Research & Education

Agent4SE-Paper-List is a dedicated repository for the paper "Large Language Model-Based Agents for Software Engineering: A Survey." This resource systematically summarizes the advancements of AI agents in Software Engineering (Agent4SE) from both SE task and agent architecture perspectives. It meticulously compiles 106 relevant papers, categorizing them to provide a clear overview of the field. Beyond just listing papers, the repository delves into the open challenges and future directions within this rapidly evolving domain, making it an invaluable resource for researchers and practitioners. The content is continuously updated to reflect the latest developments, ensuring users have access to current research.

cnn_graph is an open-source tool that implements Convolutional Neural Networks (CNNs) on graphs, utilizing fast localized spectral filtering. This repository offers an efficient generalization of traditional CNNs to arbitrary graph structures, making it a valuable resource for researchers and academics working in deep learning and graph theory. The project includes code for reproducing experiments on MNIST and 20NEWS datasets, as well as implementations of filters from other notable graph neural network papers. Users can apply the graph ConvNet to their own data by providing a data matrix, a target vector, and an optional adjacency matrix. The tool is available under the MIT license, encouraging its use and citation in academic work.

AnyModel provides a unified platform to access and compare over 50 leading AI models, such as ChatGPT, Claude, Gemini, Llama, Stable Diffusion, and DALL-E, with a single subscription. Users can send the same prompt to multiple models simultaneously and view the results side-by-side, facilitating comprehensive comparison and analysis. This approach helps users gather diverse AI responses, identify hallucinations, and combine the best elements for superior outcomes. The platform also offers AI-powered insights to pinpoint key points of agreement and consensus across multiple model responses, enhancing accuracy and reducing errors. AnyModel aims to simplify access to advanced AI technology without the need for multiple accounts or API keys, making it easier for users to leverage the collective power of various AI models.

Big Innovation Centre is a leading pioneer in AI, blockchain, digital finance, Web3, and the Metaverse, dedicated to advancing market potential for innovation adoption. It convenes global leaders from business, policy, and research to shape policy and economic transformation in a rapidly evolving digital economy. The Centre serves as the Secretariat for the UK Parliament’s All-Party Parliamentary Group on Artificial Intelligence (APPG AI), influencing UK policy. Through strategic analysis, publications, and events, Big Innovation Centre explores the political economy of artificial intelligence, intellectual property, data, digital infrastructure, and technological sovereignty, supporting decision-makers across various sectors.

WavLM is a novel pre-trained model designed to tackle full-stack downstream speech tasks by learning universal representations from complex speech signals. It jointly learns masked speech prediction and denoising during pre-training, which not only maintains speech content modeling capabilities but also enhances its potential for non-ASR tasks. The model incorporates gated relative position bias within its Transformer structure to better capture sequence ordering in speech. Furthermore, WavLM scales up its training dataset significantly, achieving state-of-the-art performance on benchmarks like SUPERB and delivering substantial improvements across various speech processing applications.

FinePDFs is a research tool developed by HuggingFaceFW, specifically designed to extract and refine data from PDF documents for AI training purposes. It tackles common challenges associated with PDF data, such as format inconsistencies, truncation, and the presence of spam, which can hinder the quality of training datasets. By processing and cleaning this data, FinePDFs aims to unlock a new tier of high-quality tokens, making them more suitable for advanced AI model development. The tool is available as a Hugging Face Space, indicating its accessibility within the ML community for experimentation and use.

SafeAppeals offers an AI-native desktop workspace designed for complex document work, such as legal appeals, dissertations, research papers, and grant applications. It allows users to open all their PDFs, Word documents, and other sources in one place, providing a unified project workspace. The AI assistant remembers the entire project context, including documents, notes, and prior conversations, eliminating the need for re-explaining or copy-pasting. The tool features native document editors for Word, Excel, and PDF files, enabling direct editing, annotation, and citation management. SafeAppeals prioritizes privacy by keeping sensitive documents on the user's machine, ensuring local-first confidentiality. It is available for Windows, macOS, and Linux, and offers flexible AI credit purchasing options or the ability to bring your own API keys for various AI models like Claude, GPT-5, and Gemini.

Qlucore provides visualization-based solutions to accelerate, simplify, and improve research and cancer diagnostics. The platform offers powerful, flexible, and easy-to-use tools for multi-omics data analysis, including bulk RNA-seq and proteomics. It integrates publicly available resources and supports a broad spectrum of omics data types like RNA-seq, NGS, proteomics, and metabolomics. Key features include real-time 2D and 3D interactive plots, powerful statistical analysis with instant visualization, and a Biomarker Workbench optimized for drug development and biomarker discovery. Qlucore also supports machine learning classifiers and offers an NGS Module for genome browser-centric analysis, including fusion gene analysis and single-cell RNA-seq data support. It is available in both cloud and local installations.

Project Flux is a platform dedicated to transforming projects with AI, offering a comprehensive resource for professionals interested in artificial intelligence. It provides the latest news, model updates, practical tips, and information on events related to AI. Users can access expert podcasts and a newsletter to stay informed on advancements and applications of AI in project delivery. The platform also features testimonials from industry professionals, highlighting its value in providing curated information and insights. Project Flux Pro offers an all-access pass to advanced AI resources for those looking to build the future with AI.

OKRA.ai, integrated within Envision Pharma Group, is an AI and technology solution designed for the life sciences sector. It combines AI-powered technology with medical, commercial, and strategic expertise to help organizations achieve smarter, faster outcomes and real-world impact for patients. The tool supports various functions across the product lifecycle, including evidence generation, medical affairs, commercialization, market access, revenue management, and compliance. It aims to accelerate access to life-changing treatments by providing integrated and scalable solutions, enabling better decisions, stronger governance, reduced risk, and accelerated impact for clients.

The SDG AI Lab is a collaborative initiative established in 2019 by UNDP's Bureau for Policy and Programme Support (BPPS) teams and the Sustainable Finance Hub. Hosted by UNDP’s Istanbul International Center for Private Sector in Development (ICPSD) with support from the Government of Türkiye, it operates under UNDP’s Digital, AI and Innovation Hub since 2025. The lab focuses on leveraging Artificial Intelligence to address challenges and advance sustainable development goals, providing research and solutions in this critical area.

AudioLCM is a PyTorch implementation of a latent consistency model designed for efficient and high-quality text-to-audio generation. This open-source tool, presented at ACM-MM'24, allows users to generate audio samples from text prompts. It provides functionalities for both single and batch audio generation, making it suitable for various applications. The repository includes detailed instructions for quick-started inference, model downloading, dataset preparation, and training of variational autoencoders and latent diffusion models. It's a valuable resource for AI researchers and developers exploring advanced audio synthesis techniques.

Deep-pwning is a lightweight, open-source framework designed for experimenting with machine learning models to assess their resilience against motivated adversaries. Built on Tensorflow, it allows users to evaluate the robustness of AI systems, particularly deep neural networks, against adversarial attacks. The framework includes modules for model implementations, adversarial sample generation, and application configurations. It supports the creation of adversarial images for various classifiers, leveraging the phenomenon of 'transferability' in machine learning. Deep-pwning was released at DEF CON 24 and is intended for security researchers and penetration testers to understand and mitigate vulnerabilities in AI applications, especially in critical sectors like medical, transportation, and finance.

Awesome-Controllable-T2I-Diffusion-Models is a comprehensive collection of resources dedicated to controllable generation using text-to-image diffusion models. This GitHub repository specifically highlights methods for controlling these models with novel conditions, drawing from a survey paper titled "Controllable Generation with Text-to-Image Diffusion Models: A Survey." The repository is structured to categorize various approaches, including generation with specific conditions (such as personalization, subject-driven, person-driven, style-driven, interaction-driven, image-driven, distribution-driven, spatial control, advanced text-conditioned, in-context, brain-guided, sound-guided, and text rendering generation) and generation with multiple conditions (like joint training, continual learning, weight fusion, attention-based integration, guidance composition, universal controllable generation, universal conditional score prediction, and universal condition-guided score estimation). It serves as a valuable academic resource for researchers and developers in the field.

Awesome-LLM-Compression is a comprehensive GitHub repository dedicated to collecting and organizing research papers and tools focused on Large Language Model (LLM) compression. This resource is invaluable for researchers and practitioners looking to optimize LLM training and inference processes. The collection covers a wide array of compression techniques, including quantization, pruning and sparsity, distillation, efficient prompting, and KV cache compression. It provides direct links to papers, and in some cases, associated code and tutorials, making it a central hub for staying updated on the latest advancements in efficient LLM deployment and operation.

AIMS5.0 is a collaborative Innovation Action focused on advancing Artificial Intelligence in Manufacturing to achieve Sustainability and Industry5.0. This project aims to bolster European digital sovereignty in sustainable production by integrating AI-enabled hardware and software components throughout the industrial value chain. With a consortium of 53 academic and industry partners, AIMS5.0 intends to boost the economy by developing new AI tools and methods, including innovative chip technology. The initiative seeks to transition manufacturers from Industry4.0 to Industry5.0, fostering human-centric workplace conditions, climate-friendly production, and increased eco-efficiency. It addresses the challenge of decoupling economic growth from resource consumption, striving for higher efficiency and reduced CO2 emissions in European industry.

Llama AI Online offers a web-based platform to interact with Meta Llama 4 Maverick, focusing on practical applications like code review, long document analysis, and image/screenshot interpretation. It provides a workspace where users can upload real-world examples such as code diffs, PDFs, and visual assets to evaluate the model's performance. The tool emphasizes a browser-first evaluation approach, allowing users to assess Llama 4 Maverick's capabilities before committing to local inference or API integrations. It clearly states its independence from Meta, positioning itself as an evaluation environment for Llama-oriented tasks. The platform supports multimodal tasks, enabling users to combine text and image inputs in a single conversation.

Awesome-LLM-Reasoning is a comprehensive, curated collection of papers and resources dedicated to understanding and enhancing the reasoning abilities of Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs). This GitHub repository serves as a valuable academic resource, covering a wide spectrum of topics from foundational Chain-of-Thought prompting techniques to cutting-edge models such as OpenAI o1 and DeepSeek-R1. Researchers and academics can explore surveys, analysis techniques, and specific methodologies aimed at unlocking and improving LLM reasoning. The collection is regularly updated with recent preprints and conference papers from leading venues like ACL, NeurIPS, and ICML, making it an essential tool for staying current with advancements in the field.

AI4Copernicus is a European H2020 project focused on integrating Artificial Intelligence (AI) with Earth Observation (EO) data. The project's primary goal is to establish the AI4EU AI-on-demand platform as the preferred digital environment for users of Copernicus data, including researchers and innovators. Key objectives include integrating resources and large EO data from existing providers, offering access to training materials and expertise, and enriching the AI4EU resources catalogue. AI4Copernicus incentivizes diverse AI4EU and Copernicus communities to address real-world business and societal challenges through a series of four Open Calls, which fund small-scale experiments and larger use-case projects. The initiative also drives the evolution and impact of platforms like AI4EU, WEkEO, and CREODIAS.

WASP-HS (Wallenberg AI, Autonomous Systems and Software Program – Humanity and Society) is a comprehensive research and education program dedicated to exploring the societal and humanistic implications of artificial intelligence and autonomous systems. The program generates research and builds expertise through various initiatives, including funding interdisciplinary collaborations, supporting guest professors, and offering a postdoctoral program. It also provides extensive educational opportunities for PhD students through its graduate school, courses, and study visits. WASP-HS actively disseminates knowledge through publications, news, and events, fostering a deeper understanding of how AI shapes and is shaped by humanity. The program emphasizes capacity building, societal collaboration, and international partnerships, with significant investment from the Wallenberg Foundations.

Deep-transfer-learning is a comprehensive PyTorch library dedicated to deep transfer learning, specifically focusing on single-source and multi-source unsupervised domain adaptation (UDA and MUDA). The library includes implementations of several prominent deep domain adaptation algorithms such as DDC, DAN, Deep Coral, Revgrad, MRAN, and DSAN for UDA, and MFSAN for MUDA. It provides researchers and developers with a robust toolkit to experiment with and apply advanced transfer learning techniques, particularly in scenarios like cross-domain image classification and fraud detection. The repository also includes performance results on standard datasets like Office31 and OfficeHome, making it a valuable resource for benchmarking and research.

CAIR-Nepal, the Center for Artificial Intelligence (AI) Research Nepal, is a leading organization focused on advancing AI research and development. Its mission is to foster innovation, build capacity, and create a sustainable AI ecosystem that benefits society. The center conducts research in areas such as Responsible AI, Human-centered AI, Applied AI, and Digital Healthcare, while also extending its work to other computer science fields like data privacy. CAIR-Nepal emphasizes education through courses and learning materials, and community impact by collaborating with institutions and promoting diversity and inclusion. It aims to become a globally recognized center of excellence in AI research and innovation.

DeepSlide is an open-source, sliding window framework designed for the classification of high-resolution microscopy images, specifically whole-slide images (WSIs) commonly found in histopathology. This tool provides the code for the Nature Scientific Reports paper "Pathologist-level classification of histologic patterns on resected lung adenocarcinoma slides with deep neural networks." It enables users to perform tasks such as train-val-test splitting, patch generation and processing, model training (with recommendations like ResNet-18 for smaller datasets), WSI testing, threshold searching, and visualization of predictions. DeepSlide is particularly useful for researchers and medical professionals working with large histopathology datasets, offering a structured approach to deep learning-based image analysis.

Design Arena is the world's first crowdsourced benchmark for AI-generated design, developed by Arcada Labs. The platform presents creative prompts to leading AI models and displays their results side-by-side. Users can then vote on which design is superior, contributing to leaderboards that indicate which AI models excel in design aesthetics. This process helps the industry understand which AI models truly have 'taste' based on millions of votes from users across 190+ countries. It offers a unique way to assess and compare the creative capabilities of various AI design tools.