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..
Pareto Optimal Risk-Agnostic Distributional Bandits with Heavy-Tail Rewards
Authors: Kyungjae Lee, Dohyeong Kim, Taehyun Cho, Chaeyeon Kim, Yunkyung Ko, Seungyub Han, Seokhun Ju, Dohyeok Lee, Sungbin Lim
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 7 Numerical Experiments Setup. We test our methods in synthetic and real-world multi-risk bandit settings with heavy-tailed rewards. We consider two settings: (1) Real-world portfolio selection, using daily returns from the top 20 S&P 500 stocks; and (2) Synthetic 20-armed Pareto bandits. [...] Real-World Results. As shown in Figure 2, MR Dist LCB consistently outperforms all baselines across 3-, 6-, and 9-risk settings. [...] Synthetic Results. Figure 3 compares cumulative Pareto regret across 3-, 6-, and 9-risk configurations under different tail indices p. MR-Dist LCB consistently achieves the lowest cumulative regret in all settings, while MR Trunc and MR LCB exhibit higher regret and larger variance. |
| Researcher Affiliation | Collaboration | Kyungjae Lee1 Dohyeong Kim2 Taehyun Cho2 Chaeyeon Kim1 Yunkyung Ko1 Seungyub Han2 Seokhun Ju2 Dohyeok Lee2 Sungbin Lim1,3 [...] 1Department of Statistics, Korea University 2Department of Electrical and Computer Engineering, Seoul National University 3LG AI Research |
| Pseudocode | Yes | Detail algorithm is described in Appendix F. [...] F.1 Distributional LCB [...] Algorithm 1 Distributional LCB [...] F.2 Median of Empirical Quantiles with Bootstrap Resampling [...] Algorithm 2 Median of Empirical Quantiles with Bootstrap Resampling [...] F.3 Multi-Risk Distributional LCB [...] Algorithm 3 Multi-Risk Distributional LCB |
| Open Source Code | No | Answer: [No] Justification: While code is not yet released, we will make it available upon acceptance, and the information in Sec. 7 and Appendix E is sufficient to reproduce all key experiments. |
| Open Datasets | Yes | We use daily stock return data from the top 20 S&P 500 companies over 3,184 trading days (2012.05.18 2025.01.14). |
| Dataset Splits | No | We use daily stock return data from the top 20 S&P 500 companies over 3,184 trading days (2012.05.18 2025.01.14). At each time step, the algorithm selects one asset and observes its daily return as the reward. |
| Hardware Specification | No | Answer: [No] Justification: Since our experiments involve simple setups and the main contributions are theoretical, the computational demands were minimal and not a limiting factor. |
| Software Dependencies | No | The paper does not contain specific software dependencies with version numbers. |
| Experiment Setup | Yes | All experiments run for 10,000 steps and are repeated with 20 random seeds. [...] Tail indices estimated via Hill s method range from p = 1.5 to 4.69, and we fix p = 1.5 for all algorithms. In the synthetic case, we vary p {1.01, 1.2, 1.5} to study the effect of tail heaviness, with p = 1.2 used as the baseline heavy-tail configuration. [...] H.2 Synthetic Experiment Setup [...] SRM1: alphas = [0.2, 0.4, 0.6, 0.8], ranges = [0.25, 0.5, 0.75, 1.0] [...] CVaR0.9: alpha = [0.9] |