We are pleased to announce that the Machine Learning and Simulation Lab has had seven papers accepted this year, including one oral presentation (Andrei Manolache) and one spotlight presentation (Jan Hagnberger and Daniel Musekamp). These works reflect the group’s contributions across equivariant modeling, diffusion generative models, PDE surrogate modeling, medical vision-language alignment, 3D point-cloud transformers, equality-constrained learning, and reward-alignment algorithms.
Accepted Papers
- Learning (Approximately) Equivariant Networks via Constrained Optimization
Andrei Manolache, Luiz Chamon, Mathias Niepert (Oral)
arXiv: https://arxiv.org/abs/2505.13631
The paper introduces Adaptive Constrained Equivariance (ACE), a constrained-optimization framework that begins with a flexible non-equivariant model and gradually tightens equivariance constraints through a homotopy-style schedule. This allows the network to exploit partial or approximate symmetries common in real-world data, improving sample efficiency, robustness, and downstream task performance compared to strictly equivariant or heuristically relaxed approaches.
- Rao-Blackwell Gradient Estimators for Equivariant Denoising Diffusion
Vinh Tong, Trung-Dung Hoang, Anji Liu, Guy Van den Broeck, Mathias Niepert
arXiv: https://arxiv.org/abs/2502.09890
This work proposes a Rao-Blackwellized gradient estimator for denoising diffusion models defined on symmetric domains, obtained by marginalizing data-augmentation orbits. The resulting estimator significantly reduces variance, improving training stability, convergence behavior, and performance on molecular and crystal benchmarks. The method additionally improves protein generation models such as Proteina, yielding higher structural fidelity and more consistent generative samples.
- CALM-PDE: Continuous and Adaptive Convolutions for Latent Space Modeling of Time-Dependent PDEs
Jan Hagnberger, Daniel Musekamp, Mathias Niepert (Spotlight)
arXiv: https://arxiv.org/abs/2505.12944
The paper presents a surrogate modeling framework for time-dependent PDEs using a latent representation constructed via continuous convolutional kernels applied to learned adaptive query points. The decoder supports querying arbitrary spatial locations, enabling efficient modeling on both regular and irregular meshes. CALM-PDE achieves improved memory efficiency and competitive or superior accuracy compared to transformer-based PDE surrogates.
- ExGra-Med: Extended Context Graph Alignment for Medical Vision-Language Models
Duy M. H. Nguyen, Nghiem Diep, Trung Nguyen, Hoang-Bao Le, Tai Nguyen, Anh-Tien Nguyen, TrungTin Nguyen, Nhat Ho, Pengtao Xie, Roger Wattenhofer, Daniel Sonntag, James Zou, Mathias Niepert
arXiv: https://arxiv.org/abs/2410.02615
This work introduces a multi-graph alignment framework for medical vision-language models. Images, instruction responses, and extended captions produced by a frozen LLM are organized into structured graphs, and a barycenter alignment procedure enforces consistency across modalities. The approach enhances performance in low-data regimes, enabling improvements over standard VLM training pipelines while using only a fraction of the data.
- How Many Tokens Do 3D Point Cloud Transformer Architectures Really Need?
Tuan Tran, Duy M. H. Nguyen, Hoai-Chau Tran, Michael Barz, Khoa D Doan, Roger Wattenhofer, Vien Ngo, Mathias Niepert, Daniel Sonntag, Paul Swoboda
arXiv: https://arxiv.org/abs/2511.05449
The authors investigate token redundancy in 3D point-cloud transformers and demonstrate that many architectures use far more tokens than necessary. They introduce a globally informed graph-based token merging technique that reduces token counts by up to 90-95% without sacrificing performance on semantic segmentation and reconstruction tasks. The results challenge common assumptions about token scaling in 3D transformers.
- Learning with Statistical Equality Constraints
Aneesh Barthakur, Luiz Chamon
arXiv: https://openreview.net/forum?id=oDA6t2RaFC
This paper studies statistical learning problems with equality constraints, such as fairness or calibration conditions. It provides generalization bounds for equality-constrained learning and shows how solving a sequence of unconstrained empirical problems can approximate the desired constrained solution. The approach offers both theoretical guarantees and a practical training procedure.
- Mitigating Reward Over-Optimization in Direct Alignment Algorithms with Importance Sampling
Nguyen Phuc, Ngoc-Hieu Nguyen, Duy M. H. Nguyen, Anji Liu, An Mai, Thanh Binh Nguyen, Daniel Sonntag, Khoa D Doan
arXiv: https://arxiv.org/abs/2506.08681
The authors address reward over-optimization in offline Direct Alignment Algorithms (DAAs) such as Direct Preference Optimization (DPO). They propose an importance-sampling-adjusted objective that incorporates clipped importance ratios to correct for distributional drift between the learned policy and the offline data. The method reduces over-optimization and outperforms alternative regularization approaches in low-regularization settings.
