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..
Fair Continuous Resource Allocation with Equality of Impact
Authors: Blossom Metevier, Dennis Wei, Karthikeyan Natesan Ramamurthy, Philip S. Thomas
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Third, we extend this approach to noisy environments with a meta-algorithm and empirically demonstrate that our methods find fair allocations and perform competitively relative to representative baselines. We test our agent s performance in various deterministic and noisy settings. |
| Researcher Affiliation | Collaboration | Blossom Metevier University of Massachusetts Amherst, MA 01002 EMAIL; Dennis Wei IBM Research San Jose, CA 95120 EMAIL |
| Pseudocode | Yes | Algorithm 1: Non-Noisy Input; Algorithm 2: next Alloc; Algorithm 3: Noisy Input |
| Open Source Code | No | Answer: [No] Justification: Does not provide open access to code yet but will upon submission sufficient instructions to reproduce the main experimental results are located in the supplemental material. |
| Open Datasets | No | Environments We consider three instantiations of reward and impact functions, which we refer to as environments. In the imbalanced-rewards environment (IRE), we consider a scenario where one group s reward function consistently yields lower rewards at the same allocation than the other, making the reward-maximizing solution favor the higher-yielding group. ... Lastly, motivated by works that develop models showing the effects of water allocation under diminishing marginal returns [Klocke et al., 2006], we consider an environment in which reward could represent the crop yield per unit of land, and impact as the log-based formulation of the relative deprivation index. We term this the water allocation environment (WAE). |
| Dataset Splits | No | Environments We consider three instantiations of reward and impact functions, which we refer to as environments. ... Assume that instead of observing the true outcomes for the reward and impact functions, the agent receives noisy outcomes of the form ˆhi(x) = hi(x)+ϵhi and ˆri(x) = ri(x)+ϵri where i {A, B}, and each ϵhi and ϵri is drawn independently from a zero-mean normal distribution. |
| Hardware Specification | Yes | All experiments were conducted on a Mac Book Air (M1, 2020 model) featuring an Apple M1 chip (8-core CPU) and 8 GB of RAM, running mac OS Sequoia 15.1.1. No external GPUs were utilized. |
| Software Dependencies | Yes | The experiments were implemented in Python 3.10.9. Key libraries utilized include Num Py (version 1.22.4), Sci Py (version 1.7.3), Matplotlib (version 3.5.1), and scikit-learn (version 1.4.1.post1), Pymoo (version 0.6.1.3). |
| Experiment Setup | Yes | The pop_size and n_generations were tuned to ensure the total number of function evaluations (queries to reward and impact functions) matched the 150-query budget of our algorithm. ... The Gaussian Process Regressor for each function was configured with n_restarts_optimizer=10 and an alpha=1e-3 (representing the noise variance assumed by the GP model, which can also be interpreted as a Tikhonov regularization term). |