Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

Transfer Learning for Benign Overfitting in High-Dimensional Linear Regression

Authors: Yeichan Kim, Ilmun Kim, Seyoung Park

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finite-sample experiments demonstrate the robustness of our methods to model and data heterogeneity, confirming their advantage. ... 5 Numerical Experiments We evaluate the finite-sample performance of our proposed TM and WTM estimates by comparing them to the target-only MNI as a baseline.
Researcher Affiliation Academia Yeichan Kim Yonsei University EMAIL Ilmun Kim KAIST EMAIL Seyoung Park Yonsei University EMAIL
Pseudocode Yes The entire procedure, from detecting informative sources to computing the WTM estimate, is outlined in Algorithm 1 (Appendix D), specifying D(0) in the set (9).
Open Source Code Yes Additionally, the authors release anonymized R sourcecodes at submission time that reproduce all figures in the paper to ensure the transparency and reproducibility of the empirical claims.
Open Datasets No We adopt the parametric configurations described below to simulate distribution shifts. Since only {(Z(q), ϵ(q))}Q q=0 in Assumption 1 are subject to randomness in our setup, all other parameters are generated with a fixed seed across the 50 simulations to ensure their deterministic nature; see Appendix F.1 for details.
Dataset Splits Yes First, partition the target dataset into K folds of equal size, each denoted by (X(0)[k], y(0)[k]) for k [K]; a common choice suggests K = 5 [17]. At each training step, we use the left-out folds (X(0)[ k], y(0)[ k]) := {(X(0)[k], y(0)[k])}K k=1 \ (X(0)[k], y(0)[k]) to train an estimate ˆβ[ k] and then evaluate the squared loss on the k-th fold given by ... For the WTM estimate, we set K = 5 and ε0 = 1/2 in Algorithm 1.
Hardware Specification Yes All numerical experiments are conducted by R using a standard laptop (ASUS Zen Book UX331UN, Intel(R) Core(TM) i7-8550U CPU @ 1.80GHz, 16 GB 2133 MHz DDR3).
Software Dependencies No All sourcecodes we release are scripted from scratch, relying on standard R packages such as MASS and mvtnorm.
Experiment Setup Yes Each setup is replicated over 50 independent simulations, and the average excess risk over the 50 simulations is plotted against each value of p {300, 400, . . . , 1000}. The mutually independent noise has i.i.d. mean-zero Gaussian components with common variance σ2 = 1. ... Let n0 = 25 and n1 = n2 = n3 = 75 for overfitting with S = 500. ... We tune the initial pre-training and fine-tuning learning rates for each experimental setup. ... for example, we use 0.001 for both pre-training and fine-tuning in Fig. 2.(a).