Neural Networks

Anthropic researchers published a paper proposing that language models have a 'global workspace' for verbal thinking, introducing the concept of J-space to identify directions in residual space that increase token output…

A developer built a 350-million-parameter Transformer model from scratch using PyTorch, detailing the architecture, training process, and performance benchmarks. The project demonstrates how to implement a large-scale la…

A developer explains the difference between LayerNorm and BatchNorm, showing that LayerNorm normalizes per token rather than per batch, which makes it ideal for Transformers. The post includes an interactive visualizer a…

Researchers have formalized a method called 'conceptors' for steering neural network activations along concept-specific directions using a soft projection operator, as detailed in a new paper. The technique allows for pr…

Researchers at Anthropic have discovered that large language models maintain a privileged set of internal representations, analogous to a global workspace in the human brain, which are available for verbal report and fle…

A new AI tool called Fable – Inner Forward Pass has been released, enabling advanced neural network interpretability through inner forward passes. The tool aims to help researchers understand model decision-making proces…

Researchers introduced Zero-Flow Encoders, a flow-inspired framework for representation learning that uses a zero-flow criterion to certify conditional independence and extract sufficient information from data. The metho…

The MiMo-V2.5 model family introduces Hybrid Sliding Window Attention (Hybrid SWA) to reduce KVCache storage to roughly 1/7 of Full Attention, sparse MoE activation to cut per-token compute, and multimodal encoders for c…

Anthropic's interpretability research reveals that its Claude AI model has self-organized an internal reasoning structure called the 'J-space,' which mirrors the Global Workspace Theory of human consciousness. The J-spac…

Researchers developed a neural completion framework for Islamic geometric patterns that enforces exact rotational symmetry by embedding formal geometric rules into the model. The system guarantees N-fold symmetry and pre…

Researchers introduced the Lifting Neural Operator (LiNO), a multiresolution neural operator built on the second-generation wavelet lifting scheme, to learn solution operators of differential equations. LiNO adaptively d…

Researchers propose MAGE, a geometry-aware framework for view-guided point cloud completion that integrates efficient modality alignment and adaptive geometry enhancement to address cross-modal geometric inconsistency. T…

Researchers found that induction heads in transformers implement two smoothing mechanisms—soft context-matching and BOS token pseudo-counts—that interpolate n-gram estimates, enabling in-context learning to regularize ra…

Researchers developed a two-stage U-Net framework for efficient pedestrian-level wind prediction across urban morphologies, using the UrbanTALES dataset. The first stage predicts wind fields patch-by-patch, while the sec…

Researchers at a university found that selectively applying monocular depth priors from Depth Anything V2 improves sparse-view neural reconstruction in driving scenes, with Splatfacto gaining 1.029 PSNR and reducing RMSE…

Researchers from arXiv found that large language models share a common geometric structure in how context transforms concept representations, with displacement patterns that are semantically organized and transferable ac…

Researchers introduced SecureCROWN, the first framework for privacy-preserving neural network robustness verification using secure two-party computation. It allows a model owner and data owner to jointly compute certifie…

Researchers at an undisclosed institution discovered that safety alignment in large language models can be bypassed by manipulating a single neuron, without training or prompt engineering, across seven models up to 70B p…

Meta's AI research division released Brain2QWERTY in early 2025, a system that decodes typed sentences from non-invasive MEG brain recordings with 61% average word accuracy. The model reads neural signals associated with…

Anthropic researchers discovered a 'global workspace' inside the Claude AI model where its reasoning surfaces, and released the Jacobian Lens tool to visualize it. A developer replicated the technique on a small model in…