ETH Zurich

AI-driven summaries of public consultations can systematically exclude dissenting voices, raising concerns about biased policy recommendations even when individual outputs seem reasonable.

Deterministic decoding can outperform stochastic self-consistency in constrained domains by systematically exploring high-probability reasoning traces, leading to better performance with less computation.

TurboQuant's claimed advantages over RaBitQ in quantization don't hold up under rigorous, reproducible comparison, raising questions about its practical utility.

Forget heavyweight processes and bandwidth bottlenecks: Proxics offers a lightweight programming model that unlocks the potential of near-data processing with efficient virtual processors and optimized communication channels.

Bridging the gap between trust region methods and PPO, this new framework guarantees performance improvements while outperforming existing algorithms in stability and effectiveness.

Early layers of language models capture human-like processing signatures in reading, rivaling traditional measures like surprisal in predicting initial eye movements.

Neural operators can achieve uniform convergence rates for approximating solution maps across diverse geometric domains, challenging traditional assumptions about shape-dependent PDE solutions.

GSQ closes the accuracy gap in low-precision quantization, achieving results comparable to complex vector methods while remaining easy to implement.

Trustworthy super-resolution in surgery is now achievable, with a model-agnostic method that identifies and mitigates unreliable reconstructions in real-time.

RadAgent doesn't just give you the answer; it shows its work, offering clinicians a transparent, step-by-step reasoning trace for AI-generated CT reports.

SCENIC delivers the best of both worlds: the high bandwidth and software integration of commercial SmartNICs, plus the customization and data processing offload capabilities of research prototypes.

Europe's collaborative EPAC chip delivers a heterogeneous RISC-V accelerator, showcasing a path towards domain-specific HPC hardware built with open standards.

LLMs struggle to simulate culturally nuanced emotional responses to bureaucratic processes, especially in Eastern cultures, suggesting current models lack the socio-cultural understanding needed for accurate policy simulation.

Blind predictions of cyclobutanone photochemistry reveal that nonadiabatic molecular dynamics can qualitatively capture experimental results, but the accuracy of underlying electronic structure calculations remains a key bottleneck.

Supercomputers can evolve beyond just pre-training to become comprehensive "AI Factories" by adopting hybrid cloud-native architectures that support the entire lifecycle of foundation models.

Continual learning just got a turbo boost: C-Flat Turbo cuts training time by up to 25% without sacrificing accuracy, thanks to a clever gradient-skipping trick.

Unlock up to 25.7% accuracy gains on frozen LLMs in knowledge-intensive domains, without any retraining, by dynamically rewarding reasoning steps.

Achieve perceptually superior video compression at extremely low bitrates by using implicit neural representations to condition diffusion models, outperforming even VVC and prior neural codecs.

Self-supervised learning on heterogeneous neutrino detector data enables foundation-style models that achieve state-of-the-art performance with an order of magnitude less labeled data.

Training 3D avatar diffusion models on millions of in-the-wild videos is now possible, thanks to a clever 3D tokenization and visibility-aware training strategy that overcomes partial observability.

Unlock interactive digital twins from messy, real-world videos: FunRec automatically turns egocentric RGB-D recordings into simulation-ready 3D scenes.

Polarization isn't always about echo chambers: Europeans can agree on *what* happened in the Ukraine war, but vehemently disagree on *why* it matters.

Scaling up avatar pre-training to 1M in-the-wild videos unlocks emergent generalization capabilities like relightability and garment support, even without direct supervision.

AI agents are far better at automating data engineering tasks than previously thought, but flawed benchmarks are obscuring their true potential.

Forget brittle, overfit skills – Trace2Skill distills diverse execution experiences into transferable agent skills that boost performance by up to 57.65% on unseen tasks, even when transferring skills learned by smaller models to larger ones.

Forget about retraining: MUNKEY offers zero-shot machine unlearning by simply deleting instance-identifying keys, outperforming traditional post-hoc methods.

Current computational aberration correction methods struggle to generalize across different camera lenses, but this new benchmark and analysis pinpoint the key factors holding them back.

Ditch the slow diffusion grind: Marigold-SSD delivers zero-shot depth completion in a single step, rivaling discriminative models in speed while retaining diffusion's accuracy.

Unlock superior trajectories in complex environments with a new ADMM-based solver that jointly optimizes spatial and temporal domains, eliminating the need for complex warm starting.

Get 6x the RLHF alignment for your LLM with a new active learning pipeline that focuses on annotating the most informative response pairs.

Clever reticle placement on wafer-scale systems can boost throughput by 2.5x and slash latency by over a third, offering a hardware-level speedup for LLM training.

LLMs can follow detailed code refactoring instructions, but still fall short of mimicking human refactoring choices in real-world codebases, highlighting a critical gap in their ability to autonomously improve code quality.

Finally, a virtual try-on system that actually works: Gaussian Wardrobe lets you swap clothes between 3D avatars with high-fidelity garment dynamics by learning shape-agnostic garment layers.

Ditch the memory banks and prototype comparisons: this method learns a compact, parametric model of normal image embeddings with an autoregressive CNN, slashing inference time and memory in unsupervised anomaly detection.

Multimodal models often exhibit lower confidence than their unimodal counterparts when they're about to fail, and this work leverages that insight to build a better failure detector.

Reasoning can boost LLM opinion alignment, but it's not a silver bullet for removing bias in political digital twins.

Forget computationally expensive fluid dynamics: this work shows that a simple, stateless model, carefully calibrated to real-world data, can create surprisingly effective digital twins for soft underwater robots.

LLM benchmark translations can be dramatically improved by test-time compute scaling, revealing a surprisingly cheap way to get more reliable multilingual evaluations.

Forget temperature scaling: JUCAL calibrates aleatoric and epistemic uncertainty in classifier ensembles, achieving SOTA results with significantly smaller ensembles and lower inference costs.

Forget solo Git tutorials—GitAcademy's split-screen view, mirroring a partner's actions in real-time, makes learning collaborative workflows feel less like a lonely commit and more like a team sport.

Unlock domain generalization with unlabeled data by exploiting the structure of anti-causal relationships, where outcomes cause covariates.

E-graphs, typically confined to isolated optimization steps, can now persist as a first-class citizen within the compiler's intermediate representation, unlocking broader and more flexible program optimization.

Forget complex architectures: RaCo achieves SOTA keypoint matching and repeatability by cleverly combining ranking and covariance estimation in a lightweight network, trained without covisible image pairs.

Context files like AGENTS.md, intended to guide coding agents, often *hurt* performance and increase costs, challenging the common practice of using them.

Achieve >97.5% of full-data VIT performance with only 16% of the data using ScalSelect, a surprisingly effective and scalable training-free data selection method.

To address the ethical imperative of improving access equality for CGT in Europe, further policy reforms are proposed including concurrent HTA, early benefit assessment, and incorporation of additional elements of value in HTA evaluations, alongside current initiatives to increase cross-border collaboration.

An interpretable deep learning model, ECG-XPLAIM, rivals ResNet in arrhythmia detection sensitivity while offering crucial insights into its decision-making process via Grad-CAM.

Multimodal LLMs often perform worse with more modalities because they struggle to jointly recognize and reason across modalities, a problem solvable with simple prompting strategies.

A new deep learning model slashes the error rate for BMI estimation from smartphone photos, opening the door to more accessible and convenient health assessments.

Automating CAD design from text prompts is now feasible, with visual feedback loops boosting performance, especially for multimodal LLMs.

Achieve faster, near-optimal path planning in complex 3D environments by combining any-angle search with multi-resolution grids, outperforming even sampling-based methods.

LLMs that excel at math don't necessarily make good math tutors, revealing a surprising trade-off between subject matter expertise and pedagogical skill.