WISE: Rethinking the Knowledge Memory for Lifelong Model Editing of Large Language Models

WISE addresses the impossible triangle in lifelong model editing of Large Language Models (LLMs) by proposing a dual parametric memory scheme, which separates pretrained knowledge and edited knowledge to enhance reliability, generalization, and locality.

Financial Statement Analysis with Large Language Models

GPT4 outperforms financial analysts in predicting future earnings changes from standardized and anonymous financial statements, even without narrative or industry-specific information.

Thermodynamic Natural Gradient Descent

Natural Gradient Descent (NGD), a second-order method, achieves comparable computational complexity to first-order methods when paired with specific hardware, sidestepping the traditional computational overhead associated with second-order training.

Show HN: Open-source real time data framework for LLM applications

Indexify is an open-source data framework designed for building data-intensive LLM applications, featuring a real-time extraction engine and pre-built extraction adapters for fast, reliable, and precise AI application development.

Low-cost shield ardEEG to measure EEG with Arduino Uno R4 WiFi

The ardEEG device transforms an Arduino Uno R4 WiFi into a brain-computer interface, enabling the measurement of EEG, EMG, and ECG signals across 8 channels, facilitating entry into neuroscience.

[R] Introducing SSAMBA: The Self-Supervised Audio Mamba!

SSAMBA, a self-supervised state-space model, outperforms or matches transformer-based models in audio tasks without relying on attention mechanisms.

Binarize CLIP for Multimodal Applications

Binarizing CLIP for multimodal retrieval and ranking involves converting high-dimensional data into binary vectors, significantly reducing memory usage by 32 times and optimizing performance.

[P] ReproModel: Open Source ML Research Toolbox.

ReproModel is an open-source, no-code toolbox designed to streamline the process of testing and reproducing ML models, addressing the significant time investment required to replicate studies from existing research.

[P] State-of-the-art, open source, Computer Vision models that are not ultra resource intensive?

Leading-edge Computer Vision (CV) models sought for inference on mid-tier GPUs like the A4000, with a preference for models beyond ResNet or YOLO and not necessarily CNN-based.

LANL Achieves Yottabyte-Scale Data Compression in Neutron Transport Equations

Los Alamos National Laboratory researchers have developed a tensor network approach for yottabyte-scale data compression in solving neutron transport equations, showcasing a significant leap in memory efficiency.

[P] Learn to Binarize CLIP (&SigLIP) for Multimodal Retrieval and Ranking

Binarizing CLIP embeddings for text or multimodal search and recommendations reduces storage needs by 32x, maintaining 87-93% fidelity of original fp32 embeddings.

DeepSeek-Prover: Advancing Theorem Proving in LLMs through Large-Scale Synthetic Data

DeepSeek-Prover significantly enhances theorem-proving capabilities in LLMs by generating and training on 8 million synthetic Lean 4 proof statements derived from mathematical competition problems.

Distributed Speculative Inference of Large Language Models

Distributed Speculative Inference (DSI) is introduced as a novel algorithm that outpaces both speculative inference (SI) and traditional autoregressive inference methods for large language models (LLMs), without needing model retraining or architectural changes.

TwoMinutePapers - 50,000,000 Point Bouncy Jelly Simulation!

The 50,000,000 point bouncy jelly simulation showcases elastic bodies, like squishy balls and armadillos, interacting in complex ways, demonstrating advanced computational physics.

Into the Omniverse: SoftServe and Continental Drive Digitalization With OpenUSD and Generative AI

SoftServe and Continental have developed the Industrial Co-Pilot, a virtual agent powered by generative AI, to streamline maintenance workflows in manufacturing, enhancing productivity and reducing downtime.