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..
Online Two-Stage Submodular Maximization
Authors: Iasonas Nikolaou, Miltiadis Stouras, Stratis Ioannidis, Evimaria Terzi
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we evaluate the performance of our algorithm on both real and synthetic datasets. Additional details on each dataset, how it maps to instances of O2SSM, our choices of k, ℓ, and T, as well as detailed descriptions of competitor algorithms, can be found in Appendix F.7. ... Results. Figure 1 shows the average Ct over time t, for each algorithm and each dataset except Hotpot QA; the latter, along with additional results with varying k and ℓ, can be found in Appendix G. |
| Researcher Affiliation | Academia | Iasonas Nikolaou Boston University Miltiadis Stouras EPFL Stratis Ioannidis Northeastern University Evimaria Terzi Boston University |
| Pseudocode | Yes | Algorithm 1 Rounding-Augmented OCO (RAOCO) 1: Uses: 2: (1) FTRL or OGA OCO policy P e Xℓ(see Appendix H) 3: (2) Randomized Pipage Rounding Ξ : e Xℓ Xℓ(see Appendix I) 4: for t = 1 to T do 5: ext P e Xℓ e F1, . . . , e Ft 1 Compute fractional solution 6: xt Ξ(ext) Round to obtain the integral solution 7: Play action xt Accrue reward Ft(xt) (implicitly) 8: Construct relaxation e Ft from ft Function ft has been revealed 9: end for |
| Open Source Code | Yes | Our code is available online at https://github.com/jason Nikolaou/online-two-stage-sub-max. |
| Open Datasets | Yes | We conduct experiments on seven datasets; five real: Wikipedia, Images, Movie Rec, Influence, and Hotpot QA, and two synthetic: Team Formation and Coverage. The Wikipedia, Images, and Movie Rec datasets are drawn from prior work [Balkanski et al., 2016, Stan et al., 2017]. ... All datasets used in our experiments (Wikipedia, Images, Movie Rec, Influence, etc.) are drawn from prior published work or standard benchmarks (e.g., VOC2012, Movie Lens, Zachary s Karate Club). We cite the original papers for each dataset in Section 6 and Appendix F. Where applicable, datasets were used in accordance with their stated licenses or public terms of use. We do not use any proprietary, scraped, or restricted-access data. |
| Dataset Splits | No | For Wikipedia, Images, and Movie Rec, we construct a sequence of T objective functions by sampling T = 4 m functions uniformly at random from the m available functions. For the remaining datasets (Influence, Team Formation, and Coverage) we set T = m and use each of the m functions exactly once, in sequence. ... For the online algorithms, we measure the cumulative average reward Ct = 1 t Pt τ=1 Ft(xt). For each dataset, we report the average Ct and its standard deviation over five runs. |
| Hardware Specification | Yes | All the experiments were ran on an Apple M2 Macbook with 16GB RAM. |
| Software Dependencies | Yes | The code is written in Python 3.11.2. ... Each iteration of RAOCO-OGA requires solving one linear program, which we do via Gurobi (using a free academic license)... |
| Experiment Setup | Yes | For each algorithm and dataset, we tested different values for the learning rate η {0.0001, 0.001, 0.01, 0.1, 1, 10} and reported the best results. In Table 3, we report the η we used for each algorithm and dataset. |