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..
PABBO: Preferential Amortized Black-Box Optimization
Authors: Xinyu Zhang, Daolang Huang, Samuel Kaski, Julien Martinelli
ICLR 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | On a benchmark composed of synthetic and real-world datasets, our method is several orders of magnitude faster than the usual Gaussian process-based strategies and often outperforms them in accuracy. ... We evaluate PABBO on synthetic functions, described in Section 5.1, and against real-world examples from the BO literature: hyperparameter optimization (Section 5.2) and human preferences datasets (Section 5.3). Subsequently, Section 5.4 presents a series of ablation studies... |
| Researcher Affiliation | Academia | Xinyu Zhang , Daolang Huang , Samuel Kaski Department of Computer Science Aalto University EMAIL Julien Martinelli Inserm Bordeaux Population Health Vaccine Research Institute Universit e de Bordeaux Inria Bordeaux Sud-ouest |
| Pseudocode | Yes | Algorithm 1 Preferential Amortized Black-box Optimization (PABBO) ... Algorithm S1 PABBO test time algorithm |
| Open Source Code | Yes | The code to reproduce our experiments is available at https://github.com/ xinyuzc/PABBO. |
| Open Datasets | Yes | The HPO-B benchmark contains multiple search spaces... (Pineda Arango et al., 2021)... The Sushi dataset collects a preference score... Available at https://www.kamishima.net/sushi/ |
| Dataset Splits | Yes | All the meta-datasets within a search space have been categorized into three splits: meta-train, meta-validation, and meta-test. ... Each meta-dataset is equally divided beforehand into two parts from which we sample either prediction set D(p) or query set D(q), so as to prevent any information leak from rewards. |
| Hardware Specification | Yes | We train our model using up to 5 Tesla V100-SXM2-32GB GPUs... We evaluate all the models on 2x64 core AMD EPYC 7713 @2.0 GHz. |
| Software Dependencies | No | PABBO is implemented using Py Torch (Paszke et al., 2019). ... For q EUBO, q EI, q NEI, we used the implementation from the Bo Torch library (Balandat et al., 2020). |
| Experiment Setup | Yes | Hyperparameter settings can be found in Appendix D.1. For q EUBO, q EI, q NEI, we used the implementation from the Bo Torch library (Balandat et al., 2020). |