Bret Taylor

Reparative Justice as a Framework for Reforming AI Data Work and Dataset Governance

This ethnographic study examines a civic-tech initiative that constructs online safety datasets collaboratively with communities most affected by online harms, using a feminist framework. The authors argue that prevailing data work norms — including safety evaluations and red teaming — fail to address foundational accountability structures, and that meaningful repair requires centering harmed communities rather than datasets or AI systems. Drawing on STS-informed reparative justice theory, they contend that just reward for data workers and collective dataset governance are the core unresolved tensions in responsible AI development.

CAMELS Gen-2: Larger Simulation Volumes Improve Neural Network Cosmological Inference, But Sub-Square-Root Scaling Reveals Information Bottlenecks

The second generation of the CAMELS project introduces 1,192 cosmological simulations varying 35 parameters of the IllustrisTNG model across (50 Mpc/h)³ volumes — eight times larger than the previous generation. These larger volumes improve parameter inference via neural networks (MLPs, CNNs, GNNs, Gaussian processes) by reducing sample variance and capturing more massive halos and diverse environments. However, the gains in constraint tightness scale sub-linearly with volume (worse than √N), suggesting that mode coupling or parameter degeneracies are limiting information extraction, not just data volume. The dataset and ancillary outputs are publicly released.

Einstein Notation as a Clearer Language for Matrix Lie Group Derivatives in Estimation

This paper proposes applying tensor notation and Einstein summation convention — already standard in physics and differential geometry — to the representation and computation of matrix Lie group operations, particularly derivatives used in gradient-based estimation. The contribution is explicitly notational rather than algorithmic: it does not introduce new mathematical capabilities but argues that existing Lie derivative expressions become significantly more perspicuous under this reformulation. The target application is state estimation in robotics, where Lie group methods have already demonstrated accuracy and consistency gains over traditional Euclidean approaches.

Accessible Name Change Policies in CS Publishing Cut Citation Errors and Nearly Eliminate Deadnaming

A mixed-methods study analyzing eight major computer science venues (2019–2025) finds that accessible, visible name change policies measurably reduce citation errors and have driven a 92% reduction in deadnaming of transgender researchers. Despite policy progress—largely pioneered through trans-led advocacy—implementation gaps persist: publisher updates remain incomplete, processing delays stretch months, and misparsed names continue appearing in citations post-update. The research links these citation errors directly to mental health harms including stress, anxiety, and safety risks for affected researchers.

Co-optimizing Diffusion and Tomographic Blur Improves Resolution in Computed Axial Lithography

Computed Axial Lithography (CAL) is a volumetric 3D printing technique that cures resin via superimposed light patterns, but free-radical quenchers like oxygen cause diffusive blurring that degrades resolution at cured/uncured boundaries. This work introduces a framework to extract a single diffusion kernel from any standard uncorrected CAL print by comparing micro-CT data against dose models convolved across a range of diffusivities. Rather than pre-compensating geometry via deconvolution, the authors co-optimize for diffusion-induced blur jointly with the inherent blur of CT reconstruction, achieving superior print fidelity over prior approaches.

OVRO-LWA All-Sky Technosignature Search Sets New EIRP Limits and Opens Path to Axion Dark Matter Detection

Using the OVRO-LWA radio array, researchers conducted an all-sky imaging search for extraterrestrial technosignatures at decametric wavelengths (50–86 MHz) with ~10 Hz frequency resolution, generating over 3 million all-sky images from a single 30-second epoch. No ETI signals were detected, but the survey established sensitivity limits of ~100 Jy per channel, translating to EIRP upper bounds of ~10¹⁴ W at 10 pc and ~10¹⁸ W at 1 kpc. Beyond SETI, the ultra-narrowband, wide-field framework is directly applicable to searches for axion-like particle (ALP) line conversion signals near neutron stars, demonstrating dual scientific utility of the pipeline.

Filling the Requirements Gap: A Structured Mission Engineering Framework for Stakeholder-Sparse Environments

Legacy requirements engineering frameworks break down when customer input is absent or ill-defined — a common condition in rapid military acquisition, space, and infrastructure programs. Fazzini & Herber propose a structured decomposition framework that translates mission intent into traceable Tier 1 and 2 requirements without relying on upfront stakeholder specifications. The method introduces quantitative tools — Best-Worst Scaling for prioritizing critical dimensions and a mission complexity factor capturing technology maturity, evidence confidence, and mission utility — providing a rigorous, adaptive alternative to traditional requirements elicitation. The framework is architected for future integration with UAF and SysML modeling artifacts, grounding it in established systems engineering practice.

Training-Free Aerial-to-Ground Geo-Localization Generalizes Beyond Urban Environments to Natural Terrain

Meridian is a cross-view geo-localization system that matches metric-semantic primitives between overhead aerial imagery and ground robot RGB-D data, enabling accurate global localization in GNSS-denied environments without environment-specific training or fine-tuning. Unlike prior art that struggles with repetitive geometries and featureless natural landscapes, Meridian uses novel consistency metrics and robust pose graph optimization to estimate and refine robot trajectory. The system was validated across structurally diverse environments — urban driving, campus/park, and wilderness — achieving a mean optimized trajectory error of 2.4 m over 19 km of traversal.

A Quantified Two-projection Theorem for Nonlinear Projections

Valorisation of Fermentation Side-Stream for Waste-to-Mycoprotein: Nutrient Composition, Metabolic Insights and Process Optimisation

SF-LIFE: A Large-Scale Simulated Movement Dataset for the San Francisco Bay Area

Do Text Edits Generalize to Visual Generation? Benchmarking Cross-Modal Knowledge Editing in UMMs

Positive Measure of Unions of Variable Surfaces

Reasoning and Planning with Dynamically Changing Norms

EventShiftFlow: Towards Hardware-efficient FPGA-based Flow Estimation

Learning the Error Patterns of Language Models

cloelib: A Flexible Python Library for Computing Cosmological Observables in the Euclid Era

cloelike: A Python Library for Cosmological Likelihood Inference in the Euclid Era

Commissioning of the large-scale lead tungstate scintillating calorimeter

Binding Visual Features Point by Point

Constraints on Dynamical Dark Energy from Multiple Probes in the Full Dark Energy Survey

AVI-HT: Adaptive Vision-IMU Fusion for 3D Hand Tracking

Efficient Agentic Reasoning Through Self-Regulated Simulative Planning

A Generalized Nash Equilibrium-Seeking Scheme for Trauma Resuscitation

Survey Footprint Explorer: A Browser-Based Interactive Tool for Visualizing and Cross-Matching Astronomical Survey Footprints

Massive Galaxies Form Early and Gray: Stellar Assembly and Dust Attenuation at $\mathbf{z>3.5}$ from CAPERS