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..
FairDICE: Fairness-Driven Offline Multi-Objective Reinforcement Learning
Authors: Woosung Kim, Jinho Lee, Jongmin Lee, Byung-Jun Lee
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Across multiple offline benchmarks, Fair DICE demonstrates strong fairness-aware performance compared to existing baselines. [...] 6 Empirical Behaviors of Fair DICE [...] 7 Welfare Maximization in Continuous Domains |
| Researcher Affiliation | Collaboration | 1Korea University 2Yonsei University 3Gauss Labs Inc. |
| Pseudocode | Yes | Algorithm 1 Fair DICE |
| Open Source Code | Yes | Our code is available at: https://github.com/ku-dmlab/Fair DICE.git. |
| Open Datasets | Yes | Environments We evaluate our method on the D4MORL benchmark [31], a standard MORL benchmark in continuous control domains. D4MORL builds upon the D4RL benchmark [32] by decomposing the original Mu Jo Co rewards into multiple objectives, such as speed, height and energy efficiency. |
| Dataset Splits | Yes | To simulate offline RL, we construct a dataset of 300 trajectories collected from a uniformly random behavior policy. [...] To simulate the offline RL setting, a dataset of 100 trajectories is collected using a behavior policy with an optimality level of 0.5. [...] The dataset for each domain consists of two types of data, collected using either expert or stochastically perturbed (amateur) behavioral policies, and is annotated with preference vectors. |
| Hardware Specification | Yes | All experiments were conducted on a single machine equipped with an Intel Xeon Gold 6330 CPU (256GB RAM) and an NVIDIA RTX 3090 GPU. |
| Software Dependencies | No | The Fair DICE algorithm, summarized in Algorithm 1, alternates between optimizing the dual variables (ν, µ) and updating the policy π via weighted behavior cloning. Both the policy π and critic ν networks are implemented as multilayer perceptrons, parameterized by ψ and θ, respectively. The scalar parameters µ are updated to maximize the desired social welfare function, and we fix α = 1 to correspond to the Nash social welfare objective. The initial state distribution p0 is estimated from the offline dataset. |
| Experiment Setup | Yes | Table 3 provides a summary of our default hyperparameters. The policy and value networks are constructed with three hidden layers, each containing 768 units. Optimization is performed using the Adam optimizer with a learning rate of 3 10 4 and a discount factor of γ = 0.99. |