🏢 ETH Zurich

Boosting reinforcement learning, this paper introduces h-PMD, a novel algorithm enhancing policy mirror descent with lookahead for faster convergence and improved sample complexity.

This paper introduces a novel notion of intrinsic similarity between language encoders, based on affine homotopy, and demonstrates its strong correlation with extrinsic similarity (downstream task per…

NEORL: Novel nonepisodic RL algorithm guarantees optimal average cost with sublinear regret for nonlinear systems!

MultiOOD benchmark and novel A2D & NP-Mix algorithms drastically improve multimodal out-of-distribution detection.

LiteVAE: A new autoencoder design for latent diffusion models boosts efficiency sixfold without sacrificing image quality, achieving faster training and lower memory needs via the 2D discrete wavelet …

JKOnet* learns diffusion processes at unprecedented speed and accuracy by directly minimizing a simple quadratic loss function, bypassing complex bilevel optimization problems.

Learned Bregman divergences significantly improve image corruption robustness in adversarial training.

Boosting deep learning generalization, this work unveils SAM’s implicit regularization using ‘balancedness’, a new metric. A resource-efficient variant, BAR, achieves 95% computational savings with i…

FUSE, a novel framework, efficiently predicts continuous fields & estimates discrete parameters in PDEs, significantly improving accuracy and robustness.

Fairness in social influence maximization is achieved via optimal transport, optimizing both outreach and a new ‘mutual fairness’ metric that considers variability in outreach scenarios.

LLM quantization, while improving efficiency, creates a security risk: attackers can craft seemingly benign models that exhibit malicious behavior only when quantized.

4DGF, a novel neural scene representation, achieves interactive-speed novel view synthesis for large-scale dynamic urban areas by efficiently combining 3D Gaussians and neural fields.

DeltaDEQ accelerates deep equilibrium model inference by 73-84% via a novel ‘heterogeneous convergence’ exploitation technique, maintaining accuracy.

LLMs struggle to cooperate sustainably; GOVSIM reveals this, showing communication and ‘universalization’ reasoning improve outcomes.

Contextual Bilevel Reinforcement Learning (CB-RL) tackles real-world strategic decision-making where optimal policies depend on environmental configurations and exogenous events, proposing a stochasti…

ConStat: Exposing hidden LLM contamination!

LLMs regulate uncertainty via specialized ’entropy’ and ’token frequency’ neurons, impacting prediction confidence without directly altering logits.

Novel compositional PAC-Bayes framework delivers data-dependent generalization bounds for persistence-enhanced Graph Neural Networks, improving model design and performance.

This paper introduces a novel algorithm to accurately compute causal effects within specific sub-populations, even when hidden factors influence the data, advancing causal inference significantly.

This paper introduces a novel quantitative definition of AI alignment for social decision-making, proposing probably approximately aligned policies and a method to safeguard any autonomous agent’s act…