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..
Prompt Tuning Decision Transformers with Structured and Scalable Bandits
Authors: Finn Rietz, Oleg Smirnov, Sara Karimi, Lele Cao
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We now present a comprehensive evaluation of our method. We first assess performance across multiple Mu Jo Co (Todorov et al., 2012) tasks from standard multi-task benchmarks (Yu et al., 2020; Finn et al., 2017). We then analyze prompt-space exploration in a custom 2D environment. Finally, we evaluate the regret behavior of our bandit architecture and compare its scalability to standard algorithms. |
| Researcher Affiliation | Collaboration | Γrebro University EMAIL Oleg Smirnov King AI Labs, Microsoft Gaming EMAIL Sara Karimi King AI Labs, Microsoft Gaming EMAIL Lele Cao King AI Labs, Microsoft Gaming EMAIL |
| Pseudocode | Yes | The algorithmic pseudocode for prompt-tuning with our bandit is provided in the Algorithm 1 in Appendix A. The complete algorithmic pseudocode for prompt selection and update with our structured bandit architecture is provided in Algorithms 2 and 3 in Appendix A. |
| Open Source Code | Yes | Work performed while the author was an intern at King AI Labs (part of Microsoft Gaming). Code available at: https://github.com/king/pdt-bandits |
| Open Datasets | Yes | We first assess performance across multiple Mu Jo Co (Todorov et al., 2012) tasks from standard multi-task benchmarks (Yu et al., 2020; Finn et al., 2017). For the Mu Jo Co tasks, we use datasets and corresponding training and testing tasks from Xu et al. (2022). |
| Dataset Splits | Yes | To separate these 3 20 = 60 tasks into the training set T train and testing set T test, all tasks with an angles greater than 1.5 Ο (independently of the radius) are treated as testing task and are not part of the training set. This split yields 48 training tasks and 12 testing tasks. Spatially, the test set is indicated by the shaded area in Figure 3. (for Sparse 2D Point) There are 40 tasks in total, 35 are used for training, and 5 are used for testing. (for Mu Jo Co Half Cheetah) There are 50 tasks in total, 45 are used for training, and 5 are used for testing. (for Mu Jo Co Ant) The benchmark consists of 50 task variations with different placement goals; 5 are held out during training for evaluation. (for Meta-World Pick-Place) |
| Hardware Specification | Yes | All experiments were conducted on an instance equipped with an NVIDIA T4 GPU with 8GB of memory, utilizing the Py Torch library (Paszke et al., 2019). |
| Software Dependencies | No | The paper mentions 'utilizing the Py Torch library (Paszke et al., 2019)' but does not provide specific version numbers for PyTorch or other software dependencies. |
| Experiment Setup | Yes | Details of the hyperparameters for the DT and PDT models are provided in Table 3, while those for the bandit model are listed in Table 4. We used standard values for each hyperparameter instead of performing an expensive hyperparameter optimization. Environment-specific hyperparameters are reported separately in Table 5. |