Neural Networks news and analysis on Web Pulse: 1263 curated articles tracking the latest Neural Networks developments, tools, and research, updated continuously from vetted sources.
Researchers introduced Learnable Rank LoRA (LR-LoRA), a parameter-efficient fine-tuning method that allows the adapter rank to be learned during training rather than fixed, enabling the optimizer to determine the appropr…
The IEEE P3109 draft standard defines a new family of parameterized binary floating-point formats designed for machine learning, allowing efficient value representation in few bits with adjustable width, precision, signe…
A new pseudospectral theory for block-triangular Jacobians in coupled gradient descent proves that transient amplification before convergence can be arbitrarily large even when asymptotic stability is guaranteed, with th…
A systematic study of query, key, and value (QKV) projection variants in Transformers found that sharing the key and value projections (Q-K=V) performs on par with or better than the standard three-projection approach ac…
Researchers introduced SDPG, a self-distilled policy-gradient framework that combines group-relative verifier advantages with normalized standard deviation and full-vocabulary on-policy self-distillation to improve reinf…
Researchers have developed LiftQuant, a framework enabling continuous bit-width control for large language models by using a "lift-then-project" mechanism that approximates low-dimensional weight vectors from a higher-di…
Researchers introduced Stein kernelized molecular dynamics (SKMD), an enhanced sampling method that uses interacting particle dynamics to acquire informative training configurations for active learning of machine learnin…
Researchers have introduced DLLG (Dynamic Logit-Level Gating), a framework that learns token-level fusion weights for combining multiple specialized large language models without requiring token-level labels or expert re…
Steve released Architecture A2, a new standard neural network recipe for modeling guitar and bass amps, on Neural Amp Modeler. The update enables users to train A2 models through TONE3000 or open-source tools, with a pre…
Large language models are made entirely of floating-point numbers called weights, not symbolic code or databases, according to a technical analysis of modern AI systems. The weights, arranged across 80 layers of matrix m…
Researchers introduced Gated DeltaNet-2, a linear attention model that decouples the erase and write operations in recurrent state updates using separate channel-wise gates. The model outperforms Mamba-2, Gated DeltaNet,…
A new scientific paper argues that a mathematical theory of deep learning, termed "learning mechanics," is emerging, drawing parallels to physics by focusing on the dynamics of training and making falsifiable predictions…
DeepSpeed has integrated the Muon Optimizer, a memory-efficient optimizer that uses a single momentum buffer and Newton-Schulz orthogonalization to improve training convergence, particularly for 2D weight matrices in tra…
Scientists have demonstrated that artificial intelligence can predict whether a person is reading a taboo word by analyzing their brain waves, according to a study published in *Psychophysiology*. The research found that…
Arcadia Science released SpectraLoRA, a method that post-trains the 22.5-million-parameter DetaNet equivariant graph neural network using SO(3)-equivariant LoRA to predict Hessians, vibrational frequencies, and full Rama…
A developer released Curatube, an open-source tool that provides a distraction-free interface for YouTube playlists. The platform allows users to import playlists and track their learning progress without YouTube's algor…
Piramidal (YC W24) is hiring senior software engineers for frontend, backend, and infra roles to work on problems at the intersection of foundational AI and neuroscience. The company is seeking candidates with experience…
Manticore Search has introduced early termination for its HNSW-based vector search, a feature that stops distance computations once the result set has converged, reducing unnecessary work. Benchmarks on a 1M-vector datas…
A developer built Vektr, a custom RAG engine implementing HNSW graphs from scratch, achieving recall@10 of 0.984 — meaning 98.4% of queries returned all 10 true nearest neighbors in the top results. The implementation co…
Jacob Lavoie and colleagues introduced rlretina, a reinforcement-learning environment that formalizes epiretinal implant output as a stroke-based rendering task, according to an arXiv paper submitted June 2, 2026. The re…