The Archive

Theoretical proof of strong laws of large numbers for Lipschitz functions with applications to subdifferential convergence.

LKValues: benchmark and fine-tuning resource for aligning LLMs to Sri Lankan cultural values in Sinhala.

SoftReason: differentiable neuro-symbolic architecture for deductive reasoning over latent perceptual facts and knowledge graphs.

Miniature humanoid robots controlled via VR teleoperation and RL for balance, extending full-sized platforms to affordable research setups.

Persian Pixel: large-scale synthetic OCR dataset for Persian language addressing script complexity and data scarcity.

HIPAA-compliant AI-augmented LIMS for clinical biomarker workflow optimization in Fragile X research assays.

Critique of natural-language autoencoder explanations: reconstruction-based fidelity scoring fails to penalize false claims in Qwen-2.5-7B.

PG-KINN: physics-informed Petrov-Galerkin Kolmogorov-Arnold Networks for forward and inverse PDE solving with improved interpretability.

Empirical study of quantum-kernel geometry preservation on IBM hardware under dynamical decoupling and gate twirling.

ARROW: first online stochastic variance reduction algorithm for MMD and CORAL domain adaptation on streaming data.

LLMs extract natural-language abstractions from solution traces to improve problem-solving via retrieval and RL-augmented training on math tasks.

Anthropic launches Economic Index connector for Claude, enabling exploration of AI and work data through conversational interface.

Anthropic outlines research priorities for its Economic Futures Research Fund, focusing on AI's labor and economic impacts.

Paired Sampling for Domain Adaptation (PSDA) applies variance reduction to unsupervised domain adaptation losses via cross-domain observation pairing.

"This is day one for cybersecurity in the age of agents," Hugging Face CEO says.

Vision-language models show consistent performance gaps based on image-vs-question modality order; test-time training method closes gap across benchmarks.

With a camera on every pair, Gogle and Samsung’s AI glasses face the same privacy problems as Meta’s. | Photo: Dominic Preston / The Verge Samsung has given us our first chance to check out its upcoming smart glasses in person, revealing two new designs and the first specs in the process, including an impressive 9-hour battery life. The glasses, developed in collaboration with Google and the eyewear brands Gentle Monster and Warby Parker, are due to launch this fall. Google and Samsung have been teasing the new glasses for some time, and revealed the first two designs at I/O in May. Now there...

A TensorRT engine build can take seconds to many minutes. Large strongly typed models, deep tactic search, and a cold timing cache on a brand-new GPU SKU can... A TensorRT engine build can take seconds to many minutes. Large strongly typed models, deep tactic search, and a cold timing cache on a brand-new GPU SKU can leave developers, end users, or AI agents staring at a frozen terminal with no idea whether to wait, retry, or kill the process. Most NVIDIA TensorRT integrations report nothing during a build or provide no way to abort early. Source

Full-text AI detection on 14k Amazon self-published books (2023–2026) shows AI-heavy titles dominate catalog but underperform in sales.

DEED framework closes sim-to-real gap for humanoid robots via data-efficient post-training and experience learning on Unitree G1-Edu with GR00T N1.6.

Interval and fuzzy physics-augmented neural networks (iPANN, fPANN) quantify uncertainty in hyperelastic constitutive modeling from sparse/noisy data.

As Chinese AI models grow in capability and popularity among US companies, the arguing over what should be done about them has reached a fever pitch.

Substack is giving readers a way to estimate how much of a newsletter was written by AI, signaling a broader shift toward transparency around AI-assisted content.

Web analytics (Aug 2023–Oct 2025) show ChatGPT, Perplexity, Gemini drive measurable referral traffic to academic library resources, particularly theses databases.

PyroDash enables cost-efficient SLM-LLM collaborative inference by training SLM to emit control tokens requesting frozen LLM handoff during generation.

OpenAI will spend the equivalent of Sweden's GDP on infrastructure through 2030.

Condition Dropout (ConD) mitigates RGB-D semantic segmentation degradation when one sensor modality fails via continued training with dropout simulation.

Multi-modal transformer jointly processes time-series, wavelet, and spectral representations for nanopore blockade signal classification in single-molecule sensing.

Graph neural networks trained on finite-volume residuals for 3D thermo-fluid prediction without labeled data.

Decentralized online optimization for strongly geodesically convex losses on Riemannian manifolds.