Research & Education

LLM-Reading-List is a curated collection of academic papers primarily focused on Large Language Model (LLM) inference and model compression. This GitHub repository serves as a valuable resource for researchers and academics looking to stay updated on the latest advancements in these specialized areas of LLM development. The list is meticulously organized into categories such as Transformer Architectures, Foundation Models, Position Encoding, KV Cache, Activation, Pruning, Quantization, Normalization, Sparsity and rank compression, Fine-tuning, Sampling, Scaling, Mixture of Experts, and Watermarking. Each section provides direct links to key papers, making it an efficient way to explore foundational and cutting-edge research without extensive searching. It is particularly useful for those delving into the technical aspects of optimizing LLMs for performance and efficiency.

Anara is an AI-powered workspace designed for scientists, students, and research teams to streamline academic workflows. It enables users to search for academic papers, extract key passages, and organize their research efficiently. The tool facilitates writing with automatic citations, ensuring academic rigor by linking every insight to original sources for easy verification. Anara helps eliminate hallucinations by limiting AI responses to uploaded files and automatically finds relevant citations based on user writing. It can understand and process various file types, including dense textbooks, clinical studies, and technical research papers, providing concise, scholarly answers to questions. Additionally, Anara can generate flashcards and multiple-choice questions from uploaded materials, and allows users to build a single, searchable library for all projects, supporting both solo and collaborative work.

Skriptr is an AI-powered platform designed to enhance the research and writing process for academics and students. It leverages AI agents that collaborate with users, helping to build on insights and streamline the entire workflow from initial research to final document creation. The platform aims to improve writing quality and efficiency by providing tools that assist with various stages of academic and research writing. Skriptr focuses on making the research and writing journey more collaborative and insightful, ensuring users can produce high-quality work with greater ease.

Finetuned Diffusion is an AI image generation tool available on Hugging Face Spaces, utilizing fine-tuned diffusion models to create images. The tool is developed by SUPERSHANKY and uses Gradio for its interface, making it accessible for users to experiment with AI-powered image creation. It operates under an MIT license, indicating its open-source nature and potential for community contributions. However, the current status shows a build error, suggesting it may not be fully functional at this time.

ml_privacy_meter is an open-source Python library designed to audit data privacy in a wide range of statistical and machine learning algorithms, including classification, regression, computer vision, and natural language processing. It helps researchers and practitioners assess the privacy risks associated with their models by enabling data protection impact assessment based on state-of-the-art membership inference attacks. The tool supports various auditing methodologies, such as membership inference, range membership inference, and dataset usage cardinality inference, and can also audit differential privacy lower bounds. It is compatible with diverse datasets like CIFAR10, Purchase, and AG News, and models such as CNN, AlexNet, and GPT-2. Users can extend its functionality to other HuggingFace datasets and transformers, or integrate custom training scripts.

HeyScience was an AI-powered personal assistant designed to support academic pursuits, primarily focusing on research. The platform has since transitioned its services and technology to Thesify.ai. Thesify.ai is positioned as a powerful AI-driven study assistant that helps users research and write essays and assignments. Existing HeyScience users are offered 10 free credits for Thesify.ai with a special promo code, encouraging a smooth transition to the new platform. This evolution indicates a shift towards broader academic writing support beyond just scientific research.

PharosBio is an AI-powered decision-support framework designed to accelerate scientific decisions in life sciences. It unifies experimental data, publications, and internal documents, transforming them into fast, reliable, and fully traceable actionable insights. The platform addresses challenges in drug development such as data silos and incomplete pipeline oversight by connecting insights from early discovery, through preclinical stages, to clinical outcomes. It supports hypothesis generation, target selection, and prioritization, and helps predict translatability of in vivo results. PharosBio emphasizes enterprise security and compliance, adhering to industry best practices for data protection, encryption, role-based access controls, and supporting certifications like GDPR, SOC2, and HIPAA.

Bytefuse is an Artificial Intelligence (AI) Research and Development company within the Alphawave Group, focusing on improving lives through contextual AI. They operate as a rapid prototyping unit, quickly iterating towards bespoke solutions by leveraging unique data and context available to each enterprise. This approach allows them to outperform most general off-the-shelf solutions and maximize insights from datasets. Bytefuse offers solutions in EdTech, such as Maski, an AI teaching enhancement solution for personalized education content, and Transport Solutions like Quebit, a smart traffic system that uses AI to manage traffic patterns. They also provide Enterprise AI Solutions, including computer vision AI for event photography and law enforcement.

QA-deep-learning is a GitHub repository that provides TensorFlow and Theano code specifically designed for insurance question answering (QA) matching. The project incorporates various deep learning models, including Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN), to facilitate effective question and answer matching. This tool is particularly useful for researchers and developers engaged in Natural Language Processing (NLP) projects, offering a foundational codebase for building and experimenting with advanced QA systems in the insurance domain. It serves as a valuable resource for understanding and implementing deep learning approaches to complex text-based matching problems.

RWKV-LM, pronounced RwaKuv, is an innovative RNN architecture designed to achieve Large Language Model (LLM) performance. It uniquely combines the strengths of both RNNs and transformers, allowing it to be directly trained like a GPT transformer while maintaining linear time complexity and constant space usage (without a KV-cache). This results in fast training, infinite context length capabilities, and free sentence embedding. The project is currently at RWKV-7 "Goose" and is a Linux Foundation AI project, making it completely free to use. It offers efficient inference and training, with specific instructions and reference implementations available for various versions, including RWKV-7 and RWKV-6.

Spatial Pixel is an AI research and design studio dedicated to developing a new form of spatial computing, specifically designed for physical spaces rather than individual users. Their core platform, Procession, combines artificial intelligence, natural language understanding, and physical sensors to create intelligent environments. This approach aims to enhance human connections and reduce screen time by seamlessly integrating technology into the physical world. Spatial Pixel offers solutions like a Spatial AI Assistant, Contextual Apps, and tools for programmable spaces, suggesting a focus on creating interactive and responsive environments. They also engage in educational initiatives, such as the SVA Spatial Pixel Lab.

Icare Institute, established in 1991, is an accredited research institute by Innosuisse and a partner of The Ark, focusing on delivering innovative IT solutions. Their expertise spans various domains of artificial intelligence, including computer vision and machine learning, as well as data management areas like big data, data exploration, and time series analysis. They are adept at adapting solutions to hardware constraints and emphasize the implementation of complete, sustainable solutions integrated into production environments to provide tangible added value. Key areas of competence include visual recognition for mobile devices, predictive algorithm creation, and the development of intelligent agents for diverse applications such as data analysis, decision support, and e-energy. They also specialize in mobile and multi-platform software development for smartphones and wearables, and the Internet of Things (IoT), covering object marking, data collection/analysis, and information presentation.

Krutrim Cloud is India's AI-first cloud platform, designed to be industry-leading, future-ready, India-native, and developer-first. It provides a unified AI and Cloud stack, battle-tested at scale, for managing compute, storage, networking, and AI workloads from a single intuitive interface. The platform supports AI workloads like distributed training on GPU clusters, low-latency inference, and model fine-tuning, alongside scalable data platforms and reliable compute for applications. Krutrim emphasizes developer experience with a simple console, powerful SDKs (Python, Node.js, Go, Java, Rust), AWS-compatible APIs, and transparent INR billing. It also offers enterprise-grade security, high availability, and 24/7 support.

The Lamarr Institute for Machine Learning and Artificial Intelligence is a leading international research center dedicated to advancing AI and ML. Researchers focus on creating future AI applications based on data, expert knowledge, and contextual information, emphasizing high-performance, trustworthy, and sustainable solutions. The institute also plays a crucial role in educating the next generation of AI experts through various initiatives, from schools to professional training. Furthermore, it fosters innovation by combining cutting-edge research with practical industry experience, developing powerful AI solutions for business and society. Key research areas include Hybrid ML, Resource-aware ML, Trustworthy AI, Human-centered AI Systems, Embodied AI, and applications in fields like Physics, Life Sciences, and Natural Language Processing.

EleutherAI is a research organization dedicated to the open-source advancement of artificial intelligence, with a strong emphasis on large language models (LLMs). Their research spans critical areas such as language modeling, interpretability, and alignment, aiming to understand how models evolve during training and to elicit latent knowledge. EleutherAI has notably trained and released several powerful open-source LLMs, contributing significantly to the AI community. They also publish research papers on topics like test set contamination and data filtering for tamper-resistant LLMs. The organization fosters community engagement through platforms like Discord, where members can collaborate and discuss ongoing projects.

VoxCPM2 is a cutting-edge, tokenizer-free Text-to-Speech (TTS) system developed by OpenBMB, designed for highly natural and expressive speech synthesis. It bypasses discrete tokenization by directly generating continuous speech representations via an end-to-end diffusion autoregressive architecture. The latest version, VoxCPM2, is a 2B parameter model trained on over 2 million hours of multilingual speech data, supporting 30 languages. Key features include Voice Design, allowing users to create new voices from natural-language descriptions, and Controllable Voice Cloning, which enables cloning a voice from a short reference clip with optional style guidance. It also offers Ultimate Cloning for reproducing every vocal nuance and outputs 48kHz studio-quality audio. VoxCPM2 is fully open-source under the Apache-2.0 license, making it free for commercial use, and supports real-time streaming with low RTF.

VITA is an open-source project focused on achieving GPT-4o level real-time vision and speech interaction. The latest version, VITA-1.5, introduces significant advancements, including a reduction in end-to-end speech interaction latency from approximately 4 seconds to 1.5 seconds, enabling near-instant user experience. It also boasts enhanced multimodal performance, with average scores on benchmarks like MME, MMBench, and MathVista increasing from 59.8 to 70.8. VITA-1.5 refines speech processing capabilities, reducing ASR Word Error Rate from 18.4 to 7.5, and integrates an end-to-end TTS module. The tool supports both English and Chinese and provides instructions for setting up basic and real-time interactive demos, making it accessible for developers and researchers.

yousan.ai is an open-source project dedicated to providing a wealth of resources for individuals studying deep learning. It encompasses a wide array of materials, including practical projects, comprehensive ebooks, and structured courses. The platform is designed to support various popular deep learning frameworks such as Caffe, TensorFlow, PyTorch, MXNet, PaddlePaddle, Darknet, Deeplearning4j, MatConvNet, Keras, Chainer, and CNTK. It covers key deep learning tasks like computer vision (image classification, segmentation, object detection), speech processing (recognition, synthesis), and natural language processing (knowledge graphs, automatic question answering). The project also offers specific resources on topics like GANs, Transformer models, and ChatGPT, making it a valuable hub for deep learning enthusiasts and professionals.

LinkDoc Technology is a data-driven and AI-powered medical technology company focused on precision medicine. It offers comprehensive solutions across the healthcare ecosystem, including the LinkCare platform for patient-centric digital care, an Internet Hospital, and an AI-assisted diagnosis and treatment system. The LinkData platform provides full-course medical data governance and AI decision-making capabilities, encompassing big data processing, natural language processing, and AI modeling. Additionally, LinkSolutions offers real-world data-driven services for clinical research, patient recruitment, data insights, and digital marketing for global life science institutions. The company aims to improve patient care and medical research through advanced data analytics and AI.

UiB AI serves as the central hub for artificial intelligence initiatives at the University of Bergen, fostering collaboration across faculties and with external partners. The platform is dedicated to advancing AI through research, education, and innovation. It provides academic programs including bachelor's degrees in AI and Data Science, a master's program in Machine Learning, and contributes to doctoral-level AI education. UiB AI aims to strengthen the university's position in the field of AI by supporting groundbreaking research and developing future AI professionals.

The University of Maryland, Baltimore County (UMBC) Center for Artificial Intelligence is dedicated to advancing AI research, application, and education. It serves as a hub for faculty members whose research interests span critical areas such as robotics, machine learning, and natural language understanding. The center actively promotes the integration of AI-related courses throughout different academic departments and disciplines, ensuring a comprehensive educational experience. By fostering a collaborative environment, the UMBC AI Center aims to drive innovation and cultivate a new generation of AI professionals and researchers.

Ponder AI is an innovative tool designed to revolutionize video editing by leveraging natural language processing. Users can control and manipulate video content through simple text commands, eliminating the need for complex traditional editing software. This approach aims to make video editing more accessible and intuitive for a wider audience. The platform is currently offering exclusive early access to its beta version, allowing interested users to join a waitlist and stay informed about its official launch and updates. Ponder AI focuses on streamlining the editing workflow, enabling faster and more efficient content creation.

Giotto.ai is a Swiss AI lab dedicated to advancing reasoning technologies beyond conventional machine learning. Their mission is to develop AI that deeply understands context, leading to breakthroughs that transform how machines learn, adapt, and serve humanity. The technology is designed to bridge the gap towards Artificial General Intelligence (AGI), focusing on AI that can adapt, learn, and reason with human-like flexibility and creativity. Giotto.ai's innovative approach combines cutting-edge methodologies, enabling their AI to generalize across tasks, solve problems elegantly, and offer transformative potential for various industries and society. They have achieved notable results on the ARC-AGI benchmark, outperforming major reasoning systems like Grok4 and GPT-5 with a 200M-parameter system and low inference costs.

Junior Associate, also known as JAX, is a secure Canadian legal AI platform developed by Main Brain AI, specifically designed to meet the unique demands of Canadian legal practice. It streamlines how law firms manage information, produce documents, and prepare cases by combining secure, responsible AI with practical tools. The platform features a Documents Hub for drafting, reviewing, and managing legal documents with automated creation and analysis. The Insights Hub provides data-driven insights for stronger strategies and negotiations, while the Operations Hub helps manage clients, matters, and deadlines with smart task management. Additionally, the Litigation Hub offers tools for mock trials, witness preparation, and cross-examination simulations. Junior Associate prioritizes data security by self-hosting its proprietary LLM within a secure environment, ensuring data never leaves controlled servers.