Bayesian inference

topic8 events

papersTODAY 04:00 UTC

Study compares optimisation and Bayesian methods for atmospheric reaction rates

A new arXiv preprint examines how well different optimisation and Bayesian inference techniques constrain reaction rate coefficients in detailed atmospheric chemical mechanisms. The work focuses on autoxidation systems, where many reaction pathways are only observed indirectly through high-resolution mass spectrometry data. The authors assess which methods best handle this indirect, uncertain evidence when estimating rates.

research/bayesian-inference/atmospheric-chemistry/arxiv

papersTODAY 04:00 UTC

arXiv Paper Proposes Bayesian Framework for Inferring Intelligence from Behavior

A new preprint develops a Bayesian account of how intelligence can be inferred from the observable behavior of agents such as language models. The authors treat each prompt as a possibly imperfect internal experiment, and the abstract is truncated before the full results are described. The work sits in the broader area of evaluating model capability without direct access to internal states.

arxiv/research/bayesian/language-models

papersTODAY 04:00 UTC

Differential Privacy of Gaussian Process Posterior Sampling

This paper studies the privacy guarantees of releasing posterior sample paths from a Gaussian process when the entire training set, including covariates and responses, is considered private. Rather than relying on standard differential-privacy mechanisms that add external noise, the analysis focuses on the randomness inherent in posterior sampling itself. It provides a formal treatment of how much privacy such released sample paths preserve.

differential-privacy/gaussian-processes/privacy/research

papersSEP 12 04:00 UTC

Bayesian Framework Unifies Point Set and Image Registration for Scientific Data

A new arXiv paper proposes treating nonrigid registration as a single unified problem, rather than splitting it into point set alignment and continuous intensity field alignment as is conventionally done. The approach, called Domain Elastic Transform, is framed in Bayesian terms and aimed at high-dimensional scientific datasets where existing grid-based and geometry-based methods struggle. It appears as a cross-listed replacement submission on arXiv cs.AI.

research/bayesian/registration/scientific-data

papersSEP 10 04:00 UTC

Settling: Equilibrium Inference for Non-Convex Validity Sets

A new arXiv paper addresses learning systems that must output a single prediction even when the set of acceptable outputs is disconnected or non-convex. The authors show that with squared loss, the Bayes-optimal conditional mean can fall outside the valid output region when the underlying distribution is ambiguous. They propose an equilibrium-based 'settling' procedure that returns a point estimate guaranteed to lie within the validity set.

research/machine-learning/inference/theory

papersSEP 10 04:00 UTC

New arXiv Paper Proposes Bayesian Adversarial Privacy, a Context-Specific Privacy Measure

A revised arXiv research paper introduces a quantitative definition of privacy designed to be tailored to specific settings rather than serving as a uniform guarantee. The framework combines Bayesian reasoning with adversarial modeling to gauge how much sensitive information an attacker could infer. The work positions itself as a unifying contribution amid the wide range of diverging approaches in privacy research.

privacy/adversarial-ml/theory

papersSEP 10 04:00 UTC

Bayes-Optimal Diagonal Regularization in Modal Inverse Problems Follows Closed-Form Power Law

A new machine learning theory paper establishes a 'diagonal saturation principle' for modal inverse problems. When truncation noise is isotropic, the optimal diagonal Tikhonov regularizer takes a closed-form power-law shape whose exponent is fixed entirely by the prior. The authors argue this explains why learned regularization converges on an analytic solution rather than a data-dependent one.

theory/inverse-problems/regularization/spectral-methods

papersSEP 10 04:00 UTC

Bayesian Tracking Guides Deep Spatial Filters to Extract Moving Speakers in Real Time

Deep spatially selective filters deliver high-quality, real-time speech enhancement for stationary speakers whose directions are known, but they struggle when sources move. A new arXiv paper introduces an autoregressive guidance scheme built on Bayesian speaker tracking that lets these filters follow moving speakers from only their initial positions. The approach preserves the efficiency and enhancement quality of the underlying architecture in dynamic scenes.

speech-enhancement/audio/bayesian-tracking/real-time