Pre-collected offline data can supercharge RL—but what if the environment shifts and your transition dynamics no longer match? Thrilled to share our NeurIPS 2025 Spotlight paper: “Composite Flow Matching for RL with Shifted Dynamics Data”. Paper: https://t.co/G5VOfZK6Sq Code:

Wonderful Collaboration with @HaichuanWang23 @tonghanwang Guojun Xiong, @MilindTambe_AI #neurips2025 #GenerativeAI #reinforcementlearning #DeepLearning #DecisionMaking

Beyond benchmarks, we apply this to patrol planning for wildlife conservation, showing how principled generative modeling helps real decisions.

Results on MuJoCo: • Best on 19/27 tasks • Ties for best on 5 more • Up to +149.8% over pure online RL under shifted dynamics Composite Flow Matching = big sample-efficiency gains without blindly trusting mismatched data.

This Composite Flow Matching setup: 1. encodes prior knowledge and help learn a better online dynamics 2. tracks a Wasserstein “gap” between old vs new 3. selectively transfers only useful offline information.

Our idea: let generative models measure & bridge the dynamics gap. We first train a flow model on offline data to learn an informative base distribution that captures rich structure in the old dynamics. Then we train a second flow on online data, coupled via optimal transport.

8/ Thanks to all my amazing collaborators: @HaichuanWang23, @CEmogor, Vincent Borsch-Supan, @lilyxu0, @MilindTambe_AI

7/ 🌍✨ As Generative AI becomes increasingly powerful, we hope it also supports efforts to protect our planet. Wildflow is a step toward that vision, offering a new perspective on how generative AI can be applied to real-world challenges in wildlife conservation.🐘🌿

6/ 🔎 Case study: Where does Wildflow shine? In cells with higher neighboring patrol effort—areas where displacement effects & spatial dynamics matter most. Flow matching captures these patterns better.

5/ 📊 Performance: Tested on two national parks in Uganda: Murchison Falls National Park and Queen Elizabeth National Park. Wildflow consistently outperforms baselines.

4/ 💡 Why latent space? Observed poaching ≠ true poaching. Wildflow models the hidden occupancy state and learns how patrols, terrain, and spatial context affect risk—even when detection is imperfect.

3/ 🌀 Enter Wildflow: A generative model that learns to predict latent poaching activity by: 1. Integrating flow matching with ecological occupancy detection models 2. Using composite flow to inject domain knowledge

2/ 🌍 The stakes: Rangers patrol huge protected areas with limited resources. But real poaching data is: 1. Sparse 2. Noisy (imperfect detection) 3. Spatially complex Forecasting risk is critical—but hard.

🧵1/ 🚨 Poaching is a major threat to biodiversity. Can we predict where it might strike next? Meet Wildflow — our new latent flow matching framework for poaching prediction in wildlife conservation. 🐘🌱 📝 Preprint: https://t.co/nab4AicEBo

Excited to present our paper at #UAI2025 on Thursday 🇧🇷 Can we predict which datasets transfer best when downstream decisions, not just predictions, matter? Can we know this before learning? We introduce OTD³, a new dataset distance for PtO tasks https://t.co/bkCVsoMIQu

Thanks for all my wonderful collaborators: Haichuan Wang, @pyrojewel419, Cheol Woo Kim, Mingxiao Song, Alayna Nguyen,@TonghanW , @haifengxu0 and @MilindTambe_AI

Empirically, we find that our game-theoretic approach yields patrol strategies with significantly lower regret than baseline methods, resulting in more effective wildlife protection in national parks.

To solve the game with diffusion models, we use the double oracle framework. This introduces challenges—constrained strategy space and sampling from unnormalized distributions. We address them via a strategy of strategies and a twisted SMC sampler for precise utility estimation.

To account for imperfections in the learned diffusion model, we formulate a robust patrol optimization problem as a two-player zero-sum game between a ranger defender and a nature adversary that selects the worst-case poaching activity within a constrained space.

We use a conditional diffusion model to forecast the number of snares in each grid cell of the national park, capturing complex high-dimensional distributions. On poaching data from Murchison Falls National Park (Uganda), our model outperforms baselines in forecasting accuracy.