ETH Zurich

A novel diagnostic pipeline unlocks physical interpretation for anomaly detection in learned latent spaces, revealing when anomaly scores diverge from their most physically relevant features.

Fresh, shallow network activations can correct diffusion model feature forecasts, unlocking significant speedups and improved fidelity without any parameter updates.

Misspecifying the coupling mechanism for ordinal data in GP-SCMs inflates counterfactual error threefold, even when observational fit is high, and this error plateaus with more data.

LLM agents can autonomously optimize RAG pipelines by diagnosing retrieval vs. generation failures, outperforming traditional hyperparameter tuning and achieving Pareto efficiency.

Amortized Bayesian causal inference for hidden confounding is now possible via a single transformer pass, achieving state-of-the-art CATE estimation without hyperparameter tuning.

Kilometer-scale 3D reconstructions are now drift-free by closing the loop with SE(3) pose optimization, bypassing Sim(3) limitations.

A single time-series language model now masters forecasting, contextual prediction, and temporal reasoning, outperforming specialized models on key benchmarks.

Critical immersions of even-dimensional manifolds into $\mathbb R^m$ are proven to be real-analytic, extending fundamental regularity results for geometric variational problems.

Value-driven optimistic exploration has historically crumbled under deep network approximations, but diagnosing and repairing failure modes in uncertainty propagation allows pure model-free RL to beat complex model-based exploration baselines.

Forcing all memory into policy weights fundamentally bottlenecks robot manipulation: offloading persistent object identity and containment to an external, LLM-queried ledger jumps benchmark success from 45.9% to 64.3%.

An LLM-discovered counterexample based on Suzuki-Tits ovoid geometry has officially refuted the dominating Hadwiger conjecture.

Scaling humanoid manipulation data no longer requires dedicated teleoperation setups: anchoring retargeting directly in the object's contact frame unlocks automated, kinematically consistent demonstration synthesis from monocular video and across varied object geometries.

Dopant site assignments in atom probe tomography can be fundamentally distorted by preferential field evaporation, proving that direct spatial reconstructions risk misidentifying atomic-scale defects without evaporation-field corrections.

This work demonstrates a compact mobile manipulator that reuses its fingers for locomotion and interaction, without a separate locomotion mechanism.

Deterministic weather models can match dedicated probabilistic ensembles within 0.04–0.13 CRPSS at a 10-day lead time simply by perturbing raw network weights at inference without any retraining.

Achieving accurate center of mass estimation from a single smartphone camera could revolutionize athlete performance analysis in real-world settings.

A new semantic correctness framework reveals that traditional evaluation metrics often overlook critical distinctions in answer quality, leading to misleading assessments of LLM performance.

Conditional Flow Matching outperforms traditional generative models, achieving 66 times the throughput of diffusion models while delivering higher quality designs for inverse problems.

Achieving an 89.2% area reduction in fault monitoring for mixed-criticality SoCs could revolutionize the design of safety-critical systems.

Reducing potential energy surface grid points by up to 81% without sacrificing accuracy could revolutionize computational efficiency in molecular simulations.

Achieving best-in-class energy efficiency of 39 pJ/b, this innovative detector redefines the capabilities of MU-MIMO-OFDM systems.

Multi-valued Byzantine Agreement can now be achieved with dramatically reduced complexity, making large-scale distributed protocols more feasible.

Communication costs for consensus protocols can skyrocket under certain adversarial structures, revealing surprising dependencies on termination requirements.

GenRec achieves unparalleled reconstruction fidelity while enhancing perceptual quality in novel view synthesis by intelligently separating reconstruction from generation.

Achieving near state-of-the-art performance with a 22M encoder, DistillPath-KS16 runs over 25 times faster than its larger counterparts while retaining critical accuracy in pathology tasks.

Financial reasoning accuracy plummets by up to 51% with deeper computations, revealing critical gaps in LLM capabilities.

Achieving state-of-the-art predictive performance while providing transparent explanations, GARLIC transforms how we handle missing data in ICU time series.

MT models show a troubling bias towards US content, with specialized models proving less robust across diverse locales.

GENCO achieves up to 85x speedups in optimal power flow analysis while maintaining higher feasibility and optimality than classical methods.

SpineSegDiff not only matches the best in vertebral segmentation but also reveals critical insights into degenerative disc conditions through uncertainty mapping.

Pre-computed embeddings can outperform traditional supervised models in estimating global biomass, reshaping our approach to carbon stock monitoring.

RAG-Stack uncovers quality-performance trade-offs in RAG systems, achieving up to 153% more effective configurations than current methods.

Adaptive sampling can cut human evaluation costs while boosting the accuracy of model rankings in NLP tasks.

Foundation model agents can achieve stable cooperation in social dilemmas by inferring behavioral similarities, defying classical game theory's expectations of mutual defection.

By refining VLM-derived reward signals with structural priors, SAFT transforms noisy feedback into a reliable guide for faster and more aligned policy learning.

Recursive interaction with time-series data enables TimeRLM to achieve unprecedented accuracy in anomaly localization, outperforming traditional models by a wide margin.

Navigating document structures with a stateful evidence approach boosts answer quality by nearly 8% in complex question answering tasks.

Achieving a Wasserstein mixing time that scales with $\sqrt{d}/\varepsilon$ could revolutionize the efficiency of sampling algorithms in high-dimensional settings.

Disfluencies are not just noise; they carry crucial meaning that, when ignored, significantly degrades translation quality.

Bridging the gap between reinforcement learning and control theory could unlock new synergies in optimizing unknown dynamical systems.

Offloading Parquet decoding to the network can nearly double query throughput, transforming cloud-native database performance.

A unified framework reveals that dexterous manipulation skills are variations of a shared task, enabling seamless chaining and robust performance across diverse objects and hand morphologies.

VidMap achieves unprecedented robustness and accuracy in metric reconstruction from uncalibrated videos, outperforming both SLAM and SfM methods under extreme conditions.

Finite precision in transformers can drastically alter memory operations, revealing a surprising hierarchy of expressivity based on attention mechanisms.

Evaluators using cESA can achieve more reliable translation quality assessments while cutting annotation time by leveraging shared context across multiple outputs.

MLLMs can encode visual evidence but often fail to control their reliance on it, revealing a critical bottleneck in multimodal reasoning.

Resilience isn't just about bouncing back; it’s an emergent property shaped by the interplay of agents, with performance often undermining long-term stability.

DVPSFormer reduces the computational burden of depth-aware video panoptic segmentation, enabling real-time decision-making for autonomous vehicles without sacrificing accuracy.

Suppressing two-body losses by orders of magnitude opens the door to exploring itinerant quantum magnetism and novel quantum droplets in polar molecule systems.

Tripody's innovative overconstrained design achieves 454% higher torsional stiffness at maximum height, revolutionizing lightweight construction robotics.

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

GNM redefines the landscape of head modeling by integrating intricate anatomical details, achieving superior geometric fidelity that existing models overlook.

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

Skill-SP not only pushes the performance ceiling of LLMs but also transforms initially misaligned models into high-performing agents through dynamic skill evolution.

Surprisal theory's reliance on language models leads to tautological predictions, undermining its empirical validity in psycholinguistics.

Larger LLMs may ignore user context more than smaller models, revealing critical insights into how linguistic framing can sway model responses.

Constant-size post-quantum quorum signatures are finally achievable, revolutionizing Byzantine fault tolerance in distributed systems.

LLMs are not just passive observers; they can uncover novel cryptographic vulnerabilities that challenge our current understanding of digital security.

NeuralChaos achieves optimal approximation rates for square-integrable processes using significantly fewer evaluations than traditional methods, reshaping the landscape of stochastic modeling.

Real-time adaptive encoding in a low-power AFE can revolutionize wireless neural signal processing for brain-computer interfaces.

Achieving a 1.6x speedup and 50% reduction in memory usage for CNNs on nano-drones could redefine their operational efficiency and application scope.

Training dynamics of Transformers can be reduced to a low-dimensional manifold, revealing how inductive reasoning emerges from data statistics and model initialization.

Temperature sampling can enhance mode diversity in diffusion models without sacrificing sample quality, thanks to a novel variance-corrective technique.

Achieving 205× latency reduction in FPGA-based neural networks could redefine the benchmarks for ultra-fast inference in latency-critical applications.

ARDY achieves real-time, controllable 3D human motion generation that outperforms existing methods in both fidelity and flexibility.

SBR reduces cognitive fatigue while achieving a 54.1% task success rate, outperforming traditional methods in real-time kinematic retargeting.

Scattering networks can achieve optimal separation capacity by strategically tuning filter frequencies and ensuring well-conditioned geometric couplings.

Only 2-4 principal components can capture 95% of the variance in the solution space of the Burgers equation, revealing a striking effective dimensional reduction.

A dual-$\boldsymbol k$-mesh strategy transforms BSE calculations, achieving unprecedented accuracy and efficiency in modeling absorption spectra.

A single minimal infinite-state tool can elevate finite-precision models to Turing completeness, while finite-state tools add virtually no expressivity.

Continuous severity scoring outperforms traditional multi-class classification in assessing lumbar spine degeneration, revealing finer distinctions in MRI evaluations.

Web agents can now safely interact with complex environments while avoiding prompt injection attacks by masking untrusted content without ever reading it.

LangLoc achieves unprecedented accuracy in indoor localization from natural language, closing the gap between coarse scene retrieval and precise pose estimation.

The learning rate is redefined as a structural parameter of training dynamics, fundamentally shaping the representations selected by gradient descent.

Integrating GNSS-derived Zenith Wet Delay into weather models boosts severe precipitation forecasts by nearly 9%.

For the first time, a spectral sparsification algorithm achieves efficiency without reliance on approximation accuracy, revolutionizing parallel graph processing.

Retaining past knowledge can actually impede real-time adaptation in dynamic environments, leading to a new framework for optimizing continual learning.

The Oracle Distance theorem reveals that all noising processes achieve the same optimal negative ELBO, linking diverse loss functions in diffusion models.

Segmentation masks can bridge the sim-to-real gap, enabling robots to achieve precise control over 23 degrees of freedom in dexterous manipulation tasks.

Synthetic sound effects can fool listeners into mistaking them for real recordings nearly 29% of the time, but a generator-specific detector can perfectly distinguish the two.

Achieving top rankings in a competitive setting, this work reveals how hierarchical soft-label learning can significantly enhance multimodal sexism detection in memes.

LIME turns ordinary egocentric video into a powerful tool for robots to dynamically adjust their camera poses based on user intent, revolutionizing how we think about robotic perception.

AEW achieves optimal performance in expectation for model selection aggregation, revealing a critical phase transition that could redefine its application in statistical learning.

Refining Cover's theory reveals that low-dimensional data structures can dramatically enhance classification capabilities, challenging traditional assumptions in machine learning.

SuperFlex achieves unprecedented reconstruction accuracy for 3D point clouds by enabling deformable superquadrics to represent complex geometries robustly.

Preemptive VCs can slash link resource usage by 76% while maintaining comparable performance in deadlock-free AXI4 NoCs.

Category theory reveals that AI identity is not a simple equality of trustworthiness but a complex interplay of transformations and histories.

Resonant modifications of London dispersion interactions can significantly enhance reaction rates in molecular systems under vibrational strong coupling.

RAISE keeps LLM-based heuristics resilient, outperforming traditional methods by up to 19 times under real-world distribution shifts.

JETO-Mine uncovers a staggering 660 execution time improvement patches in Java, revealing a critical lack of performance testing in open-source projects.