Institution profile

Munich Center for Machine Learning

Academic institutioneurope · de
Official website
Research library507linked papers
Opportunities0open roles
Selected work

Representative Papers

Examining marginal properness in the external validation of survival models with squared and logarithmic losses

Dec 10, 2022

This paper addresses the theoretical validity of two widely used external validation metrics for survival analysis models—Integrated Survival Brier Score (ISBS) and Right-Censored Log-Likelihood (RCLL)—by introducing “marginal propriety” as a novel formal criterion for scoring rule appropriateness. We prove theoretically that neither metric satisfies marginal propriety. However, Monte Carlo simulations and extensive experiments across diverse right-censored survival modeling scenarios demonstrate that RCLL consistently satisfies this property empirically, while ISBS exhibits only negligible violations under extremely small sample sizes, remaining robust in practice. This reveals a critical dissociation between theoretical impropriety and empirical robustness—a key insight with important implications for metric selection and design. Building on this finding, we propose a new class of loss-function frameworks for survival prediction, grounded in marginal propriety, thereby providing both theoretical guidance and practical foundations for developing future survival scoring rules.

3 citationsRead paper

Decoupling the Effect of Chain-of-Thought Reasoning: A Human Label Variation Perspective

Jan 06, 2026arXiv.org

This study investigates the efficacy of chain-of-thought (CoT) reasoning in probabilistic, ambiguous tasks that model human label variability, moving beyond conventional deterministic single-answer settings. The authors propose Cross-CoT, a decoupling framework that integrates distribution alignment metrics, variance contribution analysis, and reasoning trace tracking to disentangle, for the first time, the distinct influences of CoT-generated rationales and model priors on output distributions from the perspective of human annotation variation. Their findings reveal that CoT overwhelmingly governs final answer selection—accounting for 99% of accuracy variance—whereas model priors predominantly shape the output distribution’s structure, influencing over 80% of its ranking properties. Moreover, while CoT monotonically improves accuracy during inference, it exerts limited influence on distributional form, thereby challenging the prevailing assumption that CoT enables fine-grained calibration of output distributions.

2 citationsRead paper

Highly Accurate and Diverse Traffic Data: The DeepScenario Open 3D Dataset

Apr 24, 2025

Conventional onboard 3D trajectory datasets suffer from occlusions and limited field-of-view, hindering accurate modeling of distant traffic participants. To address this, we introduce DSC3D—the first open, occlusion-free, high-precision 6DoF monocular UAV-based 3D trajectory dataset—featuring 175K+ multi-class trajectories across five representative traffic scenarios. We propose the first end-to-end monocular UAV 3D tracking pipeline, uniquely covering long-tail scenarios such as complex human-vehicle interactions and full-cycle parking. Accuracy is ensured via monocular visual SLAM, multi-object trajectory optimization, UAV collaborative calibration, and multi-sensor spatiotemporal alignment. We release the dataset alongside an interactive visualization platform. Extensive evaluation demonstrates DSC3D’s effectiveness on motion prediction, behavioral modeling, and safety verification tasks, enabling research on generative, reactive traffic agents.

1 citations1 influentialRead paper

Step-resolved data attribution for looped transformers

Feb 10, 2026

This work addresses the challenge of characterizing the influence of individual training samples across the iterative steps of recurrent Transformers, a capability lacking in existing data influence estimation methods. To this end, the authors propose Step-Decomposed Influence (SDI), which unfolds the recurrent computation graph to decompose data influence at each inference step. By integrating the TracIn framework with TensorSketch approximation, SDI avoids explicit per-sample gradient computation, enabling efficient and scalable fine-grained attribution. Experiments demonstrate that SDI achieves high accuracy and strong scalability on recurrent GPT models and algorithmic reasoning tasks, facilitating multi-dimensional interpretability analyses of the internal reasoning dynamics within recurrent architectures.

1 citationsRead paper

3D Molecule Generation from Rigid Motifs via SE(3) Flows

Jan 23, 2026

Traditional 3D molecular generation methods, which treat atoms as fundamental units, struggle to efficiently model complex molecular structures. This work proposes a novel paradigm by representing molecules as collections of rigid structural motifs and, for the first time, integrates SE(3)-equivariant generative flows for 3D molecular generation. The proposed approach substantially improves both generation efficiency and representation compactness. On the GEOM-Drugs dataset, it achieves superior atomic stability compared to existing methods while reducing the number of generation steps by a factor of 2–10 and attaining a 3.5× compression ratio in molecular representation.

1 citationsRead paper
Recent publications

Latest Papers

Memorisation bias in medical AI

Sep 15, 2026

研究揭示了医疗AI模型因记忆训练数据中的患者历史记录而产生的'记忆偏差'问题，影响未来诊断准确性，并提出需改变现有模型训练和部署协议以缓解该风险。

0 citationsRead paper