The Archive

Neural networks exhibit U-shaped learning dynamics with overregularization when exceptions are rare, modeling language acquisition phenomena.

Groupoid-equivariant CNN theory for handling local symmetries on bounded/stratified domains.

Solver-level warmstarting to accelerate neural network verification by reusing solutions across specification variants.

Bayesian optimization acquisition function (JAREX) for multi-objective pharmaceutical process characterization in Quality by Design workflows.

Combining neural operators as structural priors with physics-informed neural networks resolves convergence to incorrect solution basins in PDE solving.

Nonasymptotic error bounds for conformalized quantile regression with covariate shift; minimax limits for sparse ReLU neural networks.

SE(3) neural potential fields enable 6-DoF robotic grasping from RGB images alone, avoiding classical potential field local minima via learned gradients.

G-NAC: unsupervised clustering via graph neural cellular automata achieves 0.7951 ARI across 73 tasks, matching Genie with linear scaling.

Import AI newsletter covers US superintelligence strategy, neuroscience breakthroughs in brain-in-mouse models, and machine hermeneutics with analysis of AI scaling limits.

Stochastic reconfiguration acts as statistical filter in overparameterized neural quantum states, improving finite-sample generalization.

Neural operator network (SDC-GON) for PDE solving using Green's functions with singular decomposition and consistency regularization.

Euston: 8B model trained to reject false mathematical claims instead of proving them, addressing reasoning LLM sycophancy.

U-Net post-processing pipeline for scientific data compression, correcting pixel-space residuals in RVQ-based systems.

FPGA-accelerated quantized RNN achieves sub-10μs real-time plasma state prediction for Tokamak control via hls4ml synthesis on Xilinx Alveo U50.

Vals AI is hoping to make AI benchmarking a more neutral and trustworthy resource in a world increasingly inundated by AI models.

Designer-RSI: agent framework with procedural memory learns reusable design skills from 230+ tools in professional graphics software.

BrainWideBench: benchmark for evaluating transfer learning across animals and brain regions on multi-region neural recordings.

Gricea is an open-science platform for reproducible conversational AI research enabling study artifact sharing and reuse.

Hard-label model extraction attack on ReLU MLPs via efficient sign recovery, extending Eurocrypt 2025 Carlini et al. work.

Cross-expert neural framework for lithium-ion battery RUL prediction and capacity estimation from partial charging data.

Geometric Mean Pooling operator combining sign and magnitude information for hierarchical neural network feature composition without learnable parameters.

Adaptive spatio-temporal neural field using learned tessellations to represent variable-complexity climate patterns across geographic regions.

Neural Cellular Automata transfer learning mechanism injecting pretrained hidden states to improve few-shot generalization and parameter efficiency.

Reusing ASR encoder states as sentence embeddings improves query-conditioned topic localization in transcripts without separate audio encoder.

Matrix AdaGrad: structured adaptive optimization framework for matrix-valued neural network parameters.

Statistical feature pool with LLM-generated signals exceeds neural baselines on TSB-AD-U time-series anomaly detection.

Aetiology-specific dysarthria severity models using contrastive embeddings and phonological pseudo-labelling.

ML method for accelerating plane-wave DFT in materials science by completing neural electronic initialization under PAW formalism.

PRISM-BN introduces 5054-instance benchmark for text-to-parameterized Bayesian Network extraction in neurosymbolic AI.

Riemannian Neural Hamiltonian Flows combine geodesic geometry with invertible generative modeling for improved interpretability in normalizing flows.