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..
On the Sample Complexity Bounds of Bilevel Reinforcement Learning
Authors: Mudit Gaur, Utsav Singh, Amrit Singh Bedi, Raghu Pasupathy, Vaneet Aggarwal
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We perform proof of concept experiments on Mujoco Tasks to demonstrate that the proposed first order algorithm works in practice. ... The training curves in Figure 1 illustrate the performance improvement of this approach against PEBBLE on both the Walker and Door Open tasks. |
| Researcher Affiliation | Academia | Mudit Gaur Purdue University Utsav Singh IIT Kanpur Amrit Singh Bedi University Of Central Florida Raghu Pasupathy Purdue University Vaneet Aggarwal Purdue University |
| Pseudocode | Yes | Algorithm 1 A first-order approach to bilevel RL |
| Open Source Code | Yes | The code can be found at https://github.com/Mudit Gaur/Neurips_2025_Bilevel_RL. |
| Open Datasets | Yes | We evaluate the effectiveness of this method, which solves the simplified objective, on two distinct environments: the Walker locomotion task from the Deep Mind Control Suite [37] and the Door Open manipulation task from Meta-world [27]. |
| Dataset Splits | No | The paper mentions a "fixed budget of human preference labels: 100 labels for the Walker task and 1,000 labels for the Door Open task." but does not specify any training, testing, or validation splits for these labels or any other data used in the experiments. |
| Hardware Specification | Yes | All experiments are conducted on a single machine with an NVIDIA RTX 1080 Ti GPU, and we report results averaged over multiple independent runs with different random seeds. |
| Software Dependencies | No | The paper does not provide specific version numbers for any software dependencies or libraries used in the implementation, such as Python, PyTorch, or TensorFlow versions. It mentions using 'PEBBLE [24]' as a baseline and 'B-Pref [23]' publicly released code, but these are frameworks/baselines, not specific versioned software dependencies for the authors' implementation. |
| Experiment Setup | No | The paper mentions "maintaining identical hyper-parameters and network architectures, such as the number of layers, learning rate, and the frequency of supervised reward learning." to the PEBBLE baseline, but does not provide the concrete values for these hyperparameters or other training configurations for their own method. |