About
Research & Development
Boosting Diffusion LLM Speed Without Losing Quality
TLDR: Free Draft-and-Verification (FreeDave) is a new algorithm that significantly speeds up Diffusion Large Language Models (DLLMs) by up to 2.8x through lossless parallel...
Research & Development
Unlocking Bidirectional Concepts in Large Language Models with AbsTopK Sparse Autoencoders
TLDR: A new research paper introduces AbsTopK Sparse Autoencoders (SAEs) to address a limitation in existing SAEs: their inability to represent bidirectional concepts (e.g.,...
Research & Development
Dynamic Prompt Refinement for Smarter AI Image Creation
TLDR: PromptLoop is a new framework that improves AI image generation by dynamically refining text prompts during the image creation process. Instead of directly...
Research & Development
UrbanGraph: Advanced Microclimate Prediction for Cities
TLDR: UrbanGraph is a novel framework that uses dynamic, heterogeneous graphs informed by physics to predict urban microclimates with greater accuracy and efficiency. It...
Research & Development
Dual-Architecture LLMs: Unpacking Latent Communication for Enhanced Reasoning
TLDR: A new research paper investigates dual-architecture latent reasoning in LLMs, where a Base model communicates with a Coprocessor. The study tested two hypotheses...
Research & Development
How Direct Copying Makes Language Models More Trustworthy
TLDR: CopyPasteLLM is a new framework that significantly reduces hallucinations in large language models (LLMs) used in Retrieval-Augmented Generation (RAG). By training LLMs to...