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..
Diffusion Models Meet Contextual Bandits
Authors: Imad Aouali
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate d TS using both synthetic and Movie Lens problems. In our experiments, we run 50 random simulations and plot the average regret with standard error. Our main contribution is to demonstrate that pretraining a diffusion model offline enables the construction of expressive and informative priors that substantially improve exploration efficiency in contextual bandits. We first evaluate d TS in a setting where the prior matches the true generative process (Section 5.1 to isolate the benefit of informative priors), and then consider a misspecified regime (Section 5.2 and Appendix G) where the prior is either trained on out-of-distribution data or intentionally perturbed. These experiments show that even when the prior is imperfect, d TS maintains strong performance: highlighting its robustness and practical relevance. Code can be found in this Git Hub repository. |
| Researcher Affiliation | Collaboration | Imad Aouali Criteo AI Lab CREST, ENSAE, IP Paris EMAIL |
| Pseudocode | Yes | Algorithm 1 d TS: diffusion Thompson Sampling |
| Open Source Code | Yes | Code can be found in this Git Hub repository. |
| Open Datasets | Yes | We evaluate d TS using both synthetic and Movie Lens problems. Movie Lens Dataset. http://grouplens.org/datasets/movielens/, 2016. CIFAR-10 [40] MNIST [48] |
| Dataset Splits | Yes | Table 3: Regret improvement (%) of d TS on CIFAR-10. Offline Data (%) vs. Hier TS vs. Lin TS 1% 69.11% 87.74% 5% 79.56% 92.18% 25% 80.65% 92.48% 50% 81.67% 92.88% |
| Hardware Specification | Yes | Our experiments were conducted on internal machines with 30 CPUs and thus they required a moderate amount of computation. |
| Software Dependencies | No | We used JAX for diffusion model pre-training, summarized as follows: Optimization: Adam optimizer with a 10 3 learning rate was used. |
| Experiment Setup | Yes | We vary d {5, 20}, L {2, 4}, K {102, 104}, and set the horizon to n = 5000, considering both linear and non-linear models. Non-linear diffusion. We consider Eq. (1) where fℓare 2-layer neural networks with random weights in [ 1, 1], Re LU activation, and hidden layers of size h = 20 for d = 5, and h = 60 for d = 20. Optimization: Adam optimizer with a 10 3 learning rate was used. The NN was trained for 20,000 epochs with a batch size of min(2048, pre-training sample size). |