absorb.md

Ami Vora

Chronological feed of everything captured from Ami Vora.

Browser, Location, and Jurisdiction Drive Varied Third-Party Tracking Exposure

RegTrack's synchronized crawls of 743 websites from 8 locations using 4 browsers and 2 consent states reveal distinct impacts of browser choice, user location, and hosting jurisdiction on third-party A&T exposure. Privacy browsers reduce A&T domains by up to 30% in permissive regions but less in strict ones like GDPR areas, where pre-consent A&T drops 80% and 89-91% of requests stay within EEA/adequacy countries. Hosting jurisdiction has minimal influence compared to user location.

SCoUT Enables Scalable MARL Communication via Temporal Grouping and Counterfactual Credit Assignment

SCoUT scales communication in partially observed MARL by resampling soft agent groups every K steps using Gumbel-Softmax, inducing differentiable recipient affinities and reducing critic complexity via group-aware value predictions. Agents employ a three-headed policy for actions, send decisions, and recipient selection, with precise credit assignment achieved through analytical counterfactual advantages that isolate each sender's message contribution. At test time, it supports fully decentralized execution by discarding centralized components.

Nb3Sn Superconducting Cavity Achieves >1 GHz Continuous Tuning via Mechanical Separation Without Q Degradation

Researchers demonstrate a "tuning-by-opening" mechanism in a Nb3Sn-coated 9 GHz cigar-shaped microwave cavity, enabling continuous frequency tuning exceeding 1 GHz by mechanically separating the cavity halves. FEM simulations predict that radiative losses do not degrade the intrinsic quality factor of 10^7 for apertures up to 9 mm, corresponding to a 9.0-7.5 GHz range. Experiments with copper spacers and sliding mechanisms validate Q0 exceeding dark matter search requirements across the full range, despite imperfections, offering a clean tuning method for axion haloscopes compatible with REBCO in multi-tesla fields.

Clinically Informed Multimodal Framework Boosts Fast DIR for Proton Therapy

Proposes a coarse-to-fine deformable registration framework using dual CNN encoders and a transformer decoder for longitudinal CT scans in proton therapy. Integrates clinical priors like contours, dose distributions, and planning text via anatomy-guided attention, text modulation, and foreground optimization beyond CT intensities. Evaluated on 1,222 paired scans across multiple anatomies, it outperforms SOTA methods in speed and accuracy for online adaptive workflows.

Efficient Edge-DP Power-Law Exponent Estimation via Privatized Sufficient Statistics

Traditional edge-DP power-law exponent estimation privatizes the full noisy degree distribution, incurring high distortion. This work privatizes only the low-dimensional sufficient statistics required for α estimation, enabling accurate discrete approximation or likelihood-based optimization. Algorithms are developed for both centralized and local edge-DP models, with local DP comparing degree versus log-statistic release, evaluated across graph datasets, privacy budgets, and tail cutoffs.