Great Success for the Interactive Machine Learning (IML) Department at NeurIPS 2025

Strong Presence at One of the Most Prestigious ML Conferences

NeurIPS (Conference on Neural Information Processing Systems) is among the world’s leading conferences in machine learning and artificial intelligence. With an acceptance rate of around 25%, the selection process is highly competitive. The acceptance of four contributions from the IML department highlights DFKI’s strong scientific position in the international research landscape.

Overview of New Research Results

Efficient Medical Vision–Language Models: ExGra-Med

The contribution ExGra-Med introduces a new multi-graph alignment framework for improved semantic integration of visual and textual information in medical models. Unlike existing approaches, ExGra-Med avoids the need for large-scale training datasets and instead relies on the joint alignment of images, responses, and extended image captions within a shared latent space. Key results: up to 90% less training data required compared to competing models; 20.13% performance improvement on the medical VQA benchmark VQA-RAD; outperforms established approaches such as BioMedGPT and RadFM. 

Efficiency Gains in 3D Transformers through Token Merging

A second contribution focuses on reducing computational costs in 3D point cloud transformers. The proposed token merging strategy demonstrates that up to 95% of tokens can be removed during both inference and training without degrading performance in segmentation, reconstruction, or object detection. This results in substantial efficiency gains and opens up new perspectives for scalable 3D models.

More Stable Alignment with Importance Sampling

The paper IS-DAAs introduces a new method to address stability issues in Direct Alignment Algorithms (DAAs). By incorporating an importance sampling weight into the optimization objective, the approach mitigates over-optimization, which often leads to performance degradation in existing methods. Extensive experiments show that the method yields more stable results, particularly under low regularization, and outperforms conventional approaches.

AuditCopilot: AI Support for Financial Audits

As part of a workshop, AuditCopilot was presented—an approach for fraud detection in double-entry bookkeeping using large language models (LLMs). By combining domain-specific reasoning with structured accounting data, the system enables more efficient identification of anomalies, inconsistencies, and potential fraud patterns. This opens up new opportunities to support auditors and to improve the overall quality of financial audits.

Diverse Conference Program and International Networking

NeurIPS 2025 offered a rich and diverse program featuring invited talks, tutorials, poster sessions, and a wide range of workshops. Key topics such as multimodal foundation models, graph machine learning, and GenAI for Health reflect current research priorities that are also strongly pursued within the IML group.

In addition, discussions were held with international research groups during the conference, particularly in the fields of artificial intelligence, robotics, and biomedical research. These exchanges strengthen existing collaborations and pave the way for joint projects in the coming years.

Conclusion

The successful participation in NeurIPS 2025 confirms the high international competitiveness of the IML department and highlights the relevance of the research results developed for the scientific community and for practical applications.