The Archive

IQP quantum circuits for feature engineering in credit default prediction classification.

The company also argued that its on-device models offer consumers more privacy.

RBQE framework uses cross-model agreement to estimate polyp segmentation reliability without ground truth at inference.

IBIB protocol measures enterprise AI system capability by serving route, not model checkpoint, exposing measurement error in benchmarks.

Covariance neural networks tutorial unifies GNN theory with covariance matrices as graphs and pairwise statistical dependencies.

Counterexamples to Rockafellar's sum conjecture on monotone operator theory; pure mathematics with no clear AI application.

Deep learning fault detection for 400Hz aerospace electrical systems inspired by Boeing 787 architecture.

SemiGroup-JEPA extends JEPA world models with physics parameter conditioning for zero-shot dynamics generalization.

Microsoft agreed to a set of safety and privacy principles for AI in schools a week after two major school systems announced a ban on student-facing AI. In a new agreement with the American Federation of Teachers (AFT), the second-largest teachers union in the US, and its New York City affiliate the United Federation of Teachers (UFT), Microsoft committed to ten principles that can be contractually enforced by school districts that adopt them. The terms include pledging not to train AI models on student or educator data, limiting the amount of data Microsoft collects in the first place and di...

Paul Christiano joins OpenAI Foundation Board and Safety & Security Committee, strengthening alignment expertise in governance structure.

JarvisGUI benchmark evaluates multi-device GUI agent coordination across platforms with dynamic task composition.

"We really do earnestly believe AI could kill all humans!"

Data mixing strategy enables reasoning models to perform in-language reasoning across non-English prompts consistently.

ConvMem: training-free, parallelizable framework compresses long context via convolutional memory without RL overhead.

FOM-UL: layer-selective unlearning framework removes memorized content from LLMs while maintaining utility and quantization robustness.

Probing study questions whether self-supervised speech models learn word identity or phonetic form via discriminative abilities.

Deep RL framework trains UAV agents for autonomous wildfire monitoring with stable convergence and effective navigation patterns.

General theoretical framework quantifies stability guarantees from ensembling strategies via covariance operator norms.

Study compares GPT-4 vs. clinician performance on ED revisit quality screening, finding LLMs can identify missed improvement opportunities in healthcare review workflows.

Most one-base changes to the human genome do nothing, but a few are significant.

Fortunate Recall proposes ontology-driven memory lifecycle management for LLMs, classifying facts into behavioral categories with differential decay policies to prevent unbounded memory growth.

Benchmark study compares instruction-driven vs. example-driven policy operationalization for Vision-Language Models in content moderation using ModerationBench.

HybridFLow uses SDN orchestration to optimize client partitioning in federated learning, addressing communication stragglers via network-aware synchronous-asynchronous hybrid scheduling.

Study proposes retrofitting code with exception handling using LLMs guided by Exceptional Behavior Tests, automating try/catch and error-guard generation.

Rosetta system fine-tunes NileChat-3B with LoRA for English-to-Dialectal Arabic dialogue translation, achieving 4th place in AlexandriaX-2026 shared task.

MOONWALK design framework aligns creative intent, evidence, and actions in animation/VFX pre-production review to reduce miscommunication between supervisors and junior artists.

Uses reinforcement learning to search SMEFT parameter space for new physics explanations of particle physics anomalies, addressing operator selection and loop correlations.

PACE framework minimizes Perceived Time-to-First-Response for retrieval-augmented dialogue via cascading routing, filler-response control, and latency-aware semantic caching.

Google Search adds live sports tracking and fantasy recommendations via Gemma, a consumer feature integration.

Google DeepMind and filmmakers used AI to synthesize lost footage in short film 'Love, Rendered,' a generative video application.