Black-Box Knowledge Transfer across Distinct Feature Sets

Foundation models are often used as fixed black-box predictors for downstream tasks with limited labeled data, but their predictions may be biased and unsafe to trust blindly. We study this setting through black-box assisted nonparametric regression: a learner observes labeled samples and can query a fixed predictor $f_0$, while the target $f^*$ is close to $f_0$ in $L_2(P_X)$ up to an unknown radius $δ$. We give a finite-sample minimax characterization showing a phase transition at $δ_c(n) \asymp n^{-β/(2β+d)}$, with leading risk $\min\{δ^2, n^{-2β/(2β+d)}\}$. We then analyze a Safe Residual Estimator: it learns a correction around $f_0$, initializes the residual head at zero so the initial predictor equals $f_0$, and uses holdout selection to revert to $f_0$ when the learned correction is not supported by validation data. Here, "safe" means avoiding negative transfer, i.e., performing worse than the black-box predictor alone. The estimator matches the leading minimax term up to an additive validation-selection cost. Synthetic regression experiments verify the predicted phase transition, while CIFAR-100 with CLIP and AG News with Qwen3-8B provide practice-facing evidence that the same residual-correction tradeoff is useful beyond the formal squared-loss regression setting.

Modern machine learning pipelines increasingly rely on reusing pretrained and foundation models across downstream tasks. These pretrained models can differ not only in performance but also in how they can be used: some only provide black-box predictions, while others may permit white-box access to internal representations that can be probed or fine-tuned. When deployed to the target domain in the presence of distribution shift, no single strategy, including zero-shot application, fine-tuning, or directly training a target-specific model, is uniformly the best. In this work, we propose Frontier Learning, a framework that treats a library of candidate models spanning different training histories and access regimes as complementary sources of information rather than mutually exclusive alternatives. Frontier Learning constructs a unified target-domain feature by concatenating internal representations from white-box candidates as well as prediction outputs from black-box candidates, then fits a lightweight, regularized supervised learner on this concatenated representation using labeled target data. Because the resulting hypothesis class contains predictors obtained by zero-shot reuse, fine-tuning, and direct training as special cases, empirical risk minimization over the frontier learner is guaranteed to be no worse, on the training sample, than any individual baseline. We evaluate the framework in simulations spanning varying degrees of source-target compatibility and in two real-world distribution-shift settings: visual domain adaptation on DomainNet/VisDA and clinical mortality prediction across intensive care unit domains using MIMIC-IV-Notes. Across all settings, Frontier Learning matches or outperforms the strongest individual reuse strategy, with the largest gains arising precisely when no single baseline is reliable across the range of shift considered.

Existing methods for transfer learning struggle to deal with situations where the source datasets are limited and not guaranteed to be well-aligned with the target dataset. A typical strategy is to use the empirical loss minimizer on the source data as a prior mean for the target parameters. Our key conceptual contribution is to use a risk minimizer conditional on source parameters instead. This allows us to construct a single joint prior distribution for all parameters from the source datasets as well as the target dataset. As a consequence, we benefit from full Bayesian uncertainty quantification and can perform model averaging via Gibbs sampling over indicator variables governing the inclusion of each source dataset. We show how a particular instantiation of our prior leads to a Bayesian Lasso in a transformed coordinate system and discuss computational techniques to scale our approach to moderately sized datasets. We discuss connections between the Maximum a Posteriori estimate associated with our approach and the recently proposed Trans-Lasso method and demonstrate that the MAP estimator MSE-dominates the Trans-Lasso in the normal means setting when there is no regularization on the source datasets. Finally, we perform numerical experiments finding that full Bayesian inference provides superior predictive performance relative to Trans-Lasso on a genetics application, especially when the source data are limited.