an importance-sampling-based method to mitigate over-optimization in Direct Alignment Algorithms for language model alignment

We demonstrate that the PFN-framework allows for the accurate estimation of causal effects under weakened assumptions.

MyoChallenge 2024 at NeurIPS introduced a benchmark for coordinated control of biological and prosthetic limbs through two tasks, manipulation and locomotion, using physiologically realistic models.

Recourse algorithms can induce a distribution shift that makes their own recommendations invalid.

We systematically evaluate the concerns of multiplicity and monoculture in a suite of large language models and prediction tasks.

We introduce a quantization-free way to train autoregressive transformers for continuous action decision making, improving on discretized action methods.

We present NPE-PFN, a method that uses TabPFN for training-free and simulation-efficient Bayesian inference.

We introduce the Fixed-Point RNN framework to solve state-tracking tasks by parameterizing the state transition matrix as implicitly dense.

We analyze how collective strategic behavior influences predictive models, introducing level-$k$ reasoning and showing how user coordination affects equilibrium outcomes.

whether and how transformers perform in-context counterfactual reasoning

Adam with equal betas works well, and its form can be simplified for further insights

This paper introduces distributional training data attribution, a data attribution framework that accounts for stochasticity in deep learning training, enabling a mathematical justification for why influence functions work in this setting.

We propose a novel algorithms, called NGN-M with a strong theoretical convergence analysis and extensive numerical evaluations showing the robustness of our algorithm to the choice of the learning rate hyperparameter.

BEDLAM2.0 is a significant expansion of the popular BEDLAM dataset that increases pose and body shape variation, adds shoes, and, most importantly, introduces a wide range of realistic cameras and camera motions.

We propose a method for performing simulation-based inference (SBI) of function-valued parameters, and apply it to spatial inference problems in the geosciences.

We develop DISCOVER, which enables RL agents to solve substantially more challenging tasks than previous exploration strategies in RL.

We introduce Adaptive Constrained Equivariance, a homotopy-inspired constrained optimization apprach for training equivariant neural networks.