Jacob Andreas

NUTS Is Fast in Samples, Not in Wall Time: Rethinking MCMC Choice for Neural Likelihood Emulators in Cosmology

This paper benchmarks five MCMC posterior sampling algorithms — MH, MALA, HMC, NUTS, and AIES — against fully differentiable neural-network likelihood emulators for cosmological parameter inference. While NUTS achieves the fewest effective samples to convergence, its per-step computational overhead erodes that advantage in wall-clock time, leaving simpler gradient-based methods like MALA and even standard MH competitive. Whitening and covariance adaptation are identified as broadly impactful efficiency improvements across methods. The work is accompanied by the open-source BEST (Batched Emulator Sampling with TensorFlow) package, enabling general-purpose sampling of arbitrary TensorFlow likelihoods.

Writer Retrieval Outperforms OCR for Duplicate Detection in Handwritten Civic Signature Lists

Automating the validation of handwritten signature lists for Swiss direct democracy is constrained by the poor performance of off-the-shelf OCR on short, out-of-vocabulary entries like names and addresses. A proposed pipeline combining template-based line segmentation with both OCR and writer retrieval reveals a fundamental split: OCR is insufficiently reliable for transcription (29.6% CER on first names), while writer retrieval achieves a mean Average Precision of 50.6%, making it the more viable path for detecting duplicate submissions. The findings suggest that handwriting similarity-based retrieval, rather than text recognition, is the practical primitive for supporting electoral fraud detection in this domain.

CONCUR: A Framework for Continual Constrained and Unconstrained Routing

Do LLMs Benefit From Their Own Words?

Pitfalls in Evaluating Interpretability Agents

Reaching Beyond the Mode: RL for Distributional Reasoning in Language Models