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..

Generating Informative Samples for Risk-Averse Fine-Tuning of Downstream Tasks

Authors: Heasung Kim, Taekyun Lee, Hyeji Kim, Gustavo De Veciana

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical Validation. We empirically validate our method in both synthetic and real-world settings. In a controlled regression task with highly imbalanced modes, the proposed method successfully synthesizes rare, high-loss samples that are critical for minimizing tail risk. In a real-world application wireless channel state information (CSI) compression our method consistently improves CVa R performance in the high ω regime, compared to existing robust and risk-minimization baselines.
Researcher Affiliation Academia Heasung Kim , Taekyun Lee, Hyeji Kim, and Gustavo de Veciana Department of Electrical and Computer Engineering The University of Texas at Austin Austin, TX 78712 EMAIL
Pseudocode Yes Algorithm 1 Risk-Averse Model Training via Loss-Guided Importance Sample Generation Input: Initial model ϑ0, generative model log pt(x), confidence level ω, function ς, dataset B Output: Risk-averse model ϑK+1
Open Source Code Yes Corresponding Author. Source code: https://github.com/Heasung-Kim/generating-informative-samples-for-risk-averse-fine-tuning-ofdownstream-tasks
Open Datasets Yes In this experiment, we assess the performance of the proposed method in the context of risk-averse CSI compression. We assume access to a pretrained score-based generative model trained on a CSI dataset generated by the Quadriga simulator (Jaeckel et al., 2021)
Dataset Splits No First, we obtain a pretrained (reference) model ϑ0 on 10^2 samples from p(x). Second, we use ς(x) = x1/2 to construct an importance-weighted distribution q(x) ς(ϖ(ϑ0; x)) p(x) and draw the same number of new samples from this distribution using the corresponding generative model. These samples are then used to train a risk-averse model. More detailed setup and results are provided in Appendix C.
Hardware Specification No No specific hardware details are provided in the main body of the paper.
Software Dependencies No No specific software dependencies with version numbers are provided in the main body of the paper.
Experiment Setup Yes First, we obtain a pretrained (reference) model ϑ0 on 10^2 samples from p(x). Second, we use ς(x) = x1/2 to construct an importance-weighted distribution q(x) ς(ϖ(ϑ0; x)) p(x) and draw the same number of new samples from this distribution using the corresponding generative model. These samples are then used to train a risk-averse model. More detailed setup and results are provided in Appendix C.