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..

Improved Scaling Laws in Linear Regression via Data Reuse

Authors: Licong Lin, Jingfeng Wu, Peter L Bartlett

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We also perform simulations to validate our theoretical findings. Namely, we train M-dimensional sketched linear predictors (1) via one-pass SGD and multi-pass SGD following the setup in Section 2 and 3, and analyze how their excess test errors scale with the number of samples N and the model size M.
Researcher Affiliation Collaboration Licong Lin UC Berkeley EMAIL Jingfeng Wu UC Berkeley EMAIL Peter L. Bartlett UC Berkeley and Google Deep Mind EMAIL
Pseudocode No The paper describes the multi-pass stochastic gradient descent (multi-pass SGD) and gradient descent (GD) processes using mathematical equations and prose (e.g., "vt : vt 1 γt fvt 1pxitq yit vfvt 1pxitq"), but no explicitly labeled "Pseudocode" or "Algorithm" block is present.
Open Source Code No Answer: [No] Justification: we do not release code and data for this paper at this time.
Open Datasets No In each simulation, we generate N i.i.d. samples pxi, yiq N i 1 from a linear model yi xxi, w y ϵi, where w P Rd is an unknown parameter vector and ϵi Np0, σ2q are i.i.d. Gaussian noise.
Dataset Splits No In each simulation, we generate N i.i.d. samples pxi, yiq N i 1 from a linear model yi xxi, w y ϵi, where w P Rd is an unknown parameter vector and ϵi Np0, σ2q are i.i.d. Gaussian noise. The covariates xi are drawn from Np0, Hq, and the true parameter vector w is sampled from a Gaussian prior Np0, Hwq... (This describes data generation, not explicit dataset splits like train/test percentages or counts for a pre-existing dataset).
Hardware Specification No Answer: [No] Justification: the simulations in this paper do not involve any large language models and can be reproduced on a personal computer.
Software Dependencies No The paper does not explicitly mention specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions, or specific libraries). It describes the experimental setup and parameters but not the software stack used.
Experiment Setup Yes In all experiments, we set d 10000, σ2 1 and pa, bq p2, 1.5q. (a), (b), (d): M 1000; (c): N 105. (d) γ 0.5, N 1000