3D-Learning: Diffusion-Augmented Distributionally Robust Decision-Focused Learning

Jiaqi Wen, Lei Fan, Jianyi Yang

In many modern systems—from cloud computing to smart grids—we use a "Predict-then-Optimize" (PTO) pipeline. First, a Machine Learning (ML) model predicts the future (e.g., "How many people will use ChatGPT in the next hour?"), and then an optimization algorithm makes a decision based on that prediction (e.g., "How many GPUs should we turn on?").

The problem? ML models are notoriously brittle. If the testing data looks even slightly different from the training data (Out-of-Distribution or OOD shifts), the predictions fail, leading to expensive mistakes like server crashes or wasted energy.

A new paper from researchers at the University of Houston introduces 3D-Learning, a framework that uses the power of generative diffusion models to make these decision-making pipelines "distributionally robust."

TL;DR

3D-Learning combines Diffusion Models with Decision-Focused Learning to create AI systems that remain reliable even when data patterns shift unexpectedly. By using a diffusion model to simulate realistic "worst-case" scenarios during training, it achieves a 37-51% improvement in decision performance over traditional robustness techniques in cloud resource management tasks.

What Problem Does This Paper Address?

Standard ML models are usually trained to minimize average error (like Mean Squared Error). Decision-Focused Learning (DFL) improved on this by training models to minimize the cost of the resulting decision. However, even DFL struggles when the world changes—for example, if a cloud service suddenly sees a spike in "coding" tasks when it was trained on "conversational" tasks.

To fix this, researchers use Distributionally Robust Optimization (DRO), which trains the model to perform well even under the "worst-case" distribution within a certain range (an "ambiguity set").

The gap lies in how we define that range:

  1. $\phi$-divergence (e.g., KL-divergence): Too restrictive; it can't handle "support shifts" (new types of data the model hasn't seen before).
  2. Wasserstein Distance: Mathematically powerful but often computationally impossible for deep learning or leads to "overly conservative" models that perform poorly on average.

Methods

The authors propose 3D-Learning (Diffusion-Augmented Distributionally Robust Decision-Focused Learning). Instead of using a simple mathematical "ball" to define possible data shifts, they use a Diffusion Model.

Key Innovations:

  • Diffusion-based Ambiguity Set: They search for worst-case data distributions within the parameterized space of a diffusion model. Because diffusion models are excellent at capturing the "shape" of real data, the variations they generate are more realistic than random noise.
  • Score-Matching Constraint: To prevent the "adversary" from creating impossible data, they constrain the search using the diffusion model's score-matching loss. This ensures the worst-case scenarios are still "consistent" with real-world physics or logic.
  • The IMAX Algorithm: They developed an "Inner Maximization" algorithm that uses Proximal Policy Optimization (PPO)—a technique from Reinforcement Learning—to efficiently find the most damaging (but realistic) data shifts to train against.

Datasets

The researchers evaluated 3D-Learning on a high-stakes task: LLM Resource Provisioning.

  • Azure LLM Inference Traces: Real-world data from Microsoft Azure (2023 and 2024) containing token counts for "Code" and "Conversation" tasks.
  • OOD Scenarios: They trained the model on 2023 conversation data and tested it on 2024 data and code-related tasks to simulate real-world distribution shifts.

Key Insights and Ideas

  • Generative Models as Stress-Testers: The "aha moment" is using a generative model not just to create data, but to act as a sophisticated "adversary" that knows exactly how to break the decision model in ways that are actually likely to happen.
  • Balancing the Average and the Worst-Case: Traditional robust methods often make the model so "scared" of the worst case that it performs poorly on normal days. 3D-Learning finds a "sweet spot" by keeping the adversarial samples on the data manifold.
  • Support Shift Capability: Unlike KL-DRO, 3D-Learning can handle data that is fundamentally different from the training set because the diffusion model can generalize to new areas of the data space.

Key Results

In the LLM provisioning task, the goal was to minimize "Regret" (the gap between the AI's decision and the perfect theoretical decision).

  • Superior Robustness: 3D-Learning achieved an average regret of 0.1635, significantly lower than KL-DRO (0.2612) and Wasserstein-DRO (0.3345).
  • Outperforming Standard DFL: It reduced the average regret of standard Decision-Focused Learning by over 50%.
  • Noise Resilience: When tested on corrupted datasets (adding Gaussian or Perlin noise), 3D-Learning maintained a regret ratio of only 46.8% compared to the baseline, proving it is much harder to "confuse" with noisy inputs.

Limitations and Future Work

  • Computational Heavyweight: The main drawback is the training cost. 3D-Learning takes about 2.3x longer to train than standard DFL and requires significantly more GPU memory (6.8 GB vs. <35 MB) because it has to run a diffusion model during the training loop.
  • Inference is Fast: Fortunately, the authors note that this overhead is only during training. At "test time" (when the model is actually making decisions), it is just as fast as a standard model.
  • Future Directions: The authors suggest applying this diffusion-based robustness to other complex problems, such as power grid management or autonomous vehicle path planning, where the cost of a "worst-case" error is catastrophic.

Summary updated

Abstract

Predict-then-Optimize (PTO) pipelines are widely employed in computing and networked systems, where Machine Learning (ML) models are used to predict critical contextual information for downstream decision-making tasks such as cloud LLM serving, data center demand response, and edge workload scheduling. However, these ML predictors are often vulnerable to out-of-distribution (OOD) samples at test time, leading to significant decision performance degradation due to large prediction errors. To address the generalization challenges under OOD conditions, we present the framework of Distributionally Robust Decision-Focused Learning (DR-DFL), which trains ML models to optimize decision performance under the worst-case distribution. Instead of relying on classical Distributionally Robust Optimization (DRO) techniques, we propose Diffusion-Augmented Distributionally Robust Decision-Focused Learning (3D-Learning), which searches for the worst-case distribution within the parameterized space of a diffusion model. By leveraging the powerful distribution modeling capabilities of diffusion models, 3D-Learning identifies worst-case distributions that remain consistent with real data, achieving a favorable balance between average and worst-case scenarios. Empirical results on an LLM resource provisioning task demonstrate that 3D-Learning outperforms existing DRO and Data Augmentation methods in OOD generalization performance.
BibTeX
@misc{wen2026d,
  author = {Jiaqi Wen and Lei Fan and Jianyi Yang},
  title = {{3D-Learning: Diffusion-Augmented Distributionally Robust Decision-Focused Learning}},
  year = {2026},
  eprint = {2602.02943},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2602.02943},
}
Figure