Technical University of Munich

Treating 3D Gaussian reconstruction as an active structural prior rather than a passive auxiliary target drives a 22% performance leap in category-level 6D pose estimation.

Norm2Tex is introduced, a plug-in method that augments simulations of vision-based tactile sensors with high-frequency surface details from normal map textures that preserves material-dependent tactile information across domains, improving texture recognition and producing material-dependent control behavior in the real world.

Even "leak-free" protein interaction benchmarks are riddled with non-biological shortcuts that models readily exploit—and standard negative-sampling heuristics often make the bias worse.

UAR-Net achieves state-of-the-art image restoration from adverse weather by effectively modeling both global interactions and local structures, all while estimating predictive uncertainty.

Deep learning models can accurately estimate the weight of falling particles, transforming how we approach contactless measurement in industrial applications.

Iterative text-guided image fusion can significantly enhance perceptual quality and information preservation, even in the presence of complex degradations.

Achieving 4x super-resolution in diffusion MRI with a novel transfer-learning approach that cuts training time by 6x while significantly improving image quality.

Trade restrictions emerge as the dominant risk in semiconductor supply chains, identified through a novel pipeline that combines LLMs and expert validation to score over 76,000 risk items.

CMRVision outperforms existing models in cardiac MRI analysis, achieving remarkable segmentation accuracy and setting a new benchmark for multi-task performance in this domain.

PriMD revolutionizes multimodal emotion recognition by enabling robust performance even when critical data modalities are missing.

TACTIC reveals that integrating clinical context into WB-MRI analysis can significantly boost diagnostic accuracy, even when faced with incomplete data.

Lightweight classifiers can outperform larger models in handling voice assistant fallbacks, transforming user interactions from failures to opportunities.

LLMs misjudge research ideas as "medium novel" due to a systematic bias, but a new probing method boosts their accuracy by over 22%.

Jointly modeling aleatoric and epistemic uncertainties can dramatically enhance prediction reliability in high-dimensional outputs, outperforming traditional methods.

Achieving 22% better accuracy in spine anatomy reconstruction without any synthetic training data could revolutionize intraoperative ultrasound applications.

Conceptual framing of bias definitions can lead to significant discrepancies in annotation, impacting both human and LLM assessments.

FedCC enables clients to handle ambiguity in data, leading to a remarkable 67.3% accuracy even when facing severe label distribution skews.

Neglecting cross-view correspondence can lead to misleading evaluations, with nearly 56% of trajectory pairs showing significant disagreement in agent assessments.

CEFITO's innovative action-conditioned representation space allows for precise inference-time planning by eliminating irrelevant actions, setting a new standard in procedure planning accuracy.

CUBICS reveals that context-aware performance estimation can significantly improve the reliability of safety-critical ML components by avoiding oversimplified failure models.

Open-vocabulary monocular 3D detectors mislabel correctly localized objects due to prompt sensitivity, revealing a critical gap in their semantic understanding.

Multilingual models may excel in general tasks, but they falter significantly when faced with region-specific cultural knowledge, as shown by BavGround's rigorous evaluation.

A staggering 71.6% of LLM conversations about self-treatment led to unsafe medical advice, revealing critical flaws in current safety assessments.

Decoupling coordinate frame selection from box regression leads to a remarkable 11% accuracy boost in 3D visual grounding tasks.

Earth embeddings could revolutionize satellite data analysis by enabling users to leverage compact feature vectors without the overhead of processing raw imagery.

AI agents can evolve from mere tutors to multifaceted educational tools, reshaping how students engage with complex frameworks.

CLA scores based on English can predict translation quality better than direct source-target alignment, highlighting English's role as a crucial pivot in multilingual LLMs.

Achieving a mean MS-SSIM of 0.95 for T1CE MRI, this method outperforms traditional interpolation techniques while providing reliable confidence estimates for clinical use.

Translating LTL to LTLf+ unlocks efficient finite automata techniques for a wide array of AI applications without sacrificing complexity.

Real-time 3D reconstruction is now feasible by merging visual-inertial tracking with an innovative incremental Gaussian splatting technique.

Binary classifiers misclassify nearly 40% of AI agents as humans, but a simple three-class framework achieves perfect detection across all tested evasion strategies.

NLP-driven backlog enrichment can transform how engineers identify and integrate security requirements, achieving impressive relevance ratings in real-world applications.

By framing camera trajectory generation as a language-grounded spatial reasoning problem, CinemaTraj achieves high-quality, collision-free cinematographic outputs that align closely with user prompts.

Smaller decision trees can maintain optimality while providing clearer, more understandable explanations for complex decision-making processes.

Trust in AI-assisted code reviews can be paradoxically undermined by too much explanation, as developers may question recommendations more when provided with detailed reasoning.

More accurate 3D face tracking in ViDS enables unprecedented control over expression and pose in portrait animations, setting a new standard for identity preservation.

Early design choices in the mapping pipeline can silently propagate errors, jeopardizing the reliability of large-scale geospatial maps.

Incorporating perceived transaction costs into LLM simulations can dramatically enhance the accuracy of tenant response predictions to energy policy interventions.

Achieving state-of-the-art hair dynamics in head avatars, DynHair allows for realistic and controllable hair movement that adapts to head motion.

Randomized KV-cache eviction not only enhances error attribution but also reveals the limitations of deterministic methods in managing attention-output fidelity.

PRIME-SVR achieves unprecedented reconstruction quality in fetal brain imaging, enabling the first isotropic T2 maps at previously inaccessible echo times.

A novel initialization layer enables GNNs to adaptively transfer across urban environments, achieving superior performance with significantly reduced training times.

AutoJourn not only generates balanced news summaries but also actively detects and neutralizes bias, setting a new standard for responsible automated journalism.

Hidden safety risks in AI systems may be more dangerous than the visible failures we obsess over, revealing a critical need for a new framework to diagnose them.

Current ADS testing practices are hampered by major challenges, but an evidence-centered closed-loop framework could revolutionize how we ensure their safety and functionality.

Training generative motion models with constraints directly integrated into the objective can drastically reduce collision rates and improve trajectory quality in robot navigation tasks.

Achieving real-time 4K rendering at 60 FPS on consumer hardware, this method outpaces traditional techniques by up to five times without sacrificing visual quality.

Current remote sensing models falter in hierarchical reasoning, but HieraPlan sets a new standard for cognitive analysis in geospatial contexts.

Achieving nearly 800x speedup in 3D scene modeling without sacrificing quality, AsySplat redefines efficiency in long-sequence novel view synthesis.

Achieving up to 9 percentage points improvement in latency prediction accuracy, HiFi-LLP revolutionizes hardware-aware neural architecture search efficiency.

Current RE practices fall short in supporting explainability, revealing critical gaps that could jeopardize safety in AI systems.

EquiFusion achieves unprecedented cross-dataset generalization in human motion prediction, revolutionizing how we handle diverse kinematic data.

TWIN achieves ab initio accuracy for biomolecular systems while being two orders of magnitude faster than traditional methods, revolutionizing the modeling of drug interactions and protein dynamics.

Steering driving world models in real-time without retraining could revolutionize how we control AI-generated simulations.

D-SafeMPC achieves safer and more efficient robotic planning by seamlessly integrating diffusion models with model predictive control, overcoming key limitations of both approaches.

Achieving robust robot skill generalization while maintaining critical motion geometry, SMP reduces dynamic violations and preserves end-effector paths during execution.

Automating the quality assessment of LLM-generated defeaters could revolutionize how we validate safety claims in high-integrity systems.

Achieving anatomically consistent multi-view fusion improves stenosis reporting accuracy, overcoming the limitations of single-view models.

Gaussian Splatting transforms real-world driving footage into high-fidelity simulation scenarios, drastically improving the realism of autonomous driving tests.

3DMPE reconstructs 3D point clouds from incomplete 2D views without the need for training data, setting a new standard for geometric reconstruction methods.

Predicting breast cancer treatment response just got a major upgrade, with a new framework that outperforms traditional models by effectively modeling temporal imaging data.

Energy consumption in AEVs can vary drastically based on traffic conditions, a factor often overlooked in autonomous driving research.

Real-world testing uncovers that model-level metrics can mislead safety assessments, with camera systems exhibiting failures that offline evaluations fail to predict.

Refinement complexity in automotive requirements is driven more by architectural scope than by linguistic verbosity, revealing critical insights for improving product development efficiency.

Current avatar systems are more diverse than ever, yet foundational prior learning is often overlooked in discussions of photorealistic digital humans.

SEDCoT achieves a 12% improvement in translation accuracy over existing methods while enhancing the readability of COBOL code translations into C.

ISU-Test reveals that systematic scene generation can dramatically improve the detection of failures in vision-language models used for critical in-car safety applications.

Log$_\text{b}$Quant achieves superior quantization performance, enabling efficient deployment of language models on consumer hardware without sacrificing accuracy.

Triplet recall improves by 77.4% with DeWorldSG, setting a new benchmark for 3D scene graph generation.

CommonRoad-Game enables real-time human interaction in autonomous driving simulations, overcoming the limitations of existing platforms.

Achieving state-of-the-art garment modeling accuracy without the need for physical simulations could revolutionize digital fashion design.

Models may give clearer instructions, but they fail to engage learners deeply, resulting in passive instruction-following rather than active understanding.

Semantic Reference Frames reveal that optimizing the trajectory of language model computation can significantly enhance parameter efficiency and reduce complexity.

Appropriateness in TTS varies significantly across domains, challenging the assumption that naturalness alone suffices for effective evaluation.

A holistic framework categorizing knowledge-intensive processes reveals six distinct models that can significantly enhance digital transformation efforts.

Balancing autonomy and robustness in AI agents could redefine how we implement and manage intelligent systems in dynamic environments.

Non-parametric identification of causal mechanisms from steady-state observations could revolutionize data analysis in fields constrained by experimental limitations.

Budget-adaptive routing can outperform strong models while cutting per-frame latency by nearly 30%—a game changer for edge-cloud inference efficiency.

Articulated 3D object reconstruction can now achieve high fidelity and internal structure recovery from mere text or images, thanks to a debate-driven agentic approach.

NeuReasoner reveals that while LLMs can excel in certain reasoning tasks, they still falter in critical areas like decision-making under uncertainty, challenging previous assumptions about their capabilities.

Image editing models may appear visually stunning, but they struggle with accurately reflecting real-world lighting, especially in shadowed regions.

Sparse demonstrations can now effectively bootstrap humanoid loco-manipulation learning, reducing the need for constant human oversight.

Many popular uncertainty metrics mislead decision-making, but new decision-aligned metrics show a striking improvement in utility alignment.

BtrLog slashes commit latency and boosts transaction throughput, outperforming traditional cloud logging solutions like EBS.

Generating realistic patient embeddings can yield performance on par with full datasets, even in scenarios with missing classes.

Current LLMs achieve negligible runtime and memory optimizations, while expert implementations deliver up to 15.5x speedup and 171.3x memory reduction.

A dynamic load balancer can reduce node idle time to nearly a millisecond in complex UQ workflows, revolutionizing how we approach scheduling in high-performance computing.

Triangle splats from video diffusion latents yield superior geometric accuracy and visual quality, challenging the dominance of volumetric 3D Gaussians in scene generation.

Partial data augmentation can match the statistical benefits of full augmentation, challenging the notion that complete symmetry is necessary for optimal learning.

PANY outperforms existing model-free methods by over 20% in pose accuracy, even in challenging conditions with limited reference overlap.

P-JEPA achieves state-of-the-art action classification on long procedural videos while using an order of magnitude fewer parameters than existing models.

Collapsed Effective Operators can significantly enhance spectral clustering and neural network architectures by effectively encoding long-range topological interactions in a single operator.

Misaligned perceptions of AI use among student partners can significantly hinder collaboration, especially for those with lower programming skills.