Glossary

A-MEM helps AI models remember and use information selectively over long periods by combining attention with a dedicated memory.

A system that makes AI models better at medical image segmentation by letting them skip uncertain predictions, improving accuracy on noisy data.

"Accintent" means an AI accurately understands and follows a user's goal when using tools, and methods like RISE help train AIs to avoid unexpected actions.

Making sure AI agents do exactly what you want them to do, without weird mistakes, often by training them with special synthetic examples and negative feedback.

It's a safety rule for smart helpers that makes them wait to act unless they are very sure they know what you want to do.

It's the final step in an AI hack where the attacker actually gets what they want, like stealing data or making the AI do bad things.

Activation velocity measures how an AI's internal 'thoughts' drift over a long conversation to catch hidden privacy threats.

Active Recap Learning teaches large AI models to create and use summaries of past text sections to better understand really long documents.

Active remasking is a method to improve AI text generators (Diffusion Language Models) by allowing them to refine text holistically, overcoming the limitations of current word-by-word generation.

Everyday self-care tasks that AI systems predict to help people, especially focusing on making sure those predictions are super reliable for safety.

It's a rule that ranks project tasks and their execution methods to help decide the best order to do them in complex schedules.

Actor-Critic PPO is a smart way for AI to learn by having one part decide what to do (actor) and another part judge its choices (critic), using a trick to learn steadily without making big mistakes.

Adam is a popular, fast-converging optimization algorithm for deep learning that adaptively adjusts its step size for each parameter based on gradient moments.

AdamW is a smarter version of the Adam optimizer that helps AI models learn better and generalize more effectively by correctly applying a regularization technique called weight decay.

It's a smart, efficient method for AI to analyze huge networks and make long-term predictions, overcoming the computational limits of older techniques.

It's a smart way to compress complex data in AI systems by preserving its natural structure, making knowledge retrieval more accurate.

It's a smart way to find and fix contradictions in big sets of information by breaking them into smaller parts, making it easier for AI to handle.

Adaptive Dynamic Sampling automatically adjusts how it picks data samples to learn more efficiently by focusing on the most useful information.

It's a method for AI systems to dynamically choose the best-sized AI model for a task, saving money and improving performance by not always using the biggest one.

Adaptive PCA is a smart way to simplify huge amounts of data by finding the most important patterns, making it easier for computers to understand and use.

Adaptive pruning uses an AI agent to intelligently decide which parts of a large language model to remove, making it smaller and faster while keeping its factual knowledge intact.

It's a method where an AI's reward system constantly learns and improves alongside the AI itself, making sure the AI always gets the best guidance during training.

ASCP is a framework for planning multi-step attacks to extract hidden information from AI systems by strategically asking questions that reveal more over time.

AdaReasoner helps AI models learn to use and combine tools intelligently for visual reasoning tasks, adapting to new situations without explicit instructions.

A method that makes it easier and more stable for AI to learn complex, continuous actions by cleverly handling feedback gradients.

ADMET screening checks if a new drug will get into the body, go where it needs to, be processed correctly, leave the body, and not be harmful, all before expensive testing.

ADS helps AI models accurately segment medical images by teaching them to ignore noisy or uncertain labels, making them more robust to imperfect training data.

A technique in AI that helps agents learn better by giving more weight to their most successful past actions, especially when balancing multiple goals like task completion and specific behaviors.

It's a special training method that makes AI models resistant to being tricked or persuaded by malicious inputs.

Adversarial robustness means an AI model can still make correct decisions even when someone tries to subtly mess with its input data.

Agent Code is the programming logic that allows AI agents to plan, remember, and use tools to complete tasks, but its complexity makes it vulnerable to new types of attacks.

Agent Drive is the core system that lets smart AI programs plan, remember, and use tools to complete complicated tasks on their own.

Agent QA uses AI agents that plan, use tools, and remember information to answer complex questions by thinking through them step-by-step.

The Agent Web is the complex system of interconnected steps and tools that advanced AI agents use to do tasks, making them powerful but also vulnerable to attacks that can spread through their entire process.

It's a way to use many smart AI programs to automatically apply complex rules to huge datasets, like social media, to find and understand threats quickly.

A huge multiple-choice test for AI models to see if they can think like a smart driver in various road situations.

AgentForge is an open-source framework that simplifies building complex AI agents using large language models by making them modular and easy to configure.

AgentGC is an AI-powered system that uses smart agents and LLMs to compress genetic data much smaller and faster than before, making it easier to store and share.

Agentic LLMs are smart AI programs that can plan and act on their own, but they're still learning to handle tough real-world problems with strict rules.

It's an AI system that automatically invents hard, solvable problems to teach other AIs how to think better, especially in math and coding.

An AI system that acts like a smart detective, autonomously searching and connecting clues across different levels of information to understand complex data better than simple searches.

AI agents are smart programs that can autonomously perceive, plan, and act to complete tasks, often using tools and learning from their experiences.

It's a way to teach AI agents, particularly those using LLMs, to make better decisions in multi-step tasks by optimizing their learning process at each individual turn.

AgenticRed is an AI that automatically builds and improves other AIs to find weaknesses in big language models, making them safer.

AgentOCR makes AI agents that use big language models much more efficient by converting their long text histories into small, information-rich images.

AI agents are smart programs that act autonomously to complete tasks, and researchers are trying to make them automatically learn and adapt instead of being manually built for every job.

AgentSpeak is a language for programming smart agents that can reason about their beliefs, goals, and plans to act autonomously.

AI refers to smart computer systems that automate tasks and make decisions, used to solve complex problems like detecting online manipulation and securing software agents.

AI research agents are smart computer programs that learn how scientists think to help them discover new ideas and speed up research.

AIME25 is a benchmark that measures how well advanced AI models can solve hard problems in math, coding, and science, showing their true reasoning abilities.

Algorithmic optimization means making smarter computer programs to train new AI models like Diffusion Language Models so they can work much better and reach their full potential.

Algorithms are step-by-step instructions for computers, and current research shows advanced AI models are good at simple algorithmic tasks but struggle with complex, optimized ones like dynamic programming.

AlignXplore+ makes AI personalization better by using natural language to create clear, reusable user preference profiles that work across many different AI models and tasks.

Alli is a marketing AI platform that uses a clever caching system to make its AI agents faster and cheaper by reusing common analytical steps instead of re-generating them.

It's a standard automated test that uses an AI judge to see how good large language models are at following instructions, measuring their performance with a win rate.

Amazon Alexa is a popular voice AI that lets you control smart devices and get information using just your voice, making technology more convenient.

Amazon's product search is changing from traditional lists to AI-generated recommendations, which can hide small businesses, so researchers are finding ways to make sure all products get seen fairly.

ADIR helps AI models handle changing and overlapping tasks by smartly adapting to prevent forgetting, making them more robust in real-world situations like locating sounds.

AHP is a structured way to make tough decisions by organizing choices into a hierarchy and comparing them in pairs to figure out what's most important.

It's finding weird stuff in data that stands out from the normal to catch problems quickly.

Answer Set Programming is a way to program AI using logical rules to solve problems and create clear, understandable explanations for how AI makes decisions.

Anthropic is an AI company that builds powerful and safe AI models like Claude, used by researchers for advanced tasks such as creating user interfaces.

APIs are like universal connectors that let different computer programs, especially AI models and agents, communicate and work together seamlessly.

Apple Silicon is Apple's custom-built computer chip that combines all the main processing parts into one, making their devices super fast and power-efficient.

Apple Siri is a smart assistant built into Apple devices that uses your voice to do things like send messages, play music, and answer questions.

Approximate computing makes hardware more efficient by letting it make small, acceptable errors in calculations, especially useful for AI.

A new robotics planning method that helps robots perform complex, contact-heavy tasks optimally and efficiently by pre-mapping possible movements and then finding the best path.

ARM64 is a highly efficient 64-bit computer chip design used in everything from phones to supercomputers, known for its great performance and low power use.

A system that makes assistive devices safer by only helping when it's very confident about what the user intends to do, preventing errors.

Assistive robots help people by using smart systems that only act when they are very sure of what the person wants, making them safer and more reliable.

It's a method for AI to intelligently find and use information in its memory based on connections, rather than just searching, making AI smarter over time.

AstroReason-Bench is a new test that shows current AI models aren't great at complex, real-world space planning problems with strict rules.

It's a way to train AI models where making one type of mistake is much worse than another, so the model learns to be extra careful about the costly errors.

It's a smart way for AI to combine different pieces of information by paying attention only to what's most important, making predictions better and avoiding confusion.

A technique that helps computers quickly find similar data by cleverly comparing new information against a pre-organized collection, making big AI models run faster and use less memory.

Atlas 2 is a cutting-edge AI model for pathology that uses a huge dataset to achieve top performance, robustness, and efficiency for clinical use.

Attention is a smart focusing mechanism that helps AI models understand complex data by highlighting the most relevant information.

It's a way to make AI models pay attention to all parts of a long document equally, so important information isn't missed in search.

Attention Mechanism helps AI models intelligently focus on the most relevant parts of their input data, improving their ability to understand and process complex information.

Attention sink frames help AI make long, consistent videos but can cause repetitive glitches, which new techniques are fixing for infinite video generation.

Attention tables are efficiently computed attention scores that help shrink big AI models' memory usage so they can run on smaller devices.

It's an advanced type of AI that processes sequences of information by remembering past data and also focusing on the most important bits, making it smarter at understanding patterns over time.

It's a way for AI to show its work, pointing to the exact source text or images that support every statement it generates, making its outputs transparent and reliable.

Attention-MoA is a new way for multiple AI models to collaborate using 'semantic attention' to improve performance, fix errors, and be more efficient, especially for large language models.

It's a way AI writing tools can help you write for your audience without taking away your feeling that you wrote it yourself.

These AI models detect fake voices to secure systems that rely on your voice for access or commands.

AIM is an AI module that generates lifelike talking avatars performing object interactions by intelligently combining audio and interaction information.

A technique that helps AI understand long videos by consistently tracking specific things seen and heard, preventing information from getting lost or fragmented.

Augmented HRM is a set of techniques that help advanced AI models (HRMs) make better "guesses" and avoid getting stuck on reasoning tasks, improving their overall performance.

AutoDriDM is a benchmark that evaluates how well AI models for self-driving cars make decisions, rather than just how accurately they perceive their surroundings.

A tool that automatically observes and scores AI agents' actions to check if they followed all instructions, separate from just completing the task.

AI programs can directly sign and send transactions on blockchains, letting them act autonomously but also creating new security challenges.

AutoRefine helps AI agents learn and remember how to do complex tasks by extracting and maintaining smart 'experience patterns' from their past actions.

Autoregressive models generate sequences one step at a time, using past outputs to predict the next, and are crucial for creating new content like text or video.

It's a method to detect and quantify how much a neural network's training "remembers" its past steps, proving training isn't always a fresh start.

BackdoorAgent is a tool for researchers to study how hidden malicious commands can infect and spread through advanced AI systems that plan, remember, and use tools.

Bayesian principles use probability to update beliefs with new information, helping AI models understand and manage uncertainty for better decisions.

BayesianVLA helps robots follow instructions better by making sure they don't ignore language and instead use it to guide their actions, especially in unfamiliar tasks.

Behavioral science helps design communications and systems that effectively prompt people to take desired actions, especially in critical situations like disaster warnings.

A benchmark is a standardized test that measures and compares how well AI models perform on specific tasks, revealing their capabilities and limitations.

A benchmark dataset is a standard test used to fairly compare how well different AI models perform on a specific job.

Benchmarking is like giving AI models a standardized test to see what they're good at and what they still need to learn.

BiForget helps AI models "forget" specific data they were trained on, efficiently removing its influence for privacy or compliance reasons without rebuilding the whole model.

Big data analytics uses advanced tools to find valuable insights and patterns in huge amounts of information that traditional methods can't handle.

A BiLSTM is a type of AI that reads sequences of data from both ends to better understand the full story and make more accurate predictions.

Black-box optimization finds the best solution for a problem by trying different options and learning from the outcomes, without knowing the problem's inner workings.

BLEU-4 is a common score that tells you how good a computer-generated sentence is by checking how many of its word sequences match human-written examples.

It's a way to use a blockchain to prove data hasn't been tampered with, like a digital notary, without building the whole system on the blockchain.

BluebirdDT is an AI that learns complex decision-making skills from recorded data using a transformer, enabling it to plan for future goals without real-world practice.

BM25 is a classic search algorithm that ranks how relevant documents are to a query, often used in AI systems to help them find facts.

Boolean goal variables help AI systems efficiently manage and achieve the most possible goals, even when some goals conflict, by representing many options compactly.

Boolean retrieval finds documents that perfectly match a query using logical operators, but newer systems are needed for complex, ranked searches.

BoRP is a smart way to automatically and cheaply rate how happy users are with AI chatbots by analyzing the AI's internal data, making it easier to build better conversational AI.

Boundary refinement fine-tunes the edges of detected groups in networks to make sure each item is in the best possible group, improving overall accuracy.

BREPS is a method to find tricky bounding box prompts that break image segmentation AI models, helping researchers make them more robust to real user input.

It's a computer trick that saves previous calculations or data so it doesn't have to do them again, making everything much faster and more efficient.

CaCoVID makes video AI models faster by using AI to figure out and keep only the most important video bits for correct answers, reducing wasted computation.

CADD is an AI that detects fake audio by using both the sound and its context or transcript, making it more accurate and harder to fool than audio-only detectors.

Calibrated probabilities mean an AI's confidence in its prediction matches how often it's actually right, making it safer and more reliable for important tasks.

CaMol is an AI tool that uses chemical knowledge and cause-and-effect logic to better predict molecule properties, even with limited examples.

CVC standardizes how digital identity checks are handled, making delegation of authority secure and interoperable across various online systems for both people and AI.

Capability profiles are detailed descriptions of student knowledge and skill levels used by AI to generate realistic practice materials for teachers.

CARLA is a virtual city environment where self-driving car AI can be safely developed, tested, and trained using realistic simulations.

CASCAL helps pick the best AI model for a question by using agreement among models and knowing their strengths, especially when perfect training data is missing.

Category theory is a math field that uses abstract "objects" and "arrows" to describe patterns and relationships across different areas, helping us understand how things transform and connect.

Causal feature selection picks the most important input variables for a model by figuring out what actually causes an outcome, making the model smarter and more reliable.

Causal Prompt Optimization is a smart way to automatically create the best prompts for AI models by understanding what truly makes a prompt effective for different questions.

A new AI policy that helps agents solve long, complex tasks by automatically breaking them down into a sequence of smaller, logical sub-goals before acting.

Chain-of-Thought Learning makes big AI models smarter at solving hard problems by having them explain their step-by-step thinking process.

Chain-of-thought reasoning makes AI models show their step-by-step thinking to solve hard problems, making them smarter and more accurate.

Chat is used in research to help many experts share diverse information and collectively solve hard scientific problems, especially when combined with tools like prediction markets.

ChatAD-Llama3-8B is a smart AI chatbot that's really good at finding weird patterns in data over time and explaining them, even having conversations about it.

ChatEval is a system that helps researchers understand how different types of authority given to AI agents influence their conversations and decisions.

CircuitJSON is a structured format that helps AI design correct and readable electronic circuits from text, preventing common AI errors.

Claude 3.5 and 3.7 are powerful AI models from Anthropic that excel at using tools reliably in automated business systems.

Claude 3.7 is a proprietary AI model benchmarked for its reliable tool-use capabilities in complex automated systems for enterprises.

Claude Opus 4.5 is a top-tier AI model used as the core intelligence for smart coding agents that can solve hard problems and use tools flexibly.

Claude Sonnet is a smart, fast, and cost-effective AI model from Anthropic that helps businesses do many tasks, like writing and analyzing data, without breaking the bank.

CLEAR-Mamba is an AI system that improves medical image diagnosis by adapting to various image styles and learning better from difficult cases to give more reliable results.

CliConSummation is a way to summarize doctor-patient talks that also points out exactly where each summarized fact came from, making it super reliable for medical use.

Climate RADAR is an AI system that uses data and smart language models to give people personalized, actionable advice during disasters, making early warnings more effective than just sending out alerts.

Clustering is an AI method that automatically sorts data into groups based on similarities, revealing hidden patterns without needing pre-defined labels.

Code-Grounded Vistas (LCV) teaches AI models to override their old knowledge and follow new, contradictory rules by representing those rules as executable code during training.

It's when an AI system can rewrite its own basic programming rules instead of just tweaking its settings or outputs.

Codex is an AI that writes computer code, helping developers build software faster, and its performance can be optimized with specific project instructions.

It's about making AI models, especially new types like Diffusion Language Models, smarter by giving them the ability to think in complex, non-linear ways, like a human solving a puzzle.

The Cognitive Universal Agent (CUA) is a structured AI system that makes LLM reasoning more transparent and trustworthy by separating complex cognitive tasks into distinct, manageable stages.

CogToM is a big, new test for AI models to see if they can understand what others are thinking, much better than older, simpler tests.

A new theory explains how AI models can get better on their own by finding the most consistent and predictable ways to behave, especially useful when there's not much training data.

CoLLM-CC helps multiple AI language models collaborate better on tough, long-term tasks by using a central system to guide their learning and reduce confusion.

CoLLM-DC helps AI language models work together efficiently and independently by using a special learning method, especially for tasks where they get quick feedback.

Algorithms that make computers faster by minimizing data transfers, which is often the slowest part of big computations like those in AI.

COMPAS is a dataset from a criminal risk assessment tool that AI researchers use to find and fix biases in algorithms.

Techniques that shrink AI models to make them faster and fit on small devices, like phones or IoT gadgets, by reducing their size and computational needs.

It's a way to understand and prove that human-AI teams are reliable and trustworthy by looking at their decision-making process, not just their final answers.

ComputePN is an algorithm that lets search engines quickly and efficiently answer really complicated questions that involve deep logic and math.

Conditional Marginal Gain helps attackers strategically plan queries to steal sensitive information from AI systems that use document retrieval by maximizing new data acquired.

CVaR95 measures the average of the worst-case outcomes to help understand and manage extreme risks in AI systems.

This module makes 3D face animations more realistic by giving custom movement instructions to each facial part, rather than a generic one, preventing blur and distortion.

A rule that makes an AI system only take action if it's super confident in its prediction, making it safer and more reliable.

It's a smart way for AI systems to pick the right-sized model for each job, saving money and speeding things up without sacrificing accuracy.

A method for AI to solve problems more reliably by making sure it's confident in each step it takes, preventing errors from piling up.

It's a math method to find the absolute best solution to a problem, but only among the solutions that follow all the given rules or limits.

Containerized execution wraps an application and its dependencies into a self-contained package that runs reliably and consistently across any computer system.

A context graph helps AI models predict molecular properties by organizing chemical knowledge to find causal links, especially useful when data is scarce.

CSF is a tool to check if AI models show different biases when given different contextual information, revealing that simple bias tests aren't enough.

Contextual bandits are algorithms that learn to make the best personalized choices in the moment by considering the current situation and past results, balancing trying new things with sticking to what works.

It's how AI figures out what something means based on the whole conversation or what it already knows, especially important for smart AI agents that remember things over time.

A new benchmark that reveals how AI model bias dramatically changes based on subtle contextual cues, showing that current bias tests might not be enough.

Continual learning teaches AI models to learn new skills sequentially without forgetting old ones, like how humans learn throughout life.

Continuum Memory Architecture lets AI agents continuously learn and update their knowledge over many interactions, unlike current systems that just look up static information.

Contrastive learning helps AI models learn to tell things apart by making similar things look more alike and different things look less alike, improving their ability to handle new situations.

It's a smart check-and-balance system for AI models, making sure they use external information wisely and don't get confused, especially in critical areas like healthcare.

A type of AI model that's really good at finding patterns in images and other grid-like data, making it useful for tasks like recognizing objects or classifying signals.

Cor2Vox generates realistic, anatomically precise synthetic brain MRI scans by using the brain's cortical surface as a guide, solving real-world data scarcity and quality issues.

CORD is a framework that helps audio-based AI models reason more effectively by learning from their own text-based understanding.

CorpusQA is a benchmark that tests how well big AI models can reason by combining information from extremely large sets of documents, using automatically generated questions and answers.

It's a method to pick and use AI models and computing power efficiently to get top performance while staying within a budget.

Counterfactual explanations tell you the smallest changes needed to an input to make an AI model change its mind.

Cross-entropy is a common loss function that tells a machine learning model how wrong its classification predictions are, helping it learn to be more accurate.

A technique that tweaks how AI language models process information at runtime to make them better at less common languages without retraining.

CrossAdapt is a two-stage system that helps big AI models efficiently transfer knowledge between different architectures, reducing retraining costs and improving performance in user prediction systems.

CrossMPT is an AI error corrector that works great in ideal conditions but fails significantly when faced with unexpected noise or attacks, raising concerns about its real-world reliability.

Crowdsourcing is a method where tasks are outsourced to a large, often online, community, commonly used by researchers to find participants for studies, particularly those involving human-AI interaction.

An AI agent that uses a library of human-like computer interaction skills to autonomously operate software, making it more successful and robust at completing tasks.

CUPED helps make online experiments (A/B tests) more precise by using past data to filter out random variations, so you can spot real changes quicker.

Curriculum learning trains AI models by showing them easy data first, then gradually harder data, to improve learning and performance.

Cyber-Physical Systems are smart networks that combine physical things (like sensors on cows) with computer systems to monitor, control, and automate real-world processes.

D-STAR is a smart robot control system that helps humanoids interact better by separating the timing of actions from the actual movements, making them more responsive than just copying.

DAgger is a method that teaches an AI by repeatedly having it try a task, getting corrections from an expert on its mistakes, and then learning from those corrections.

It's a smart way to find the best solution in math problems by taking calculated steps, making sure it doesn't get lost or go too far, especially useful in AI for better training.

DARA is a two-step AI system that uses language models for initial planning and then a precise optimizer to refine ad bidding strategies, especially useful with limited data.

A dataset is a collection of structured information used to teach and test AI models, making them smarter and more capable.

The Davies-Bouldin Index tells you how good your data clusters are by measuring how compact they are and how well-separated they are from each other, with lower scores being better.

DeBERTa-base is a powerful AI model that understands text well, often used as a base for tough language tasks, even when data is tricky.

Decision Trees are flowchart-like models that make predictions by asking a series of yes/no questions about data, making them easy to understand.

It's a way to test AI weather models by seeing if they actually help people make better decisions, rather than just checking if their weather predictions are technically correct.

DDPG is a smart AI algorithm that teaches agents to perform smooth, continuous actions in complex environments, like driving or controlling robots, by learning both what to do and how good that action is.

DeepASMR-DB is a huge dataset of ASMR sounds in English and Chinese used to teach AI how to generate new ASMR in anyone's voice.

DeepBound uses AI to learn how to solve hard math problems (MILPs) more efficiently by intelligently picking the best path, replacing old, unreliable human-made rules.

DeepResearch Bench is a test for AI models that generate complex research papers, helping scientists see which models are best at writing detailed, factual reports.

DeepSeeks Engram is a way to make AI models better at pure reasoning by intentionally removing specific factual knowledge, helping them think more logically and avoid mistakes.

Delegation Grants are secure digital permissions that let you give limited access to others, including AI, without sharing your private login details.

It's a specialized AI model that uses a DenseNet and a U-Net together to accurately draw outlines of anatomical structures in medical scans, like bones.

Dep-Search helps big AI models solve tough problems by explicitly breaking them down, remembering what they've learned, and figuring out the best way to search for answers, rather than just guessing.

A strategy for finding the absolute best decision tree by exhaustively building one branch at a time, but it's slow and doesn't offer good intermediate results.

It's a smart way to make AI models faster and smaller by splitting a complex image processing step into two simpler, more efficient parts.

It's a theory explaining that AI models improve themselves by finding the simplest, most predictable ways to connect inputs to outputs, especially useful for learning with little data.

It's a guide to help developers choose between using advanced AI or simpler computer programs for a task, to avoid wasting resources and improve efficiency.

DETR is a transformer-based AI model that directly finds objects in images without needing complicated intermediate steps, making object detection simpler and more efficient.

DevRev Search is a benchmark for AI customer support that automatically creates data and shows how to efficiently update AI models without costly re-indexing.

DiffCR is a fast and efficient AI-powered image compression technique that makes image files much smaller while keeping them looking good.

It's a smart way to train AI to draw continuous lines and networks in images, making sure they don't break apart, especially for medical and scientific applications.

It's a smart guesser that predicts small, immediate movements to help robots act smoothly and on time, even when things are moving fast.

A training method that teaches AI models to use external tools more wisely by penalizing excessive or unneeded information requests.

Diffusion Language Models are a new way to create text by cleaning up noisy data, allowing for more flexible and parallel generation than current word-by-word methods.

Diffusion LLMs can generate text out of order, but this flexibility currently makes them less good at reasoning by letting them avoid difficult but important parts of a problem.

Diffusion LLMs are AI models that use a step-by-step "denoising" process to create new designs, particularly good at handling complex, interconnected relationships that regular AI models struggle with.

Diffusion models are AI systems that generate new, realistic data by gradually removing noise, useful for creating synthetic data or planning safe actions.

AI models that generate images and videos by denoising, used for tasks like extracting concepts from pictures and creating specific training data to make other AI systems smarter.

A new method helps slow but smart AI models control robots and other systems in real-time by separating their thinking from their quick actions.

DMA lets computer devices talk directly to memory, bypassing the CPU to speed up data transfers and make the system more efficient.

DistilBERT is a smaller, faster version of BERT that uses less computing power, making it great for AI tasks on small devices.

It's an AI component that actively manipulates data distributions to uncover real cause-and-effect links, especially in complex scientific data like molecular structures.

A smart way to pick a small, accurate sample from a huge dataset so that any insights gained from the sample apply to the whole dataset.

Molecular docking is a computer method that predicts how drug molecules attach to proteins, helping scientists design new medicines more efficiently.

A way to make specialized knowledge databases more comprehensive by intelligently adding relevant facts from bigger, general databases.

A math theorem that splits a random process into a predictable part and an unpredictable part, used in AI to figure out the 'advantage' of different decision paths.

DrawSim-PD generates realistic fake student science drawings and explanations to help teachers practice understanding student thinking without using real student data.

DrivoR is an efficient AI system for self-driving cars that uses transformers to process camera data and intelligently plan driving paths, even explaining its choices.

DSXFormer is a new type of AI model that uses advanced attention and pooling to classify super-detailed images more accurately and efficiently.

It's an AI design that takes different kinds of information, processes each separately, and then cleverly merges them to make better decisions.

A smart pathfinding method that lets robots quickly adjust their routes in real-time when things move or change around them.

Dynamic Context Attention helps AI models efficiently understand local patterns in complex data, making them more accurate and faster, especially for detailed image analysis.

Environments where conditions constantly shift, requiring autonomous systems to adapt their behavior in real-time to stay safe and functional.

Dynamic interception tasks are about getting AI to react quickly and accurately to moving targets in real-time environments, which is hard for current slow AI models.

A smart system that predicts future events in real-time by learning and adapting to how data patterns change over time.

DynTS helps big AI models think faster and use less memory by only keeping the crucial information needed for their reasoning process.

E5-base-v2 is a text embedding model that turns text into numerical vectors, enabling AI systems to understand and compare meanings for tasks like advanced search and information retrieval.

EAPO is a special training method that teaches AI models to find and use evidence better when reading really long documents, making their answers more reliable.

ECCT is a powerful AI error corrector that works great in ideal situations but fails when signals are slightly off or tampered with.

EDL is a formal measure of how much predictive knowledge a model gains from training data, helping to understand its learning process and generalization ability.

EDSR is a powerful AI model that makes blurry, small images look sharp and big by using a special, simplified deep neural network.

EEG measures brain's electrical signals from the scalp, allowing researchers to study brain function and develop systems where thoughts control technology.

EHR-RAG helps AI models accurately interpret long patient medical histories by intelligently finding and using relevant information, improving predictions.

An EM-based algorithm is a two-step iterative process used to find the best parameters for statistical models, particularly useful when some data is unobserved, like in complex causal models.

An ELM teaches a big AI language model to read and write directly from the hidden numerical codes (embeddings) that represent information, making those codes understandable and allowing the AI to create new things from them.

eMFD is a lexicon that provides simple word-based clues to help AI models detect human values in written text, even when the values are hard to spot.

An encoder converts raw data into a meaningful numerical representation, helping AI models understand and process information.

These are advanced AI models that intelligently fill in missing visual or text information for other AI systems, making them more robust when inputs are incomplete.

EDVEX is a plan to make sure data creators get paid fairly when their data is used for AI, fixing the current system where most of the money goes to big tech companies.

ErrEval is a system that finds specific errors in automatically generated questions and uses those findings to give a much more accurate quality score.

EFSGD is a smart way to train AI models that generate realistic fake data, making sure private information stays private without sacrificing the data's quality.

These special cameras only record when something changes in their view, making them super fast and efficient for robots in tricky situations.

It's a way to make big AI models better at complex tasks with lots of text by specifically training them to find and use the right information more precisely.

Evolutionary search helps AI find optimal solutions by trying many ideas, learning from their performance, and iteratively improving them, especially useful for training large AI models.

It's a way for multiple AI models to team up and share their smarts to solve problems better and adapt to new challenges.

ExeFuse is a smart way to combine general knowledge with specialized knowledge graphs by treating general facts like tiny computer programs that must prove their relevance and fit.

Experience Replay helps AI models remember old information by replaying past experiences, preventing them from forgetting when learning new tasks.

A dataset meticulously labeled by human experts to provide a highly accurate benchmark for evaluating AI models on challenging, detailed tasks.

Explainable AI helps us understand why an AI made a certain decision, making complex AI systems more transparent and trustworthy.

It's a way to put a hidden digital signature into AI-generated text to prove its origin, but it breaks easily if the text is translated.

A method to watermark AI-generated text that works well normally but breaks when text is translated, requiring more advanced, layered approaches.

FadeMem is a smart memory system for AI agents that actively forgets unimportant information to improve reasoning and reduce storage, inspired by how human memory works.

Fair-Eye Net is an AI tool that uses multiple types of eye scan data to detect glaucoma early, track its progression, and ensure fair diagnoses for everyone.

It's a technique to make AI text-understanding models fairer by making sure all parts of a long document, including later sections or less common languages, are equally represented.

FairGU is a system that helps AI models forget specific data from social networks fairly, preventing biases and protecting sensitive user information during the unlearning process.

FaLW is a smart technique that helps AI models forget specific, often rare, data more effectively to comply with privacy rules.

FastInsight is a new, efficient way for AI to quickly find useful information in large interconnected datasets by cleverly mixing graph structure and meaning.

FC-r is a smart system that dynamically adjusts the complexity of video analysis models to match available computing power, balancing performance and efficiency.

FEATHer is a tiny, efficient AI model for accurate long-term predictions on small devices, especially useful in factories and industrial settings.

Feature extraction turns raw data into useful information that AI models can learn from more effectively.

Feature learning is when computers automatically figure out the important patterns in data instead of humans having to tell them what to look for.

Federated Learning lets many computers work together to build a better AI model by only sharing what they learned, not their private data, keeping everything confidential.

FedKDX helps hospitals build better, private AI models together by sharing what their models learned, not patient data, and focusing on unique "negative" information.

It's a method where an AI system repeatedly refines its decisions by using its own previous outputs as input, making small models perform like big ones efficiently.

FENCE is a smart way to fill in missing data in time-and-location-based information, especially when there's a lot missing, by adaptively guiding its predictions.

Few-shot settings are when AI models learn new tasks using only a tiny amount of example data, often by leveraging knowledge from much larger pre-trained models.

FGM is a technique that makes tiny, specific changes to data to fool AI models, helping researchers understand how easily these models can be tricked.

FINCH is a smart system that merges different data sources, like audio and location, for better predictions by figuring out how reliable each source is for every situation.

A Finite State Machine is a model that describes how a system changes between a fixed number of defined states based on specific triggers.

It's a special type of logic that lets scientists formally ask questions about how AI models, especially neural networks, make decisions by treating them as weighted graphs.

It's a method to figure out which input words most affect what a big AI language model predicts next, especially its overall confidence across all possible words.

It's a math trick that finds a stable answer by repeatedly doing the same calculation until the answer stops changing.

FlashAttention makes big AI models run much faster and process longer inputs by optimizing how they calculate attention on GPUs.

FLASK is a benchmark used to thoroughly test and compare advanced AI language models, especially those with multiple agents, across various complex abilities.

A tool that allows various prediction models to be easily swapped and tested in a simulated business setting to see their actual impact on operations, not just their statistical accuracy.

It's a method to check if a set of equations using computer's decimal numbers has a valid answer, often used to verify software.

Flow Matching is a fast and efficient AI method that directly learns to generate complex data in one or a few steps, outperforming slower multi-step generative models.

It's a part of AI speech systems that helps make voices sound very natural and specific to a person, even for tricky styles like ASMR.

Flow-matching policies are powerful AI controllers for continuous actions that are hard to train, but new methods make their sophisticated capabilities usable.

Fog computing brings computing power closer to devices and users, making applications faster and more efficient by processing data locally instead of sending it all to a distant cloud.

Forest-Chat is an AI tool that uses natural language to help analyze satellite images for changes in forests, like deforestation or growth.

A smart computer model that finds changes in pictures, like satellite images, without needing to be taught every single type of change beforehand.

Big, versatile AI models that can be adapted for many different jobs, making systems like robots smarter, more flexible, and cheaper to operate.

A framework is a structured plan or system used in AI to build, test, or understand models more effectively and reliably.

FDA is a technique that helps diffusion models compress images better and faster by carefully aligning different frequency components of the image data.

FSX helps explain GNN predictions by combining internal data flow analysis with a game-theory approach to pinpoint important graph structures.

FunCineForge is a system that creates better training data and uses an advanced AI model to produce higher quality, more realistic movie dubbing for all kinds of scenes.

A method to detect if an AI model was copied from another by checking its internal 'brain activity' patterns, helping protect intellectual property.

FunHSI is a computer program that makes 3D animated people interact with objects in a scene correctly and realistically, just by telling it what to do.

GAC is a technique that helps AI models learn accurately from messy, noisy data by teaching them to ignore unreliable labels, particularly useful in medical imaging.

GAIA is both a smart system that helps AI agents avoid errors in computer interfaces by learning to check their own moves, and a tough test used to measure how well advanced AI agents perform complex tasks.

GANs are AI models that generate realistic fake data, like images or videos, by having two parts of the model compete to get better at generating and detecting fakes.

GLUs are special neural network layers that use a gate to make big AI models run faster and more efficiently by controlling how information moves through them.

A mathematical model that uses simple linear equations and random noise to understand how things cause each other, especially when some causes are hidden.

Gaussian priors are like giving an AI model a starting guess for its internal settings, assuming those settings are usually centered around a certain value, which helps the model learn better and not get confused by too much detail.

GaussianSVR is a self-supervised AI that builds detailed 3D fetal MRI scans from 2D slices, solving the problem of needing perfect 3D examples for training.

A fast search algorithm that uses a "best guess" to quickly find a solution, but might not find the absolute best one.

GCC-PHAT helps pinpoint sound sources by measuring arrival time differences, and its use in data augmentation improves accuracy in challenging real-world scenarios.

GCFX helps explain complex AI models for graph data by showing how small changes to the input graph would change the model's prediction.

GDPR is a strict EU law that protects people's personal data and privacy, making companies and researchers handle information very carefully, especially in AI and healthcare.

Gemini 2.5 Flash is Google's speedy and cost-effective AI model, balancing powerful capabilities with efficient operation for various applications and research.

Gemini-2.5-Pro is a Google AI model that understands both text and images, showing strong performance in complex tasks but struggling with unclear inputs and tiny visual details.

Gemini-3-pro is a top-tier AI that can deeply understand scientific papers, including pictures, and show exactly where it found its answers, proving real comprehension.

Gemma-3 is Google's open-source AI model, used by researchers to make large language models run more efficiently.

It's a way to figure out the exact reasons an AI agent acts the way it does, whether it succeeds or fails, to make it more transparent and trustworthy.

Generative AI is a type of AI that can create brand new things, like stories or pictures, by learning from lots of examples.

A Generative Application Firewall (GAF) is a central security system that combines all safety measures for AI apps, making them more secure and easier to protect.

A type of loss function that helps generative AI models, like voice cloners, learn to produce high-quality, specific outputs by optimizing key components end-to-end.

AI models that create new images and videos, helping artists and filmmakers quickly visualize and prototype ideas.

A method to teach large AI models specific domain knowledge, making them much better at tasks in specialized fields like medicine or coding.

Genetic Algorithms are like a computer simulation of evolution that finds good solutions to hard problems by breeding better and better answers over time.

GeoDynamics is a smart AI that uses advanced math to map how brain connections shift, giving a truer picture of brain activity and how it supports thought.

GAT is a special AI module that uses geometric data and a Transformer to make incredibly detailed and lifelike digital faces from videos, especially for virtual reality.

GFM4GA is a specialized AI model that uses graph data to find groups of unusual activities, outperforming models that only look for individual anomalies.

GlimpRouter makes big AI models more efficient by using a small model to quickly check if a thinking step is hard, then only sending hard steps to the big model.

An algorithm that summarizes how an entire graph-based AI model makes decisions by showing global 'what if' scenarios to improve understanding.

Go is a fast, concurrent programming language used in many software projects, and it's a key language for testing how well AI models can maintain code.

Google Home is a smart speaker or screen that uses your voice to play music, answer questions, and control other smart gadgets in your house.

It's an unchangeable rulebook for AI systems that automatically enforces consequences to stop them from doing bad things together.

GPT-3.5-Turbo is a popular, powerful AI model from OpenAI used for many text tasks, balancing capability and cost, but it can be tricked by advanced attack methods.

OpenAI's smartest and most current AI model for developers, GPT-4-turbo, can understand huge amounts of text and costs less to use.

GPT-4.1-mini is an AI model used to automatically catch other AI agents doing bad things, like sabotage, proving very effective in tests.

GPT-4o is a super smart AI that can understand and generate content using text, images, and sound, making it useful for many advanced and complex tasks.

GPT-5.1 is a top-tier AI model being tested to see how close its thinking is to human understanding, revealing both advanced capabilities and remaining challenges.

GRACE is either a tool for comparing how robots plan movements or a technique to make big AI vision-language models run efficiently on smaller devices.

Gradient regularization helps AI models learn more effectively and generalize better to new data by making their training updates more stable and robust.

A 2-billion parameter AI model trained to be more reliable by learning to admit when it doesn't know an answer, reducing made-up information.

Graph convolution lets AI models learn from interconnected data by processing information from a node's neighbors, similar to how image filters work but for graphs.

Graph models map relationships, and dynamic graph models are specialized to predict outcomes in networks that are constantly changing and growing.

GNNs are AI models that learn from connected data, like social networks, by having each data point share information with its neighbors to understand complex relationships.

GNNs are AI models that learn from connected data, like social networks or chemical structures, by having each part of the network share information with its neighbors to understand the bigger picture.

A pre-computed map of all possible object movements that helps robots find the most efficient way to manipulate objects by touching them.

Graph retrieval helps AI models find and use specific facts from interconnected knowledge bases to give smarter, more accurate answers.

AI methods that learn from data structured as graphs (like networks of atoms or people) are good for understanding relationships, but sometimes struggle when small changes in the graph lead to big changes in what they predict.

GRanker is a smart search tool that uses graph connections to find better and faster answers in complex data for AI models.

GRIP helps AI models truly forget specific information by directly erasing it from their "expert" components, rather than just superficially redirecting requests.

GroundedInter is a new standard test for AI models that generate virtual characters performing realistic, text-guided interactions with objects.

GRPO is a special training method that uses reinforcement learning to make big AI models better at complex tasks like reasoning and generating specific things, often by learning from groups of information.

GDPO is a new way to train AI models with multiple goals by keeping each reward signal distinct, leading to more stable and accurate learning, especially for language models.

These are clever algorithms that efficiently figure out why groups get certain recommendations by showing what past actions influenced the outcome, balancing explanation quality and cost.

It's a special training trick for big AI models to make them smarter and more accurate at tasks like online ad bidding, even with little data.

AI models with built-in rules to make sure their advice is safe, reliable, and directly tells you what to do, especially in important situations.

GUITestBench is a new test for AI programs that helps them get better at finding bugs in software interfaces automatically.

GUITester is an AI that automatically finds bugs in software interfaces by intelligently separating exploration from defect checking, outperforming previous AI testing methods.

gWorld is an AI model that predicts future mobile app screens by creating renderable code, improving how AI agents interact with phones.

H-SecCoGC helps AI models learn securely and accurately from many devices, even when their internet connection is shaky or some devices are offline.

H.265 is a modern video compression standard that makes video files much smaller without losing quality, so you can watch high-definition movies online or on your phone without buffering, but it needs more powerful devices.

H3 is a system that maps the entire Earth using a grid of hexagons, making it easier to analyze location data at different levels of detail for things like city planning or transportation.

HaC is an AI layer that automatically adjusts its internal settings based on the input data, making the model much better at handling diverse information, especially in complex fields like medical diagnostics.

HalluGuard is a new AI tool that uses advanced math to detect and prevent language models from making up false information, making them more reliable for critical tasks.

HalluJudge helps detect when AI models make up false information in code review comments, making AI-powered code reviews more trustworthy.

Hashing is like assigning a unique, short fingerprint to any data, allowing computers to quickly organize, check, and approximate information.

HATCC is a smart algorithm that turns complicated, cyclic networks into simpler, tree-like ones to perform exact and faster probabilistic calculations.

HateXScore is a tool to evaluate how well AI models explain *why* they identify hate speech, making content moderation more transparent and reliable.

HCVR is a system that uses AI and spatial analysis to design VR environments that perfectly match physical spaces, making 'redirected walking' safer and more immersive.

Health-SCORE helps AI models in healthcare get evaluated and trained more cheaply and effectively by automating the creation of evaluation rules.

HealthBench is a benchmark for testing healthcare AI models to make sure they don't give wrong or unsafe medical advice.

HERMES is a new, efficient, and training-free AI system that helps large language models understand live video streams quickly and accurately with less computer memory.

HeterCSI is a smart way to train AI models for 6G wireless networks to handle diverse signal data efficiently and generalize across many different scenarios.

A smart AI component that learns from complex, multi-part graphs by focusing on important connections, helping other AI systems solve hard problems like route planning.

Hierarchical Federated Learning trains AI models by grouping devices into layers that aggregate data locally, making training more efficient and private, especially with shaky internet.

It's a component of an AI system that reviews an AI's actions to correctly tell if a software bug is the software's fault or the AI tester's fault.

A type of AI model that learns to find and organize simple, explainable concepts within complex data, making big AI models easier to understand.

A way to organize long videos into a detailed, multi-level structure so AI can understand and navigate them more effectively than simple segmenting.

High-frequency actuation means a robot or AI can make many rapid, continuous control adjustments, essential for dealing with fast-moving situations.

HIPAA is a U.S. law that protects your private health information by setting rules for how doctors, hospitals, and health apps handle your medical data.

HMDB51 is a standard collection of videos showing 51 types of human actions, used by AI researchers to train and test models that recognize what people are doing in videos.

HPAC helps navigation AI plan better by grouping actions and understanding how those action groups change the visual world over longer periods.

HRM is a high-performing AI for complex reasoning that sometimes 'guesses' solutions and can fail unexpectedly, prompting research into improving its reliability.

It's a system that makes robots interact smoothly and safely with humans by using cameras to track people and position the robot correctly.

HumanDiffusion is a system that lets drones use camera vision to safely fly around people and help in emergencies, like delivering medical aid, without needing maps.

It's like finding the perfect recipe adjustments (hyperparameters) for a cake (AI model) to make it taste the best, often using smart trial-and-error methods.

It's a way to merge diverse data sources, like medical images and text, into a unified mathematical representation to improve AI prediction accuracy.

ICON is an AI framework that makes finding people using text descriptions much more robust and accurate by teaching models to ignore distractions and focus on essential features.

IFP(SUM) is a powerful logic for formally querying and analyzing machine learning models, particularly neural networks, by treating them as weighted graphs.

Image processing helps computers 'see' and 'understand' pictures by analyzing and manipulating visual data to extract useful information.

IBMDP helps AI make good decisions in complex situations without needing a simulator, by learning from past data and updating its beliefs like a Bayesian system.

Implicit Q-Learning is an AI technique that trains agents to do tasks with specific styles using only existing data, solving the problem of balancing performance and behavior.

It's a technique that helps AI models learn properly when their memory is compressed, preventing them from getting confused by missing information.

Large AI models can learn new tasks on the fly by seeing examples directly in the input, making them super flexible without needing to be retrained.

Large AI models can learn new tasks from examples given in the prompt itself, making them adaptable without needing to be retrained.

It's a smart way for AI to learn from its own past attempts and results to get better at tasks, without needing to be completely re-taught every time.

A systematic method for designing and optimizing rewards in research studies, particularly those involving humans and AI, to ensure accurate and valid results by guiding participant behavior.

IB-GRPO helps big AI models create customized learning plans that are effective and adapt to each student's needs by balancing multiple goals and learning from limited examples.

INFA-Guard is a security system for AI agent networks that stops malicious attacks by identifying and fixing 'infected' agents that have been tricked by attackers, not just the attackers themselves.

Initial Access is the first step in attacking an AI system, usually by tricking it with a special text prompt to gain control.

Insight is an AI model that explains how vision systems make decisions by showing human-understandable concepts directly on the image.

A method to make AI models forget specific data more effectively by individually adjusting how much each data point is 'unlearned', especially for imbalanced datasets.

Institutional AI uses a system of explicit, enforceable rules and consequences, like a digital constitution, to govern how groups of AI agents interact and prevent them from causing harm.

Integrated Assessment Models combine different computer models to help plan long-term climate adaptation strategies for cities and infrastructure, especially under uncertain future conditions.

It's a system component that improves AI recommendations by fixing data loss and better understanding item relationships when converting complex information into simple codes.

InteractAvatar is a system that generates lifelike talking avatars that can interact with objects based on text, by splitting the task into understanding the environment, planning actions, and then creating the video.

It's a way to use computers and interactive visuals to find and understand stories hidden in massive amounts of information, like news or social media.

Intervention Training teaches AI models to self-correct their reasoning steps, improving their ability to solve complex problems by fixing mistakes as they go.

An AI model that acts as a 'critic' to predict if a GUI agent's next action is correct, helping it avoid errors and perform tasks more reliably.

The Invariant Extended Kalman Filter is a smart math tool that helps robots figure out where they are and how they're moving more accurately, even when things go wrong like slipping.

A software tool that simulates how inventory is managed using different forecasts to see their real-world business impact, not just their statistical accuracy.

IoT connects physical devices and sensors to the internet, letting them collect data and communicate to create smart, automated systems.

Isaac Lab is a computer program that lets scientists teach robots how to move and react in difficult situations, like climbing stairs or not falling over, before trying it on real robots.

A system that uses authors' self-rankings to adjust review scores, making best paper selections more accurate and incentivizing honesty in large academic conferences.

The IVLN dataset helps AI learn to navigate realistic 3D places by following written instructions, like a robot finding its way through a house based on a description.

JADE is an AI system that makes the planning and action-taking parts of complex AI models learn and work together perfectly, solving a common problem in advanced AI.

Jenks optimization is a way to sort numbers into the best possible groups by finding natural dividing lines in the data.

Researchers are using ideas from Jungian psychology to make AI models have more realistic, adaptable, and evolving personalities for better human interaction.

JitRL lets big AI models learn and adapt instantly without needing to be retrained, saving a lot of time and money.

Justificatory AI helps AI systems explain their decisions with good reasons, making them more trustworthy and understandable for people.

KAGE-Bench is a special test for AI agents that helps scientists figure out exactly why they struggle when the visual appearance of their environment changes.

Kakugo is a cheap way to build AI models for languages with little data by having a big AI make the training data.

KernelSHAP is a smart way to quickly estimate how important each piece of input data is to an AI model's decision, making complex models understandable.

KGW is an AI text watermarking technique that's effective normally but struggles significantly when text is translated and re-translated, especially in less common languages.

A method that turns noisy continuous data into robust tokens representing motion, helping AI models make better decisions, especially in finance.

A math penalty in AI models that forces hidden data to be organized in a simple, predictable way, often like a bell curve.

It's a way to teach AI models specialized knowledge using text first, so they can work well in areas where speech data is hard to get, like medical conversations.

It's a method to train small AI models to be smart like big ones by teaching them what the big model knows, even what it *doesn't* know, which is great for privacy-sensitive applications like healthcare.

A Knowledge Graph is like a smart, organized database that helps AI models understand complex information and make better decisions by giving them specific facts and rules.

Kolmogorov-Arnold Networks are a new kind of neural network that learns flexible math functions on its connections, making them potentially smarter and more transparent than standard AI.

Koopman operators turn complicated system behaviors into simpler linear ones, helping AI models predict complex dynamics more accurately and generalize better.

Krippendorff's α is a statistical tool that measures how much different people or AI systems agree when they're judging or classifying something, showing if their ratings are reliable.

KRPO is a system that teaches AI models to improve how they pull out facts from text by letting them check their own work and then automatically refine their instructions.

Giant AI models trained on massive datasets that can do many different tasks, like writing or coding, by adapting to new instructions or data.

Using AI chatbots to ask questions and get answers from large amounts of information, especially scientific documents, by turning text into searchable data.

Super smart computer programs that can talk and write like humans, used for many tasks but still learning complex social thinking.

Powerful AI models that reason step-by-step are being made faster and cheaper by dynamically stopping their reasoning when they're confident, based on how certain their outputs are.

LANO helps AI models called neural operators learn from incomplete data by filling in the blanks and reconstructing full solutions, making them useful for real-world scientific problems.

Latent diffusion models generate high-quality images efficiently by performing their core operations in a compact, hidden data space instead of directly on pixels.

Latent thinking is a proposed method for Diffusion Language Models to process information more broadly and iteratively, allowing them to generate text with better overall structure and coherence than current word-by-word AI.

In LLM security, lateral movement is when an attacker uses a compromised AI to spread their attack to other computer systems or user accounts.

Lean is a computer program that lets you write and check mathematical proofs to make sure they are perfectly correct, often with the help of AI.

Learn Before Represent is a two-step process that first teaches large AI models specific domain knowledge, then refines their understanding to excel in specialized fields.

It's a smart way for AI to find the key atoms in a molecule that cause its specific traits, helping predict properties better with less data.

Learning loops enable AI systems to continuously learn and modify their own internal workings and behaviors based on feedback and experience.

LifeAgent is an AI agent that improves digital health assistants by using smart evidence gathering and combining techniques to give better personalized health advice.

LifeAgentBench is a big test to see if AI models can give good, personalized health advice by understanding lots of different health information over a long period.

An AI model that designs specific small molecules to help the body get rid of bad proteins linked to diseases like Alzheimer's.

A linear probe classifier is a quick test to see how much useful information a pre-trained AI model's internal features contain for a new task.

Linear reasoning is the step-by-step way current AI models generate text, which limits their ability to see the big picture and refine their output holistically.

Linear transformers are efficient AI models that can process super long sequences of data much faster than regular transformers by using a smarter, simpler attention mechanism.

LIWC-22 is a program that counts different types of words in text to reveal psychological insights and can help AI understand human values.

LLADBench is a new test for AI models that find anomalies, checking how smart and versatile they are, especially when using large language models.

LLaMA is a popular, powerful family of AI language models from Meta that researchers use to build and test new AI technologies.

Llama 3.1 is a new, advanced AI language model from Meta that's open for everyone to use and research, capable of complex tasks and even showing signs of 'self-awareness.'

LLaMA-2-7B is a free-to-use, 7-billion-parameter AI model from Meta that can understand and generate human-like text, making it a popular choice for building new AI tools.

Llama-3-8B is a large AI model used by researchers to find and fix security weaknesses in AI systems.

LLaMA-3.3 is a very advanced, open-source AI model from Meta that can talk, write code, and reason, making it a powerful tool for many smart computer programs.

LLaMA2 is a popular open-source AI model used by researchers to test and improve how large language models are made smaller, trained better with less data, and run securely.

Llama3 is a big, powerful AI model from Meta that researchers use to build new AI tools, make them run faster, and understand how they behave.

Llama3.1 is an open-source 8-billion parameter AI model used by researchers to improve AI efficiency across languages and simulate human social behavior.

LLaVA is a popular open-source AI model that combines vision and language, used by researchers to make AI smarter and more efficient at understanding images and text.

LLaVA-1.5-13B is a big AI model that combines vision and language, used to improve spatial understanding and make AI processing of images faster.

A method that lets big AI models learn and fix their own mistakes during use by giving them an editable text-based memory, making them act more like a continuously learning system.

Tools that use AI language models to speed up and improve the process of checking if electronic hardware designs are correct and secure.

LLM-AutoDP uses AI to automatically clean and optimize data for training other AI models, making the process faster, cheaper, and more private by avoiding human data access.

Software that uses big AI language models to do smart things, but needs new ways to stay safe from complex digital attacks that act like computer viruses.

An AI system that uses a smart language program to understand what you want and then tells other AI tools how to analyze things like satellite pictures.

It's a way for AI agents to use big language models to smarten up their memory, combining similar thoughts and forgetting what's not important, so they don't get confused or overloaded.

Logic solvers are efficient computer programs for formal reasoning, and when combined with AI models that provide common sense, they can solve much harder logical problems.

Logic Tensor Networks teach AI models to learn from data and follow logical rules at the same time, making them smarter and more reliable.

A mathematical model that predicts the probability of a binary outcome (like yes/no) based on input factors, often used for its simplicity and effectiveness in predicting rare events.

A special loss function that helps AI models truly forget specific data, even when the models are shrunk down for faster use, preventing the forgotten info from accidentally coming back.

AI models designed to read and understand really long documents or conversations, helping them grasp the full picture instead of just small parts by improving their memory and context handling.

A system that helps robots understand their environment and follow instructions by using geometric clues, making them better at doing tasks in new places.

LOOKAT is a method to shrink the memory footprint of large AI models, especially for their attention mechanism, so they can run on small devices without needing a lot of memory speed.

LoRA is a technique that makes fine-tuning large AI models much more efficient by only updating a small fraction of their parameters.

LPCORP is a two-stage AI framework that accurately predicts rare events by first reasoning and then correcting predictions to overcome data imbalance bias.

LTI-Bench is a test for AI agents to see how good their memory systems are at remembering and connecting information for complex tasks.

LTLf synthesis automatically designs systems to meet complex, time-based rules, and a new method can handle many conflicting rules efficiently by finding the best possible compromise.

LUMOS is an AI framework that uses machine learning and physics to efficiently design new fluorescent molecules with specific desired characteristics.

Machine learning teaches computers to learn from data to solve problems and make predictions, rather than being told exactly what to do.

Macro-F1 Score measures a model's accuracy and completeness for each type of output equally, which is great for datasets where some output types are rare.

It's a system that lets many small AI models collaborate and refine their answers over time, making them perform like one giant AI without needing massive computing power.

MAGE-KT is a smart way to predict student learning by building a focused map of knowledge and interactions, avoiding common pitfalls of large data analysis.

It's the first test specifically designed to see if big AI models can recognize and avoid dangerous website links, finding they often can't.

MANGO is an AI tool that helps robots learn better by creating diverse, realistic camera views from simulations, making them more robust to changing perspectives.

Manifold geometry studies how AI models organize information internally, showing that bigger models restructure their 'thinking' in task-specific ways, which predicts how well they learn.

Maniskill is a virtual playground for training robots to do tricky tasks, making it easier and safer to develop advanced robot skills.

MapEval-API is a benchmark that tests if AI models can truly perform geospatial calculations or if they just make up spatial answers.

MapQA is a benchmark that tests how accurately AI models can answer questions requiring real understanding of maps and spatial data.

MapViT is a Vision Transformer AI that helps robots understand their environment and predict radio signal quality in real-time for better navigation.

MarScope is an AI system that lets scientists search and map Martian landforms using everyday language, making planetary exploration much easier and faster.

A math theorem that guarantees certain random processes (martingales) will eventually stop changing and converge to a specific value.

Martingale Foresight Sampling helps AI models think multiple steps ahead using advanced probability theory to find the best reasoning paths, overcoming their usual short-sightedness.

MASBENCH is a benchmark that helps researchers figure out when and why multi-agent AI systems are more effective than single-agent systems by testing them on tasks with controlled characteristics.

A new AI model that generates text by filling in missing parts all at once, making it potentially faster and smarter at learning.

Matrix factorization is a math trick to simplify big data tables into smaller ones, revealing hidden connections and making predictions easier, especially for things like recommending movies.

The MAX-CUT benchmark is a hard graph problem used to test how fast and effectively new computer systems can solve complex optimization tasks.

MaxEnt RL teaches AI agents to be good at a task while also being creative and robust in their actions, and FLAME is a new technique that makes this learning process much better and faster.

A framework that lets AI agents use external tools to give factual, non-hallucinating support, especially in emotional conversations.

MDD helps AI models learn from different kinds of data without needing new labels, and also helps solve complex scheduling puzzles by figuring out when tasks can happen without clashing.

MedMamba is an AI model used as a building block for medical image analysis, making diagnostic tools more robust and accurate, especially for complex conditions.

Megalodon was a giant, extinct shark that was the top predator in ancient oceans, known for its massive teeth and immense size.

Mem0 is a framework that gives AI agents a persistent memory, allowing them to remember past interactions and learn over time, making them much more useful and intelligent.

MemCtrl helps AI agents, especially robots, remember only the important stuff online so they can work better with limited memory and processing power.

Memo-SQL helps computers turn spoken questions into database queries more accurately and efficiently by using smart ways to break down questions and learn from its own past errors.

Memory benchmarking tests a computer's RAM to see how fast it can read, write, and access data, helping improve overall system speed.

These AI systems use external memory and smart planning to help language models understand and reason about massive amounts of text, far beyond what they can hold in their immediate memory.

Memory-V2V adds a memory to AI video editors to ensure videos stay consistent across multiple editing steps, making iterative video creation much smoother.

A merge operator is a method to combine multiple AI models into one, saving training costs and improving efficiency, with new tools helping pick the best way to merge them.

A method that uses evolutionary principles to automatically design and refine AI agents, making them more effective and less biased than human-designed ones.

MetaboNet is a big, organized dataset of Type 1 Diabetes patient info, helping researchers build smarter AI tools to manage the disease.

MG-Data-240K is a new dataset that helps AI models understand and pinpoint objects in complex multi-image scenarios, improving their visual reasoning abilities.

MGU is a smart way to make AI models forget specific data in graph networks, especially important for privacy and accuracy in web applications.

MHub.ai is an open-source platform that puts medical AI models into standard containers to make them easier to use, share, and verify for doctors and researchers.

Tiny, on-demand AI writing suggestions that help improve text but need personalization to keep writers feeling like the true authors.

Microsoft-GraphRAG is an AI system that builds knowledge graphs for advanced reasoning, but it's vulnerable to attacks that can secretly extract its internal graph data.

MIMIC-CXR is a large dataset of chest X-rays and their reports, used by AI researchers to build and test smart systems that can understand medical images and explain their findings.

Min-Seek helps big AI models reason more accurately and stably for longer, even beyond their normal limits, by efficiently managing their memory.

A lightweight AI model that helps robots precisely locate themselves in complex farm environments using simple laser sensors.

mIoU is a common score that tells you how well an AI model can draw accurate outlines around different objects in an image.

MIRACLE is an AI tool that helps doctors predict surgery risks for lung cancer patients by analyzing their medical data and offering understandable advice.

Mistral-7B is a 7-billion parameter AI model used by researchers as a versatile foundation for tasks like building smart chatbots and improving complex reasoning.

MixDPO trains AI models to better understand videos by learning from both text and visual examples, reducing errors like making up details about actions or timing.

Mixture-of-Experts models use a smart "router" to send different parts of a problem to specialized "experts," making huge AI models efficient and powerful.

Mixture-of-Models dynamically combines several smaller AI models at runtime to achieve the performance of much larger models more efficiently.

MJD lexica are specific word lists that help AI models better detect human values and moral cues in text by providing clear, lightweight signals.

It's a way to make computer programs that solve tricky logic puzzles much faster by teaching them to make smarter decisions using AI.

A Multi-Layer Perceptron is a basic but powerful type of neural network that learns complex relationships in data by passing it through several processing layers.

MMLU is a tough test for big AI models, checking if they truly understand and can reason across many school subjects and job fields, not just memorize facts.

It's a smart way to teach AI models to understand speech using very little speech data, especially after they've already learned a lot from text.

The Model Context Protocol helps AI models intelligently find and use important background information to give smarter, more relevant answers.

MoE routing helps huge AI models run efficiently by letting them use only the specific parts they need for each task, saving energy and cost.

MolecularIQ is a specialized test for AI models to precisely check their ability to understand and reason about molecule structures, helping improve AI in chemistry.

MDNF helps explain why complex AI models make certain decisions by turning their internal logic into easy-to-understand rules using simple concepts.

Monotonicity is a property used in algorithms to simplify complex problems by compactly representing many options, leading to much faster solutions.

Monte Carlo methods use lots of random guesses to get a good estimate for problems that are too hard to figure out precisely.

A technique that uses randomness during a neural network's prediction phase to estimate how confident the network is in its answers.

MCTS is a smart search method that uses simulations to find the best decision in complex situations, like playing chess or driving a car.

MonuMAI is a real-world architectural style recognition system used as a tough testing ground for AI to learn how to reliably detect weird or unexpected inputs.

MT-Bench is a benchmark that uses a strong AI to evaluate how well other AI models perform in multi-turn conversations.

It's a smart system that maps out how different knowledge topics connect, helping AI predict student learning better and more efficiently.

It's a framework where multiple AI agents learn to make the best choices from many options, often needing to balance individual goals with fairness for everyone involved.

MARL teaches multiple AI agents to collaborate and learn together, making them better at complex tasks like coordinating LLMs, especially in decentralized setups.

A multi-agent system is a group of smart computer programs or robots that collaborate to tackle big, complicated tasks more effectively and efficiently.

MATTRL helps teams of AI agents make better decisions by letting them learn from new information and discuss problems together right when they need to solve them.

It's an AI system that uses a smart vision-language model to judge other AI's work on complex visual tasks across several quality measures.

A technique to train AI agents to solve complex problems by letting them make decisions at both high-level strategic and fine-grained operational levels, helping generate better training data for other AIs.

It's a fancy statistical method to check if a scientific model works the same way for different groups of people, like comparing how stress affects students in different countries.

A mechanism in AI models that lets them look at different parts of information at the same time to understand complex relationships better.

A part of AI models that uses multiple parallel attention mechanisms to understand complex data and generate sequences more effectively.

A smart way for AI models to handle massive amounts of data efficiently by using multiple attention mechanisms to focus on a small, fixed set of learned summaries instead of the entire input.

It's a way to make smart training data for medical AI so it can better understand and use complex health information from many sources.

A sophisticated AI method that merges multiple images into a single, high-quality picture, overcoming challenges in consistency and realism.

It's like having multiple independent groups trying to solve a puzzle, but they occasionally swap their best pieces to help each other find the overall best solution.

Multi-layer perceptrons are basic neural networks that learn by passing data through multiple layers of connected processing units, enabling them to identify complex patterns and make decisions.

A new AI system that uses advanced vision and language models to understand and describe changes in satellite images of forests, letting users ask questions naturally.

MCI is an AI system that helps analyze changes in satellite pictures, like deforestation, by letting you ask questions in plain language and getting detailed answers.

It's a smart way to mix different kinds of information, like audio and video, by paying more attention to the most important parts, making AI models better at tasks like creating emotional talking faces.

AI models that combine language with other data like images to understand and interact with the world more comprehensively, used in areas like automated software testing.

A multi-modal neural network is an AI that learns from different kinds of data at once, like pictures and text, to understand things better and make more accurate decisions.

These algorithms use evolution-inspired rules to find a range of good solutions when you have multiple conflicting goals, helping you pick the best compromise.

Multi-Persona Thinking makes big AI models less biased by having them think from different perspectives to find and fix their own unfair assumptions.

It's a way for AI to understand text by looking at it from many different zoomed-in and zoomed-out perspectives at once, especially useful for new types of text-generating AIs.

It's about continuously analyzing live data from many sensors that keeps changing, to predict things like when a machine might break down.

Multi-Session Chat is when an AI needs to remember and use information from past conversations over a long time, requiring smart memory systems to avoid forgetting or getting overwhelmed.

A technique that smooths graph data features across multiple layers to create better representations for AI models, especially in graph contrastive learning.

Multilingual BERT is a powerful AI model that understands and processes text in over 100 languages, especially good at handling mixed-language content.

AI that understands information from multiple types of data (like text, images, and sound) to get a more complete picture of the world.

AI models that combine the power of large language models with the ability to understand and generate content across different types of data like text, images, and audio.

A system that predicts a user's next move using various data, but only acts when it's very sure, making assistive tech safer.

Multiple Instance Learning helps computers learn from groups of data where only the group has a label, often by finding the few important items within the group that determine its label.

Multitask learning teaches one AI model to do many related things at the same time, making it smarter and more efficient than separate models.

MultiVis-Agent is an AI system that reliably generates complex data visualizations from diverse inputs by using smart rules to guide its reasoning.

MultiVis-Bench is a large test set for evaluating how well AI systems can generate complex data visualizations using multiple types of input like text, images, and code.

The Myers-Briggs Type Indicator is a personality test used to categorize people into 16 types, and it's now being used by researchers to see if AI can develop consistent human-like personalities.

NSED is a smart way to make many small AI models work together dynamically to achieve the performance of a single giant AI model, but more efficiently.

A smart way to train AI models faster and more stably by understanding the 'shape' of the problem's parameter space, rather than just the simple gradient.

NLI is about teaching computers to figure out if one statement logically supports, contradicts, or is irrelevant to another statement.

NEON SIMD is an ARM processor feature that speeds up tasks by doing the same thing to multiple data items simultaneously, making devices like phones faster for multimedia and AI.

Neural Nets are computer systems that learn from data by mimicking brain cells to find patterns and make smart decisions.

Neural networks are AI models that learn from data, mimicking brain structure to recognize patterns and make decisions, powering most advanced AI applications today.

NeurIPS is a huge yearly conference where AI researchers present their latest work, setting trends for the entire field.

A method that blends neural networks with symbolic logic to make AI models more robust and accurate by fixing how they understand spatial and semantic relationships.

NeuroFilter is a fast and robust privacy guardrail for AI language models that detects privacy violations by looking at internal AI thought patterns, not just the words used.

NeuroVLA is a brain-inspired AI system that makes robots move more stably, react quicker, and use less energy by mimicking biological nervous system structures.

Newmark's CSI framework is a tool to spot and understand culture-specific details in language, helping evaluate how well AI handles different cultures.

Next Key Point (NKP) helps predict long-term movements accurately by using a high-level goal to guide the prediction, preventing unrealistic paths.

Nishpaksh is a web-based tool that helps engineers check if AI models are fair and comply with local regulations, particularly for critical applications like 6G networks.

It's a rule in scheduling that makes sure tasks don't happen at the same time, and scientists use clever math to solve these complex timing puzzles efficiently.

Recent AI research is pushing boundaries in LLM intelligence, efficient computing, and smarter data handling for diverse applications.

It's a special similarity score that helps AI find rare anomalies in messy data, making active learning smarter.

A novel ML algorithm is a newly invented way for computers to learn and make decisions, often leading to breakthroughs in how we solve complex problems.

Numerical optimization is about finding the best possible answer to a problem by repeatedly making small adjustments until a target goal (like minimizing an error) is met.

It's an AI that uses a general coding model to solve advanced math problems and prove theorems by interacting with specialized tools, making it adaptable and powerful.

NVIDIA Triton is a free software that helps deploy and run AI models from any framework on any computer chip super fast and efficiently.

o1 is a way to make AI models perform better by using more clever computations during inference, instead of just making the models larger.

It's a way to make AI understand text better, especially tricky product descriptions, by treating the text like a picture and using image recognition.

OctoBench is a benchmark that tests how well AI coding agents follow complex instructions and rules within coding environments, revealing a gap between task completion and rule compliance.

Offline RL lets AI learn from existing data without needing to try things out in the real world, making it safer and faster to train.

It's a design for AI writing tools where the AI only helps when you specifically ask it to, so you still feel like the true author.

It's a smart way for AI models to learn by comparing how well they understand audio versus text, making sure they don't miss important details.

Online model updates let AI models automatically learn and adjust to new information as it comes in, keeping them smart and accurate even when things change.

ONNX Runtime is a universal engine that makes AI models run super fast on almost any device by optimizing them and using special hardware.

The Open Research Knowledge Graph converts scientific papers into structured, queryable data to make research findings easier for computers and people to use and integrate.

It's a public tool that lets advanced AI language models interpret and create text from numerical data representations, making those numbers more understandable and useful.

AI models whose code and data are publicly shared, enabling widespread use, research, and community-driven improvements.

OpenAI is a top AI company that creates powerful AI models like GPT, aiming for safe and beneficial superintelligent AI.

A massive AI language model from OpenAI, likely with 120 billion parameters, capable of understanding and generating highly sophisticated text.

OpenCV is a free software library that helps computers process images and videos for tasks like object detection and robot navigation.

OpenMP is a tool that lets programmers easily make their code run faster by using all the cores in a computer's processor simultaneously.

Optical flow is a computer vision method that tracks pixel movement in videos, helping AI understand motion for tasks like action recognition and robot navigation.

It's a math rule that tells you the best time to stop a random process without messing up its average outcome, used in AI to smartly cut off bad search options.

ORACLE-CT is a smart AI system that helps doctors quickly and accurately analyze CT scans by focusing on specific organs to find problems.

It's a system that lets many different AI programs securely check and learn from each other, making them collectively smarter than one big program.

ODEs are math equations that describe how things change continuously, helping us model and predict the behavior of dynamic systems.

OSW helps big AI models learn new tasks without forgetting their old skills, especially complex ones like coding, by protecting specific parts of their memory.

OurBench is a new benchmark that tests how good AI models are at writing and fixing complicated SQL code for businesses, showing they still have a lot to learn.

PAC learning is a mathematical theory that proves if a computer can reliably learn a concept from data, guaranteeing it will be mostly right most of the time.

Paired sampling improves how accurately AI models explain their decisions by making feature importance calculations better, without extra computational cost.

Parameter updates are the core way AI models automatically tweak their internal settings during training to get better at their tasks.

It's a method that quickly finds good explanations for group recommendations by balancing different group members' needs, especially when there's not much data.

PID is a math tool that precisely shows how different information sources contribute to a result, helping to see if AI models truly forget sensitive data.

PLS is a smart way to find hidden connections in complicated data to make better predictions, even when many factors are related, like mapping LLM states to satisfaction scores.

A POMDP is a way for an AI to make smart decisions when it can only see parts of the situation, using its observations to guess what's really going on.

PCL-Reasoner-V1.5 is a huge AI model that's really good at math, trained with a special, more stable method to achieve top performance.

PIM is a module that helps virtual avatars understand their surroundings and perform actions with objects based on text instructions, making interactive videos more realistic.

PerfGuard is a framework that helps AI agents choose and use tools more effectively by understanding their actual performance, especially for tasks like generating images.

A testing method that shuffles parts of a document to check if AI models unfairly prioritize some sections or languages over others when creating document summaries.

In AI security, persistence means an attacker can make their malicious commands stick around in an LLM's memory or data, letting them keep control over time.

It's an AI writing feature that gives advice based on a specific role or character, but it doesn't always make you feel more like the original author.

A robot AI component that figures out the timing of actions independently of their location, making human-robot interactions more natural and responsive.

Microsoft's Phi series are compact AI models that perform surprisingly well on reasoning tasks because they're trained on very high-quality, educational data.

Phi-4 is a smaller, focused AI model that performs like bigger ones, helping make advanced AI more affordable and sustainable.

PhyloEvolve uses an AI agent to automatically and intelligently optimize GPU code by learning from the step-by-step changes and results, much like how species evolve.

PIMC is a computer simulation trick that uses random sampling to figure out how quantum particles act in materials by treating them like tiny, wobbly chains.

It's a system that helps big AI models avoid mistakes and be more trustworthy by having one part create plans and another part strictly check and correct them.

A module that makes AI agents better at finding software bugs by proactively searching for defects using embedded testing goals, rather than just focusing on task completion.

PLATE helps big AI models learn new things without forgetting old knowledge, even without access to past data, by cleverly updating only specific, redundant parts of the model.

POBench-PDE helps AI models solve scientific equations using only partial data, making them useful for real-world problems with missing information.

It's a design rule for AI writing tools that shows you exactly what the AI suggested and where it came from, helping you feel like the true author.

A Policy Decision Record is a detailed log that shows how an AI agent's request was evaluated against rules, explaining why it was allowed or denied, especially for important actions in secure systems.

A new evaluation method that accurately measures where non-verbal sounds occur in audio by separating speech recognition errors from sound detection errors.

It's a method to make an AI model's confidence scores more honest, so you can trust its predictions more, especially in important situations, without making the model less accurate.

AI models trained on huge amounts of text that can understand and create language, then quickly learn new tasks with minimal examples.

Predictive analytics uses past data and smart algorithms to predict what might happen next, helping people make better decisions before things occur.

PrefixRL helps AI models learn complex tasks more efficiently by using successful partial solutions from past attempts to guide new learning, making the process more stable and faster.

Train a model on a big general task first, then tweak it for a smaller specific task to make it perform better and learn faster.

PrevizWhiz is an AI system that helps filmmakers quickly turn rough 3D ideas into cool-looking video previews, making it easier and faster to plan movies.

PCA is a math trick to shrink big datasets into smaller, more manageable ones by finding the most important patterns, making them easier for computers to understand.

A Principal Semantic Vector (PSV) captures the core meaning from an AI's thinking process to help put a hidden digital mark on its text without breaking its logic or quality.

PRISM is a new AI framework that improves recommendation systems by making sure the computer accurately understands items and doesn't lose important details when suggesting things.

A privacy budget is a quantifiable limit on how much personal information an AI system can reveal, ensuring privacy while sometimes also improving fairness.

It's when hackers trick an AI into ignoring its rules, giving them more control to do bad things like steal info or make unauthorized actions.

PSCE creates reliable 'what if' explanations for AI models that are guaranteed to stay accurate and consistent even when the AI model gets updated.

Procedural fairness ensures AI decisions are made fairly by giving everyone an equal voice in the process, not just by balancing the final results.

It's about picking the most important parts of data to clean or fix so AI models learn more effectively, especially with messy or private information.

A method to compress high-dimensional data by breaking it into smaller pieces and quantizing each piece, making large AI models more efficient for memory and speed.

Product-game states help AI systems figure out what goals they can actually achieve when dealing with many rules, even if some conflict, and then create plans to meet as many as possible.

It's a math trick that finds the best answer to a problem by taking small steps and then forcing the answer to stay within specific rules.

It's a super-efficient way to build small AI models for predicting future trends, perfect for tiny computers in industrial settings.

ProMem helps AI remember better by making it actively review and question its past conversations to fill in gaps and fix errors.

Prompt-based methods use small, task-specific input cues to help AI models learn new things without forgetting old ones, making them more adaptable.

P2O uses an AI to write and train small programs that find the best solutions for hard problems, especially when testing those solutions is costly.

PromptSplit helps researchers understand why different AI models give different answers to the same questions by analyzing their prompt-response patterns.

ProP helps AI models learn new things continuously without forgetting old knowledge by using unique task-specific prompts and prototypes, avoiding common interference issues.

A math tool that helps solve tough optimization problems by making "soft" versions of rules, rather than strict ones.

PPO is a smart way to train AI agents to learn new skills by making sure they improve step-by-step without messing up their progress, making it stable and efficient.

PPO is a stable and efficient AI learning method that helps robots and autonomous systems learn complex behaviors by making careful, small adjustments to their decision-making rules.

A technique where an AI model generates its own 'fake' labels for unlabeled data and then uses them to learn more, especially when real labeled data is scarce.

It's a method where an AI model automatically labels huge amounts of unlabeled data to create bigger datasets, helping train even better AI models without needing people to label everything.

PsyFIRE is a detailed system for identifying specific types of client resistance in online therapy chats, helping AI understand and explain these behaviors better than before.

PUMA is an AI system that lets robots see and decide where to step in real-time for agile movement over obstacles, like a parkour athlete.

It's a smart way to make sure the digital labels for items in AI recommendation systems are clean and accurate, helping the system suggest better things.

Putnam 2025 is a tough 12-problem math test that a new AI system passed perfectly, proving AI's growing ability in advanced mathematical reasoning.

Python packages organize code into reusable modules, making it easy to share, install, and deploy software for various applications.

A Q-Former is a smart component that helps AI models understand and respond to spoken instructions in multiple languages more accurately by managing language differences.

Q-learning is a smart way for AI to learn the best actions in any situation by trying things out and remembering what worked best to get rewards.

QFed uses quantum computing to make federated learning models much smaller and faster for edge devices, while keeping their accuracy.

It's a voting method where the cost of a vote increases quadratically, letting people show how much they care, and in AI, it helps combine many small models' opinions effectively.

It's a special function that helps a team of small AI models combine their 'votes' in a smart, non-linear way to make better decisions than one huge AI model.

Quantization shrinks big AI models by simplifying their internal numbers, making them run efficiently on small devices.

QVLM is a smart AI that generates code to precisely count and measure objects in images by working directly with pixel-level details, fixing a major flaw in current visual AI.

Qwen 2.5 is a smart AI model from Alibaba that's good at understanding and following instructions, often used by researchers to test AI capabilities.

Qwen 2.5 7B is a free-to-use, 7-billion parameter AI model from Alibaba that's good at understanding language, writing, and coding, and it's efficient enough for many uses.

Qwen-3-4B-Instruct is a medium-sized, instruction-following AI model used by researchers to make AIs more honest by teaching them to say "I don't know" when uncertain.

Qwen-8B is an 8-billion parameter, open-source AI model from Alibaba Cloud that's great at understanding and generating text in many languages.

Qwen2.5 is a family of open-source AI models, with some versions performing as reliably as top commercial AIs when it comes to using tools in automated systems.

Qwen2.5-7B is a versatile 7-billion parameter AI model that performs well in specialized tasks like anomaly detection and engineering simulation, often after being fine-tuned.

Qwen2.5-VL is an AI model that understands videos and text, but needs improvement in counting and avoiding made-up details, which can be fixed with special training.

Qwen2.5:32B is a 32-billion parameter open-source AI model that reliably uses tools as effectively as GPT-4.1, making it great for business automation and private data tasks.

Qwen3 is a series of powerful AI models from China used as a base for many advanced language tasks and for studying AI behavior and efficiency.

The Qwen3 Embedding model turns words into smart numbers so computers can understand what text means, making search and AI better.

Qwen3-Max is a very advanced AI language model that researchers use to see how smart AI can get and where it still needs to improve.

R3 is an AI framework that improves multimodal models' ability to both generate and understand by making them iteratively generate, reflect on, and refine their outputs.

RAGCRAWLER is a smart attack method that can slowly steal private information from AI systems by asking carefully planned questions, exposing a privacy flaw.

RAILS is like an AI teaching another AI how to be better by giving it feedback, making the learning process faster and more scalable.

Random Forests combine many simple decision-making trees to make more accurate and reliable predictions than any single tree could alone.

RaP is a training technique that helps AI models become more reliable by specifically learning from data points they are less confident about.

RRTMG is a slow but important radiation calculator in big weather models, and AI is being used to make it much faster.

A philosophical idea that a fair society guarantees basic liberties for all and arranges inequalities to help the least well-off, stressing the importance of fair processes and equal voice in decision-making.

RBench is a new standardized test for robot video generation models, helping researchers see which models are best and where they need to improve, especially for realistic physics.

It's a way to combine or refine AI models to make them faster and smarter, particularly for big reasoning problems, by improving how efficiently they learn and operate.

ReAct lets AI models reason through problems step-by-step and use external tools, making them smarter and more reliable at complex tasks.

Getting instant predictions from live data streams to make quick decisions, like knowing a machine might break down right now.

A method called RELIEF helps large AI models improve their reasoning and efficiency by subtly changing their internal 'beliefs' about how they think, without needing expensive human examples.

ReasonMark is a smart way to watermark AI-generated text, especially from reasoning AIs, by embedding a hidden signal without messing up the AI's logic or making it slow.

An AI system that uses "Theory of Mind" to understand what reviewers are thinking and then writes smart, persuasive responses for academic paper rebuttals.

RebuttalAgent is an AI that helps researchers write better responses to paper reviews by understanding what reviewers are thinking and crafting smart, persuasive arguments.

RECAP is an AI system that accurately detects and explains different types of client resistance in online therapy chats, helping counselors do their job better.

Reciprocal Rank Fusion combines ranked lists from different search engines or retrieval methods to create a better, more comprehensive single list.

ReCon is a post-processing tool that makes community detection in networks with positive and negative links more accurate and reliable by iteratively refining the results.

ReCreate helps AI agents learn to build and improve themselves automatically by analyzing their past actions and outcomes.

Recurrent models are AI systems that use an internal memory to process sequences and adaptively update what they remember, proving robust for complex memory tasks.

RNNs are neural networks with memory that process sequences of data, like words in a sentence or measurements over time, by using past information to understand the present.

Recursivism is a framework for analyzing art where the creative process itself learns and changes its own rules, particularly relevant for AI-generated art.

Methods for creating realistic and diverse datasets or instructions for AI by either mimicking existing examples (reference-guided) or generating novel ones (reference-free).

It's an AI learning method where the AI 'reflects' on its own performance to get better at learning and adapting.

Reflexion lets AI agents learn from their mistakes by thinking about what went wrong or right, then using that knowledge to do better next time.

A training method for AI models that makes them forget specific facts so they can focus on pure reasoning, reducing errors and improving logic.

Regularization constraints keep AI models stable and prevent them from overfitting by limiting how extreme their internal values can get during training.

A new AI optimizer that uses a Bayesian approach to make deep learning models train faster and generalize better by simplifying complex calculations.

It's a way to train smart AI models to do specific jobs by letting them learn from rewards, making them better at following instructions for complex tasks.

A method to train AI models to be more honest and reliable by teaching them to say "I don't know" instead of guessing wrong.

RLVR trains AI to admit when it doesn't know an answer, making it more reliable and less likely to make up facts.

REKD helps small AI models better explain their decisions by learning from a big, expert AI model's explanations and predictions.

RelayLLM lets a small AI model ask a big AI model for help only on the hardest words, making complex tasks much faster and cheaper without losing much accuracy.

RENEW is a smart navigation system for self-driving boats that helps them find safe, diverse, and energy-efficient paths through tricky ocean waters with currents.

RepoReason teaches AI models to understand and work with whole code projects, not just snippets, to help developers with complex tasks like fixing bugs or improving code.

It's about making AI models smart enough to act like software engineers, working on real code projects and following all the rules and guidelines.

A request classifier is an AI tool that figures out what a user really wants, helping other AI systems give smarter, more accurate responses, especially when dealing with complex topics.

A research field in AI/ML is a focused area where experts develop new techniques to solve specific, complex problems with intelligent systems, like making them more trustworthy or robust.

A type of deep neural network that uses shortcut connections to train very deep models, improving performance and efficiency for complex tasks like simulating weather.

ResMAS helps make groups of AI language models more reliable by proactively designing their communication networks and individual instructions to withstand failures.

ResNet-18 is a popular, relatively small but powerful AI model that uses special "skip connections" to learn effectively, making it great for many image-related tasks and as a base for other AI systems.

A brain-inspired AI network that uses sparse, dynamic connections in a special geometric space to process information much more efficiently than standard models.

Responsible AI means making sure AI systems are fair, safe, and understandable so they help people without causing harm or bias.

RAG helps AI models give better answers by finding relevant information first, like a smart student checking their notes before speaking.

It's an AI system that uses medical facts and multiple AI agents to automatically create specific checklists to check if other AI models are giving safe and correct advice, especially in healthcare.

It's a way to make AI models smarter by giving them specific, helpful examples from a memory right when they need to answer a question.

RGMP-S helps humanoid robots perform complex tasks by improving their understanding of the world and making them learn new movements efficiently using geometric rules and spiking neural networks.

RHealth is a powerful software framework that simplifies building AI tools for healthcare, making it faster and more accessible for everyone.

RLSTG is a smart graph network that uses curved math to accurately model how complex, real-world networks change over time, fixing issues with older 'flat' models.

It's a fancy math concept for a curved space that helps AI models better understand complex, non-flat data, like how different parts of the brain connect.

RIFT is a smart way to train AI language models that uses all the model's practice attempts, even the wrong ones, to make them better and more efficient.

RISE trains AI models to use tools more accurately by generating virtual examples of correct and incorrect actions, helping them avoid subtle misunderstandings of user intent.

A smart routing system that uses AI to predict traffic and dynamically reroute deliveries to avoid congestion and improve reliability.

It's how robots figure out where they are, where they need to go, and how to get there without bumping into things, now also considering fairness and privacy.

RoPE helps AI models understand sequence order by rotating data based on its position, allowing them to process longer texts and complex data more effectively.

RTCE is a test for AI models that write code, checking if they can consistently understand and reverse coding operations, revealing their struggles with reliable code reasoning.

A high-end NVIDIA graphics card for professionals that uses new tech to make AI, design, and science tasks much faster in workstations.

It's a method to make AI models better at ignoring irrelevant location details and noise, so they can accurately identify things like people no matter where they are in an image.

It's a smart way to make many small AI programs work together dynamically to solve big problems, achieving top performance without needing one massive, expensive AI.

SafeRemind is a way to get smart reminders without giving up your privacy, by keeping your personal data secure on your own device.

Systems where mistakes are dangerous, so AI in them needs to be super sure of its predictions by checking its own confidence before doing anything.

"SAGE" is either a system that makes AI better at understanding emotions from messy data, or a smart tool that automatically creates hard questions and answers to train advanced search engines.

It's a way to train AI models to look at the whole image or scene, not just tiny parts, making them more robust and less prone to errors.

A method in AI model training that adaptively weights samples to help smaller models keep their strong points and even outperform bigger teacher models.

A smart way to train small AI models to be as good as or better than big ones for specific jobs, by learning from the big model's training process.

ScienceWorld is a benchmark environment used to test how well AI agents can solve complex, multi-step scientific problems and learn from their interactions.

SD 1.4 GLIGEN is a text-to-image model that lets you tell it exactly where to put objects in the picture using boxes, not just words.

SD 1.5 BoxDiff is a specific AI model used as a benchmark to see how accurately text-to-image AIs can follow spatial instructions like 'object A above object B'.

SDA2E is an AI model that uses a smart autoencoder and active learning to efficiently find rare, hidden threats in complex, imbalanced data, especially in cybersecurity.

Search-R1 is a popular toolkit for training AI agents to search and reason through vast amounts of information, particularly for answering tough questions in scientific and technical domains.

SearchGym is a simulated training ground for AI search agents that creates reliable, fact-checked data to teach them complex reasoning without the cost or errors of real-world data.

It's a special training method that uses a fake but realistic internet to teach AI search agents how to find information accurately without costing a lot or getting confused by bad data.

A specialized AI model that processes long audio by analyzing short segments and their long-term relationships, primarily used for detecting AI-generated music.

The SEIR model is a math formula that predicts disease spread by tracking people who are susceptible, exposed, infectious, and recovered, especially useful for diseases with an incubation period.

Self-attention helps AI models understand context by letting them weigh how important different parts of an input are to each other, enabling better understanding and generation.

Self-MedRAG is an AI system that answers medical questions by repeatedly searching for information and checking its own logic to avoid errors, making it more reliable for doctors.

Self-Refine is when an AI model checks its own answers, finds errors, and then fixes them to make its output much better.

It's a system where experts use play money to bet on a problem's solution and chat about it, helping everyone figure out the best answer together, even if there's no clear right or wrong.

Self-supervised learning lets AI models learn from data by generating their own practice problems, reducing the need for human-labeled examples.

Semantic embeddings turn complex data into meaningful numbers, helping AI understand context and relationships for tasks like content evaluation, robotics, and fraud detection.

Semantic encoding is how computers understand the meaning and relationships in information, not just its structure, making AI smarter at answering complex questions.

Semantic segmentation is a computer vision task that assigns a specific category label to every pixel in an image, providing a detailed map of what's where.

STeX helps AI systems search complex information networks faster and more accurately by intelligently blending text meaning with network connections.

SemanticALLI makes AI systems faster and cheaper by caching and reusing common internal thought processes instead of re-generating them every time.

SemExplainer is an AI tool that explains complex social network recommendations by identifying the specific network connections that work together to influence suggestions.

Semi-supervised learning trains AI models using a mix of a little bit of labeled data and a lot of unlabeled data to make them smarter and more accurate.

SemiMol is a smart way for AI to predict molecule properties better, especially when data is scarce or molecules are tricky, by learning from unlabeled examples and gradually improving.

SeNeDiF-OOD is a smart way for AI to spot weird or unknown data by looking at it in layers, making AI systems more reliable in the real world.

Sequence modeling teaches AI to understand and generate ordered data, like text or time series, by learning how elements relate to each other over time.

SeqWalker is a way to analyze or create sequences by having a virtual "walker" explore the data step-by-step to understand complex patterns.

SFAA uses AI to automatically and smartly remove private details from research documents, making sure personal information is protected while keeping the document useful.

Shapley values fairly measure how much each input feature contributes to an AI model's output or a specific event, helping us understand its decisions.

SHARP is a system for evaluating AI models that deeply analyzes various types of social harm and focuses on worst-case failures, rather than just average performance, to uncover hidden risks.

SAM is a training trick that helps AI models learn to generalize better by finding "smoother" solutions, and newer versions like X-SAM improve on this idea.

ShopSimulator is a challenging virtual shopping environment used to test and train AI shopping assistants, revealing their current limitations in complex e-commerce tasks.

ShortCoder makes AI-generated code shorter and more efficient by simplifying its structure without changing what it does, helping AI models generate code faster.

ShorterCodeBench is a dataset of simplified Python code that helps AI models generate code faster and with less computing power by reducing the code's length while keeping its meaning.

A Siamese Network uses two identical neural networks to learn how similar two things are, making it great for tasks like face recognition or comparing data points.

A score that tells you how well your data points are grouped into clusters, with higher scores meaning better, more separated groups.

A smart way for AI to find rare anomalies in big, uneven datasets by strategically asking for labels based on data similarities, saving expert time.

SimMerge helps combine big AI models by predicting the best way to merge them, avoiding costly trial-and-error experiments.

SimuAgent lets advanced AI models understand and automate complex engineering designs in Simulink by translating its technical language into something easier for the AI to process.

SimuBench is a big test for AI models to see how good they are at designing and simulating engineering systems using Simulink.

It's an algorithm that finds good solutions to difficult problems by randomly trying options and slowly refining them, like cooling a hot metal.

Simulation is like building a computer model of a real system to safely test ideas and predict outcomes without having to do them in the real world.

A simulation platform is a virtual world used to safely test and improve AI systems, like self-driving cars, by letting them practice in a fake environment before going real.

A simulation software framework lets you test prediction models in a fake but realistic environment to see how they impact real business outcomes like costs and service, not just how accurate they are statistically.

SIN-Bench is a benchmark that tests if AI models truly understand scientific papers by making them point to the exact evidence (text and figures) for their answers, not just provide a correct answer.

SIN-Data is a dataset of scientific papers with text and images used to test if AI models can truly understand and connect information across different parts of a document, not just find keywords.

It's a smart way to make AI image compression much faster and more efficient by refining the model's predictions.

SMART is an AI that predicts physics stuff on complex shapes using just a basic 3D scan, skipping the hard part of making a detailed digital grid.

SMCR is a framework used to study how easily AI models can be convinced to change their answers or "beliefs" by analyzing the source, message, channel, and the model itself.

SMOTE creates synthetic data points for rare classes to help AI models learn better from imbalanced datasets, improving prediction accuracy and influencing explanation consistency.

SMT solving checks if complex logical rules involving numbers and other data types can be satisfied, often used to find bugs or verify designs in computer systems.

SMT-LIB is a standard computer language for writing logic problems that specialized programs called SMT solvers try to solve.

Games like Mafia where players have secret roles and must lie or figure out who's lying are used to test how good AI is at deceiving people.

A new test for AI that measures how well language models can use dialogue to change other characters' minds, revealing a big gap between AI and human social skills.

It's a smart way to combine multiple AI models' best guesses to get a much more accurate answer, especially for tricky classification problems.

A Software Engineering Framework provides a structured way to build software, especially for creating smart AI agents that can learn and use expert human knowledge to solve complex problems.

A big video dataset that helps AI learn to recognize detailed actions, often without being told what each action is.

Sophia is a clever algorithm that speeds up and improves the training of big AI models by efficiently using insights into how the model learns.

Space Planning Problems are complex, high-stakes tasks, like managing satellites, that are hard for general AI to solve due to strict physical rules and long-term planning needs.

SPARQL is a query language for knowledge graphs, letting you find and manage information in a web of interconnected facts, similar to how SQL queries traditional databases.

A smart part of a forecasting model that helps predict repeating patterns in data very efficiently, so it can run on tiny computers.

A method to train big AI models with less memory by smartly handling compressed data, preventing errors, and keeping performance high.

A formal framework that helps AI understand and correctly process spatial data, preventing errors and enabling genuine geographic reasoning.

Spatial reasoning helps AI understand 3D environments, and new techniques allow AI models to actively look around to better grasp spatial relationships.

Spatial-temporal learning teaches AI to understand moving things by looking at both what's in each picture and how it changes over time.

It's the data about where things are and how they move over time, crucial for AI to understand videos and dynamic scenes.

Spectrograms turn sound into a visual map of frequencies over time, helping AI spot hidden details in audio, like whether speech is real or fake.

Speech Language Models let AI understand and talk using spoken language directly, making interactions feel more natural and efficient.

SPF is a way to update big AI models for new jobs so they get smarter without accidentally becoming dangerous or easy to trick.

SPIKE improves AI models for physics by finding simple, sparse linear rules in complex systems, making them much better at predicting future behavior and generalizing to new situations.

Spiking Neural Networks are energy-efficient, brain-inspired AI models that use discrete 'spikes' for communication, perfect for low-power devices and fast, real-time tasks.

Splines are smooth mathematical curves used to represent continuous, noisy data, helping AI models better process real-world signals for more stable decision-making.

SEResNet is an improved deep learning model that uses an attention mechanism to better identify and use important image features, leading to higher quality results.

SQuID is a new dataset of satellite image questions and answers designed to challenge AI models to get better at precise counting and measuring in pictures.

Stable Diffusion 1.5 is a widely used open-source AI that turns text into high-quality images, enabling creative work and research.

It's a way to understand how complex systems work by looking at how variables change over time, even when they influence each other with delays.

It's a system that intelligently picks the best-sized AI model for each step of a task to save money and improve results.

Statistical feature extraction turns raw data into useful numbers that AI models can learn from more effectively.

It's a computer method that uses the 3D shapes of molecules to predict how they interact, which helps in designing new drugs.

AI models with organized memory systems that help them remember and adapt, but currently struggle with updating old information effectively.

AI writing tools can adapt to your personal writing style to make suggestions feel more like your own, helping you maintain ownership and use more AI help.

Subword tokenization splits words into smaller parts so AI models can understand new words, slang, and mixed languages more effectively.

SFT teaches a big AI model to follow instructions by showing it many examples of what to do, making it more useful for specific tasks.

SVG is a smart image format that uses math to draw pictures, so they always look sharp no matter how big or small you make them.

SVGFormer is an efficient AI model for 3D medical images that analyzes meaningful regions (supervoxels) using a special encoder to better understand diseases like brain tumors.

It's a tool that makes up fake but realistic customer orders to help companies test how good their prediction software is for managing inventory, especially for hard-to-predict items.

T-Retriever helps AI language models find better information by organizing complex data into trees and intelligently compressing it, leading to more accurate and relevant answers.

T2QBench is a benchmark that quizzes AI agents on their understanding of virtual environments, revealing that task success doesn't always mean they truly comprehend the world.

Talos is a special math formula that helps recommendation systems pick the best few items for you much faster and more accurately, even when your tastes change.

TAM-Eval is a benchmark that evaluates how effectively AI models can create, fix, and update software tests in real-world programming projects.

It's a way to train AI models to understand categories and subcategories, making them less likely to make big mistakes in complex fields like medicine.

TCAV helps us understand what specific human-defined concepts an AI model uses to make its decisions, making black-box models more explainable.

TEA-Dialog is a dataset that helps train AI emotional support agents to use tools for factual information, making them more reliable and less prone to errors.

An Indian national standard for evaluating AI system fairness, particularly for telecom and 6G applications, ensuring compliance with local regulations.

It's how AI agents connect past information to current interactions, letting them remember and learn continuously over time.

Temporal-Difference learning is how AI programs learn to predict future outcomes and make better choices by comparing their current predictions with what actually happens next.

A training method that helps robots make quick, correct decisions by learning to use slightly delayed high-level plans alongside real-time sensory data.

Tencent WeChat Channels is a huge platform within WeChat that uses AI to predict user behavior, serving as a critical real-world testbed for efficient and adaptive machine learning models.

Tensors are like smart, multi-dimensional containers that hold all the numbers and data AI models use to learn and make decisions.

AI systems that let you chat with computers using text to get information, complete tasks, or have a conversation, often becoming smarter when multiple AI agents collaborate.

Text-to-image models are AI systems that turn written descriptions into unique pictures, like a super-smart digital artist following your instructions.

Text2Cypher lets AI understand regular questions and turn them into graph database queries, making AI more reliable for scientific research by using factual knowledge graphs.

TF-IDF assigns a score to words in documents to show how important they are, helping computers understand text better for tasks like searching or categorizing.

It's a smart computer system that automatically finds and names the main topics in online conversations, even across different social media sites, using AI to make the results easy to understand.

It's a method that makes AI models explain *why* they choose certain actions or parameters when using tools, making them more accurate and transparent.

TDCCM is a technique that figures out cause-and-effect in complicated industrial systems by considering time delays and how everything is connected, making predictions more reliable.

It's a smart way to figure out what causes what in complicated systems, even when things happen with a delay and everything is connected, making monitoring tools much better.

TDPCM helps figure out direct cause-and-effect relationships in complicated industrial systems, even when things happen with a delay and are all connected.

TD-DFT is a computer method used to figure out how molecules react to light, helping design things like glowing materials or better solar cells.

These models learn from many different types of time-series data to make good predictions on new tasks, especially when only a little specific data is available.

TimeCast is a smart system that predicts future events by learning from changing sensor data patterns in real-time.

It's a way for AI models to keep learning and adjusting on a specific task during testing to find the absolute best answer, rather than just a generally good one.

TKTO is an optimization method that helps AI models for anomaly detection become better at understanding and solving various time series problems, making them more versatile.

TSR is an AI pipeline that helps machines write persuasive academic rebuttals by understanding a reviewer's thoughts, planning a strategy, and then crafting a response.

A guide for developers to pick the right AI tool for a job, preventing them from using overkill (and expensive) AI for simple tasks.

ToolBench is a benchmark that evaluates how effectively large AI models can use external tools and APIs to solve complex, multi-step problems.

A PyTorch library that speeds up big scientific calculations with sparse data on GPUs and makes them compatible with machine learning's automatic differentiation.

TraDo is a specific AI model that generates text using a time-based process, helping researchers create more varied and high-quality written content for creative and reasoning tasks.

It's a math-based way for robots and autonomous vehicles to figure out the absolute best way to move from one point to another while following rules and avoiding problems.

A standardized framework that lets AI agents securely communicate their desired actions to blockchains, preventing errors and malicious activity.

Transformer-based models are advanced AI systems that use a special "attention" trick to understand complex relationships in data, making them great for things like language and images, but they might struggle with constantly updating memories.

Transformers are a type of AI model that uses a special "attention" mechanism to understand relationships in data, making them super effective for language and other complex tasks.

TGA is an efficient AI model that uses a special graph to understand complex user behavior sequences on platforms like e-commerce, making it faster and more scalable than traditional methods.

It's a method to catch AI language models that secretly memorize test answers, even when the tests are translated into different languages.

TransportAgents is an AI system that combines multiple specialized AI agents to accurately predict traffic crash severity, outperforming single large language models.

TravelPlanner is a tough benchmark that tests how well AI agents can plan and execute complex, multi-step tasks, especially those requiring learning from experience.

Tree ensembles combine multiple simple decision-making trees to make much more accurate and trustworthy predictions than any single tree could alone.

It's a smart way to organize complex information into a tree structure so AI can find and use it better, especially for answering tough questions.

It's a method that proved that finding the right information is the biggest challenge for AI when dealing with long documents, leading to new ways to train AI.

TREX is a smart system that predicts the best combination of languages to train language model tokenizers, making them more efficient and performant.

TNC is a new way to find negative emotions in spoken language, helping scientists study how AI models understand feelings and relate to human brain responses.

TriPlay-RL is an AI system that uses three collaborating AI agents to automatically make large language models safer by iteratively identifying and mitigating harmful content generation.

It's a smart way to train AI models on graph data to group similar items together and push different items apart, making the training faster and more effective.

A Trust Gateway is a system component that securely manages and verifies permissions and identities in complex digital environments, making sure delegated authority is safe and auditable.

TSDA is a method for multimodal sentiment analysis that improves performance by separating and aligning temporal and spatial information from different data sources before combining them.

A new dataset that teaches AI models to reason about and discuss unusual patterns in time series data.

TSEvol is a multi-agent AI algorithm that makes large language models much better at finding and explaining anomalies in time series data.

It's a method to make complex AI image generation and compression super fast by cutting down the reconstruction process to just two essential steps.

UA-3DTalk is a system that creates highly realistic 3D talking faces with accurate emotions by better aligning audio with expressions and intelligently combining multiple camera views.

UCF101 is a standard dataset of 101 human action videos used to train and test AI models for recognizing what people are doing.

ULP optimization precisely tunes floating-point calculations to minimize errors, ensuring bit-level accuracy in complex numerical problems.

It's a smart system that quickly and reliably ranks AI language models by only testing what's necessary, saving a lot of evaluation effort.

Specialized parts of AI models that figure out how reliable information is, helping the model combine data smartly and make better, more confident predictions.

A U-shaped neural network that's really good at precisely outlining things in images and is a key part of many modern image-generating AI systems.

A new system for securely and verifiably sharing digital access rights between people and AI across different online platforms without exposing sensitive credentials.

UMAP helps you see hidden patterns in really complicated data by squishing it down into a simple picture, making it easy to spot groups and differences.

These are ways to numerically measure what makes a city neighborhood feel unique, often using AI to analyze virtual versions of places.

URLGuard is a small security tool that helps AI web agents spot and block tricky bad website links to keep users safe.

An easy-to-use website that helps people quickly get answers and instructions from smart computer programs, even for complicated tasks like fixing medical machines.

A new way to measure fairness in AI that considers user needs and data randomness, often by linking it with privacy, to make robots and AI more ethical.

V-CAGE is a framework that generates high-quality, realistic training datasets for robots by making sure virtual scenes are physically sound and that AI actions correctly match complex instructions.

VAEs are AI models that learn to create new data similar to what they've seen by understanding the underlying patterns and variations in a compressed, probabilistic way.

vecTrans is a smart, fast way to analyze many interacting data streams in time series forecasting, making complex models run much quicker.

Verilog is a language used by engineers to design and describe digital electronic circuits for computer chips, allowing them to simulate and build hardware.

It's a smart system that makes video AI run better on devices with limited power by dynamically picking the best-sized model for the job.

AI models that edit videos by generating new content, but need help remembering past edits to keep everything consistent in multi-step changes.

VideoLLaMA-3 is a smart AI that understands and talks about videos by combining a powerful language model with video-seeing capabilities.

VideoMaMa uses AI to precisely cut out subjects from videos, even without specific training data, by leveraging synthetic data and advanced generative models.

It's an AI component that intelligently picks camera views in 3D spaces to help other AIs find all the information they need for a task.

VIOLA helps big AI models learn from new video types with very little labeled data by smartly choosing what to label and carefully using unlabeled examples.

Virtual Urbanism uses AI to generate artificial city environments to scientifically measure and understand what gives a real city its unique identity.

Big AI models trained on tons of images and videos that can understand and process visual information for almost any task, making them highly versatile.

AI models that combine visual understanding with language processing to interpret and reason about the world like humans do.

Vision Transformers are AI models that use a global self-attention mechanism on image patches to understand visual data, making them powerful for many computer vision tasks.

VLMs are AI models that can understand and work with both pictures and words, making them great for tasks like building app interfaces from descriptions.

VLA lets robots understand spoken commands, see their environment, and then act on those commands, making robots easier to control for complex tasks.

A VLA policy lets robots understand spoken commands and visual cues to perform actions, learning efficiently in virtual worlds to work better in the real world.

Visual servoing uses camera vision to guide robot movements, making them precise and adaptable for tasks like picking up objects or interacting with people.

Rules for good data visualization, when taught to AI, let anyone create expert-level charts and graphs automatically.

VisWorld-Eval is a study framework exploring how AI models can use visual generation to improve their understanding and reasoning about the physical world, similar to how humans think.

The Vision Transformer (ViT) uses the same AI tech that powers ChatGPT to understand images by treating parts of an image like words in a sentence.

VAD models are like smart filters that tell computers exactly when someone is speaking in an audio recording, ignoring background noise.

A design trick for AI writing tools that makes sure the AI writes in your style, so you still feel like the author.

Waterfall is a way to secretly mark AI-generated text for ownership and misuse detection, but it breaks easily if the text is translated.

WaveLSFormer is a smart AI model that combines wavelets and Transformers to predict profitable stock trades in fast-moving markets better than other AI models.

WESR-Bench is a new dataset and evaluation tool that helps AI accurately find and pinpoint non-verbal sounds in speech, like laughs or cries, by providing clear definitions and a smart way to measure performance.

Wikidata is a giant, free, online database of facts organized so computers can easily understand and use the information.

Researchers use a "Wizard-of-Oz" setup by having a hidden person act like a computer to test new tech ideas with real users before the tech is actually made.

WoW-bench is a new benchmark that tests how well advanced AI models can act as agents in complex business systems, revealing their difficulty in predicting hidden effects of their actions.

X-SAM is a smarter way to train AI models that helps them generalize better by explicitly guiding them to find flatter, more stable solutions in the learning process.

XFactors is a weakly-supervised AI framework that disentangles specific data characteristics, allowing explicit control over them, solving issues with prior unsupervised and supervised methods.

YOLOv8 is a fast and accurate AI system that finds and draws boxes around objects in pictures or videos, and can even outline them precisely.

A mathematical tool that helps computers smooth images and remove flaws by finding the fewest necessary changes to image gradients, keeping important edges sharp.

A technique to help AI systems better understand and use complex, structured information by organizing it based on both connections and meaning.

Virtual 3D spaces used to train AI models to navigate, explore, and understand complex scenes, especially for answering questions about what they perceive.

A technique using many tiny 3D shapes (Gaussians) to render realistic 3D scenes and animated characters very quickly and with high detail.

A 3DGS-based neural renderer uses a collection of optimized 3D colored 'splats' to quickly generate photorealistic images of a scene from any angle.

A method using 4D Gaussian representations and multi-resolution code-books to make digital talking faces show very precise and realistic emotions from audio.