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..
Bigger, Regularized, Categorical: High-Capacity Value Functions are Efficient Multi-Task Learners
Authors: Michal Nauman, Marek Cygan, Carmelo Sferrazza, Aviral Kumar, Pieter Abbeel
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We test our approach on 7 multi-task benchmarks with over 280 unique tasks, spanning high degree-of-freedom humanoid control and discrete vision-based RL. We find that, despite its simplicity, the proposed approach leads to state-of-the-art single and multi-task performance, as well as sample-efficient transfer to new tasks. |
| Researcher Affiliation | Academia | Michal Nauman1,2 Marek Cygan2,3 Carmelo Sferrazza1 Aviral Kumar4 Pieter Abbeel1,5 Pieter Abbeel holds concurrent appointments as a Professor at UC Berkeley and as an Amazon Scholar. This paper describes work performed at UC Berkeley and is not associated with Amazon. Marek Cygan was partially supported by National Science Centre, Poland, under the grant 2024/54/E/ST6/00388. We also gratefully acknowledge the Polish high-performance computing infrastructure, PLGrid (HPC Center: ACK Cyfronet AGH), for providing computational resources and support under grant no. PLG/2024/017817. |
| Pseudocode | No | The paper describes methods and architectures but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' block. |
| Open Source Code | Yes | We open-source our code under the following link: https://github.com/naumix/Bigger Regularized Categorical. |
| Open Datasets | Yes | Benchmarks. We consider a wide range of tasks, with a total of 283 diverse, complex control problems spanning five domains: Deep Mind Control (DMC) [101], Meta World (MW) [121], Humanoid Bench (HB) [93], Atari [10], and Shadow Hand (SH) [49]. |
| Dataset Splits | Yes | We list all the task sets considered in Appendix E. Transfer learning. In our transfer experiments, we evaluate three adaptation protocols inspired by previous work. ... The multi-task model is not trained on the transfer tasks, mimicking the train-test split used in supervised learning [14]. We report results for transfer experiments in Figures 2, 10 and 11. We list the tasks used in multi-task and transfer learning in Appendix E. |
| Hardware Specification | Yes | Hardware Information & Reproducibility All experiments were conducted on an NVIDIA A100 and H100 GPUs with 80GB of RAM and 16 CPU cores of AMD EPYC 7742 processor. |
| Software Dependencies | No | We would like to thank the Python [109], Num Py [42], Matplotlib [50], Sci Py [110] and JAX [16] communities for developing tools that supported this work. The paper lists software tools used but does not specify version numbers for these components, which is required for a reproducible description of ancillary software. |
| Experiment Setup | Yes | We discuss our experimental setting in Section 4 and Appendix D. We detail hyperparameters in Appendix F. F Hyperparameters We detail the hyperparameters used in our experiments in Tables 3 and 4 below. As discussed in Section 4, we use a single hyperparameter configuration across all tested tasks, showcasing robustness of our approach. |