The Archive

Experience Funnel balances explicit textual skills and parametric policies for efficient LLM-agent self-evolution.

BatchNorm artifacts confound machine unlearning evaluation: single forward pass on retain data reverses forgetting metrics.

Auditable evidence framework improves speech deepfake detection by replacing single scores with explainable decision signals.

Methodological comparison of MAE, Solar, and UniMMQA frameworks tracing evolution of multimodal QA architectures.

Q2D-Web large-scale benchmark for agentic RAG evaluates first-stage retrievers on agent-reformulated queries at production scale.

Statistical mechanics framework for sparse sampling on Hamming slices; applies to Ising models and Bayesian regression but lacks direct AI systems relevance.

OntoKG-EQ: knowledge graph system for reproducible equity market analysis with provenance tracking and competency-question governance; niche domain application.

API benchmark scores for ChatGPT, Claude, Gemini systematically diverge from chatbot interface behavior by 3.4pp average; questions reliability of model eval standards.

Action-conditioned world model for Earth-system emulation enabling what-if interventions; ML accelerator for climate simulation with user-driven state perturbations.

Google is speeding up Chrome’s release schedule to ship security patches and new features faster.

Theoretical analysis of transformer length generalization via C-RASP language; resolves contradictions in prior length-generalization bounds and empirics.

Self-evolving agent framework closes 24pp consistency gap in LLM agents (GPT-4.1 on AppWorld); addresses production reliability of agentic systems.

Closed-form estimator for LLM-judge panel error decomposition under single-common-factor model; addresses contaminated anchor detection in eval.

Ostrich: GPU-accelerated differentiable rigid-body simulator optimizing contact gradients via tape compression; enables gradient-based robotics optimization.

On-Policy Reverse Distillation enables stronger models to surpass weak supervisors via teacher policy shifts; reduces cost of successive model generation post-training.

Rater Ising-Potts model with LLM-derived weights for multi-category scoring reliability; extends classical statistical model for LLM-based evaluation.

Hi-FLoop: hierarchical state-feedback framework for multi-timescale traffic simulation with branch-consistent agent coordination.

AHLERT system automatically generates environment-aware hunt leads from cyber threat intelligence reports for security investigations.

Silent Revision: metric measuring undisclosed changes in frontier AI safety frameworks; corpus of versioned developer commitments.

Adaptive Anisotropic Attention: attention mechanism for structured, low-SNR signals like EEG with electrode and time-axis dependencies.

Machine translation evaluation metrics accounting for terminology variation in English-French scientific translation; cross-term variation diagnostic.

LLM prompt representation effects on type 1 diabetes glycemic-event prediction; zero-shot/few-shot evaluation across open-weight models.

Cybersecurity LLM benchmarks show 80-point score swings from pipeline choices; audit of 8 benchmarks reveals 15 systematic failure modes.

ZK-Trace: zero-knowledge credential system for tracing GNSS classifier leaks in federated monitoring using Tardos fingerprints.

Google DeepMind's AlphaGenome Atlas predicts effects of 9B DNA variants using deep learning, advancing genomic prediction but outside core AI architecture/systems focus.

Google DeepMind has unveiled an AI tool that its scientists claim could help unravel the mysteries of the human genome and transform our understanding of biology, accelerating scientific research and ultimately paving the way for new treatments for diseases. The platform, called AlphaGenome Atlas, contains a "predictive map of every possible DNA letter change in the human genome," the researchers said in a blog post published on Tuesday. DNA is written in an alphabet of four chemical "letters" - usually shortened to A, C, G, and T - and the human genome contains roughly three billion letter p...

Randomly rotated quantization schemes connect MMSE reconstruction scaling to classical CDEF Pythagorean relations in signal processing.

Kairos: fine-grained video-language dataset with time-resolved annotations across 10–30 min videos for temporal dynamics modeling.

OpenAI argues that increased AI capability and lower costs enable new workflows and economic growth for businesses and individuals.