Research & Education

sumeval is an open-source, multi-language evaluation framework designed for text summarization. It allows users to test and compare various text summarization algorithms with high accuracy. The framework is thoroughly tested, with ROUGE-X scores validated against the original Perl script (ROUGE-1.5.5.pl) and BLEU scores matching the official mteval-v13a.pl script via SacréBLEU. Beyond English, sumeval supports Japanese and Chinese, with an extensible architecture for adding other languages. It is implemented purely in Python and can be used programmatically or via the command line, making it a versatile tool for researchers and developers in natural language processing.

tab-ddpm is the official open-source implementation of the paper "TabDDPM: Modelling Tabular Data with Diffusion Models" presented at ICML 2023. This tool provides researchers and developers with the necessary code to train, sample, and evaluate TabDDPM for generating synthetic tabular data. It includes scripts for hyperparameter tuning, evaluation against baselines like SMOTE and CTGAN, and privacy calculation. The repository also offers pre-tuned hyperparameters for evaluation models and provides access to datasets used in the paper, making it a comprehensive resource for experimentation and development in the field of AI and machine learning, particularly for those working with tabular datasets and diffusion models.

Poetry3D is an innovative artistic visualization project that transforms user-entered poems into unique 3D semantic trees. This tool serves as a practical demonstration of core AI concepts, including tokenization, vector embeddings, vector databases, and cosine similarity. By representing each word as a point in a multi-dimensional space, where words with similar meanings are positioned closely, Poetry3D visually explains how AI processes and understands language. Connections between words form branches, and parallel branches indicate similar phrases, revealing hidden patterns and semantic structures within the poem. The project flattens AI's 1,536-dimensional word space into a 3D visualization, offering a 'shadow of meaning' that is unique to each poem.

SVD_Xtend offers comprehensive training code and extensions for Stable Video Diffusion, allowing users to finetune SVD models for customized video generation. A key feature is tracklet-conditioned video generation, which provides precise control over object movement within videos using bounding box information. The tool supports various video data processing methods, including the use of datasets like BDD100K, and offers detailed training configurations. It also integrates methods from Boximator and TrackDiffusion for enhanced control and instance-level manipulation. SVD_Xtend is ideal for AI researchers and developers looking to experiment with and advance video diffusion models.

Veritas Q-AI is an advanced AI-powered legal document analysis platform that leverages a Quantum Simulation Engine and Quantum Entanglement Logic to analyze complex legal documents. It simultaneously processes information across multiple jurisdictions, including Turkey, the United States, the United Kingdom, and Germany, delivering precision beyond classical AI. The platform offers features like Quantum Probability Mapping for precise risk scoring, Global Jurisdictional Reach with live data streams, Entanglement Analysis to link contracts with high court precedents, and Cross-Border Compliance for international agreements. It helps detect hidden conflicts and risks by simulating legal probabilities in a Multi-Jurisdictional Superposition, ensuring legal security in international operations.

Tarsier is a family of large-scale video-language models developed by ByteDance Research, designed to generate high-quality video descriptions and provide comprehensive video understanding. It utilizes a simple model structure combining CLIP-ViT for frame encoding and an LLM for temporal relationships, along with a meticulously designed two-stage training procedure. Tarsier models have demonstrated superior video description capabilities compared to existing open-source models and are comparable to state-of-the-art proprietary models like GPT-4o and Gemini 1.5 Pro. Beyond description, Tarsier is a versatile generalist model achieving new state-of-the-art results across numerous public benchmarks, including multi-choice VQA, open-ended VQA, and zero-shot video captioning. The project also introduces DREAM-1K, a new challenging benchmark for evaluating video description models, and AutoDQ for interpretable automatic evaluation.

TeaCache, or Timestep Embedding Aware Cache, is an innovative, training-free caching approach designed to significantly accelerate the inference process for various diffusion models. It achieves this by estimating and leveraging the fluctuating differences among model outputs across timesteps. While primarily focused on Video Diffusion Models, TeaCache also demonstrates effectiveness with Image Diffusion Models and Audio Diffusion Models. The project is open-source and available on GitHub, offering support for a wide range of models including Open-Sora, Latte, CogVideoX, and many others. It has been recognized as a highlight in CVPR 2025, underscoring its significance in the field. TeaCache also encourages community contributions and provides instructions for supporting new models, making it a versatile and evolving tool for researchers and developers.

3DTopia-XL is an AI-powered tool designed to streamline the creation of 3D models from 2D images. Users can upload an image, and the application automatically removes the background, generates a corresponding 3D model, and provides both video renders and a GLB file of the final model. The tool offers adjustable settings such as steps, seed, and resolution, allowing for fine-tuning of the generation process to achieve better results. This makes it a valuable asset for anyone looking to quickly convert images into usable 3D assets for various applications.

Minyma Technologies is a research-focused organization dedicated to advancing machine learning and knowledge representation systems, with a particular emphasis on the intricate dynamics of human-AI interactions. Their flagship project, Quino, is designed to facilitate fluid, AI-assisted human learning, suggesting an innovative approach to educational technology and cognitive enhancement. Beyond their internal research, Minyma Technologies also offers partnerships for deep-tech development projects, indicating their expertise in applying advanced technological solutions to complex problems. This positions them as a key player in both academic research and practical deep-tech applications, aiming to decipher and shape the future of AI.

Undermind is an AI-powered research assistant designed to revolutionize scientific research and discovery. It autonomously reads and evaluates hundreds of academic papers, follows citation trails, and delivers precisely relevant insights. The tool helps users describe their research needs, explores the literature by asking follow-up questions, and allows for iterative report building and refinement. Key features include brainstorming research directions with an AI expert, generating custom tables, verifying responses with in-line citations, and gauging paper relevance. Undermind also keeps users updated on new publications in their areas of interest, making it ideal for assessing novelty, scoping complex topics, uncovering cross-disciplinary insights, and identifying literature gaps.

Aveksana is an AI-powered platform designed to support students and researchers throughout the initial stages of their academic work. It specializes in helping users develop compelling research topics, identify critical gaps in existing literature, and formulate robust research proposals. By providing structured assistance, Aveksana aims to increase graduation rates for students and alleviate the workload for supervisors. The tool focuses on streamlining the ideation and proposal development process, making it easier for academics to get their research approved and off the ground efficiently. Its core functionality revolves around guiding users through the often challenging initial phases of academic inquiry.

The Clinical NER Leaderboard is a platform designed for evaluating and comparing Named Entity Recognition (NER) models specifically within the clinical domain. Users can explore existing leaderboards to understand the performance of various models and submit their own models for evaluation against established benchmarks. This tool is hosted on Hugging Face and aims to track progress in clinical Natural Language Processing (NLP), helping researchers and developers identify top-performing models for healthcare applications. It facilitates detailed analysis of model results, contributing to advancements in medical text processing.

Constellation is an AI research tool hosted on Hugging Face Spaces, designed to visualize the evolutionary tree and graph of large language models (LLMs). It aims to help researchers understand the origins, relationships, and development pathways among a vast collection of LLMs, specifically noting its capability to analyze 15,821 different models. While the tool's live website currently indicates a runtime error and scheduling failure, its intended purpose is to provide a comprehensive overview for academic research and analysis within the rapidly evolving field of AI.

The Icelandic Institute for Intelligent Machines (IIIM) is a newly established center dedicated to research in artificial intelligence, robotics, and advanced simulation. IIIM aims to catalyze innovation and high-technology research within Iceland, fostering advancements in these critical fields. The institute actively publishes scientific and technical papers, shares news on AI policy, and engages in featured projects. It also maintains a focus on AI ethics and gender equality, reflecting a commitment to responsible technological development. Through its work, IIIM contributes to the global understanding and application of intelligent machines.

papers is a GitHub repository that curates and summarizes research papers on deep learning. It offers a valuable resource for researchers and practitioners to quickly understand the core concepts of various academic papers without having to read the full text immediately. Each entry in the repository includes a link to the original paper and a review, making it easy to access both the source material and a concise overview. The repository is open-source, encouraging community contributions to expand its collection of summaries. It covers papers from various years, including significant works from 2012 to 2018, across topics like computer vision, neural networks, and natural language processing.

trashnet offers a comprehensive dataset of trash images, categorized into six classes: glass, paper, cardboard, plastic, metal, and general trash. This dataset, comprising 2527 images, was collected using various iPhone models under natural lighting conditions and is available for download via Google Drive. Alongside the dataset, trashnet provides the code for a Torch-based convolutional neural network (CNN) designed for garbage image classification. The CNN, developed as a final project for Stanford's CS 229, has achieved approximately 75% test accuracy. The repository includes installation instructions for Lua and Python dependencies, as well as guidance for setting up CUDA for GPU acceleration, making it a valuable resource for students and researchers in machine learning and environmental studies.

TTRL (Test-Time Reinforcement Learning) is an open-source research project focused on advancing Reinforcement Learning (RL) techniques, particularly for scenarios where ground-truth labels are unavailable during inference. The project investigates how common practices in Test-Time Scaling (TTS), such as majority voting, can generate effective rewards to drive RL training. TTRL has demonstrated significant performance improvements, such as boosting the pass@1 performance of Qwen-2.5-Math-7B by approximately 211% on AIME 2024 using only unlabeled test data. The project provides code and experimental logs, with an implementation based on the 'verl' framework, making it accessible for researchers and developers to reproduce results and further explore test-time reinforcement learning.

TinyZero offers a minimal reproduction of DeepSeek R1-Zero, focusing on reinforcement learning tasks. Built upon the veRL library, this tool allows 3B base Large Language Models (LLMs) to independently develop self-verification and search capabilities. The project provides scripts and instructions for data preparation and training, including configurations for single GPU and multi-GPU setups, and supports instruct ablation experiments. While the repository is no longer actively maintained, it serves as a valuable resource for understanding and replicating the core concepts of DeepSeek R1-Zero, particularly for researchers and developers exploring advanced RL techniques for LLMs.

Video-MME is the first-ever comprehensive evaluation benchmark designed to assess the capabilities of Multi-modal Large Language Models (MLLMs) in video analysis. It covers a wide range of visual domains, temporal durations, and data modalities, including short, medium, and long-term videos (from 11 seconds to 1 hour). The benchmark comprises 900 videos totaling 254 hours and 2,700 human-annotated question-answer pairs. It integrates multi-modal inputs beyond video frames, such as subtitles and audios, to provide a full-spectrum evaluation. Video-MME is suitable for both image MLLMs and video MLLMs, offering a robust framework for evaluating model performance in understanding and processing sequential visual data.

Grand-Challenge.org offers a comprehensive platform for the development, evaluation, and deployment of machine learning solutions specifically tailored for biomedical imaging. It enables users to manage and upload medical imaging data securely, control access, and view data using browser-based workstations. The platform facilitates the training of expert annotators by allowing the creation of question sets for datasets and providing immediate feedback. Users can also gather annotations, customize hanging protocols, and benchmark algorithms for fair assessment. Furthermore, it supports the deployment of algorithms by allowing users to upload container images and manage access for researchers, making it a robust environment for collaborative AI development in the medical field.

Gem is a live AI assistant designed to operate in physical locations such such as buildings, campuses, venues, and hotels. It serves as a virtual concierge, answering visitor questions, providing guidance, and capturing leads. The tool supports over 20 languages, enabling organizations to serve a diverse audience without needing additional bilingual staff. Gem aims to reduce repetitive questions and administrative tasks for staff, freeing them up for more complex duties. It also helps recover missed revenue opportunities by capturing bookings, upgrades, and requests that might otherwise be overlooked. Organizations can start with a single GEM Console and expand its deployment based on proven demand and return on investment.

WMD (Word Mover's Distance) is an open-source implementation of the Word Mover's Distance algorithm, as described in Matthew J Kusner's paper "From Word Embeddings to Document Distances." This tool provides both Python and Matlab code, making it accessible for researchers and practitioners in natural language processing. It allows users to compute the distance matrix between documents based on their word embeddings, offering a robust method for comparing textual content. The repository includes scripts for extracting word vectors, computing WMD, and even a KNN function for classification. It also provides access to datasets used in the original paper, facilitating replication and further research. Prerequisites include Python 2.7 packages like gensim, numpy, and scipy, along with pre-trained word2vec embeddings.

Artificial Intelligence TU/e - WUR - UU - UMC Utrecht is a collaborative initiative between four prominent Dutch universities: Eindhoven University of Technology, Wageningen University & Research, Utrecht University, and University Medical Centre Utrecht. This alliance focuses on advancing AI and data analysis research, with a particular emphasis on developing trustworthy AI solutions that contribute to societal transitions. The platform connects scientists, researchers, and students to foster interdisciplinary collaboration, facilitate research grant applications, and promote knowledge sharing across various fields, including Preventive Health, Circular Society, and Living Technologies.

Read Their Lips is a specialized video processing tool designed to facilitate lip-reading from video content. Users can upload video files and define precise start and end times for the analysis. The platform features an intuitive interface that enables users to accurately frame the subject's face within the video. It also offers a multi-face detection toggle for more complex scenarios, streamlining the lip-reading experience. The service is currently under construction, but once live, it will provide a straightforward way to analyze video for lip-reading purposes, with pricing based on video duration.