Early-Signal Detector

method · first seen 8 Oct 2026

score matching

Also: Hyvärinen score, score-based generative modeling

Estimating the gradient of the log density of a distribution without computing the normalizing constant.

Score matching, introduced by Hyvärinen in 2005, fits unnormalized models by matching the gradient of the log density, avoiding the partition function. It underpins score-based diffusion models. The sampled papers are all arXiv work that uses it as a standard tool: a Princeton paper on training-free change detection with Hyvärinen scores, a causal discovery method under selection bias, convergence analyses of alpha-stable and jump-diffusion generative models, a Stein-operator framework unifying score matching with other estimators, diffusion-free conditional sampling, and a self-supervised regularizer co-authored by Yann LeCun. Nothing in the sources marks the core idea itself as new.

The small weekly rise reflects steady output from a mature diffusion and energy-based modeling literature rather than a new trend. The more specific extensions, such as training-free score detectors and score-based entropy regularization, are the parts worth tracking.

Analyst note (established), 17 Nov 2026.

Mentions per week

0319 May11 Aug10 NovPapersCodeDiscussion

PapersCodeDiscussion

Rank 136 on 14 Nov 2026 · 6 mentions in 28 days · 6 independent groups · papers · first seen 5 weeks ago · acceleration +2.20

Earliest sources

Latest sources

Mentioned together with

diffusion modelsCUSUM change detectiondiffusion smoothingMcDiarmid inequalitycausal discoveryselection biasalpha-stable Levy processStein's methodnoise contrastive estimationconditional sampling