training the model, shape, and size while preserving conformal coverage guarantees.
ProTakeaways
- New method learns minimum-volume multivariate conformal prediction sets by treating volume as an explicit training loss
- Prediction regions are p-norm balls with learnable shape parameter p, center, and transformation matrix
- Jointly optimizes the predictor, the norm structure, and a covariate-dependent shape function for local adaptivity
- Beats naive quantile regression and covariance-based baselines on 16 of 18 real-dataset experiments while preserving coverage
- Robust to outliers up to fraction alpha because loss uses order statistics rather than mean residuals
- Code released at github.com/ElSacho/MVCS; paper at arxiv.org/abs/2503.19068
Conformal prediction has become the default tool for wrapping a machine learning model's output with rigorous confidence guarantees. When the response is vector-valued, current approaches force a painful tradeoff: either assume a rigid shape like an ellipsoid, or use flexible density-based methods that scale badly and often produce bloated regions. A new paper from researchers at Inria Paris and UC Berkeley proposes a different route by treating the shape of the prediction set as something to be learned, with volume itself as the loss function.
The method, called Minimum Volume Conformal Sets (MVCS), comes from Sacha Braun, Liviu Aolaritei, Michael I. Jordan, and Francis Bach. Code is available in the official GitHub repo.
Why multivariate conformal prediction is awkward
For a scalar output, quantile regression with conformal calibration gives tight, adaptive intervals. Extending this to a vector-valued response is where things break. A straightforward extension that builds Cartesian products of marginal intervals fails to account for dependencies across dimensions, resulting in overly conservative and inefficient prediction sets.
The existing alternatives each have a catch:
- Ellipsoidal sets shaped by empirical residual dependencies incorporate covariance information but assume elliptical symmetry, limiting their adaptability to more complex distributions.
- Copula-based methods and density estimators relax geometric constraints but often require accurate joint distribution estimation, which is computationally expensive in high dimensions and typically produces high estimation variance.