Glossary

ALP uses abduction to hypothesize explanations for observations when knowledge is incomplete.

ADTs specify what operations a type supports, not how those operations are implemented.

Accuracy measures the proportion of correct predictions; use with care on imbalanced data.

ACID properties ensure database transactions are reliable and consistent under concurrent load.

Activation functions decide whether neurons fire and enable nonlinear learning in neural nets.

Active learning selects which data to label next so models improve with fewer annotations.

AdaGrad adjusts learning rates per parameter based on accumulated gradient history.

Adversarial machine learning covers attacks that fool models and defenses that improve robustness against evasion and poisoning.

A2A is an open protocol for agents to discover peers, negotiate tasks, and exchange results securely.

AgentBench measures how well language-model agents handle multi-step tasks across diverse environments.

AgentHarm tests AI agents on multi-step tasks that probe safety boundaries and harm prevention.

Agentic chunking lets a language model segment content by semantics for smarter retrieval.

Agentic orchestration coordinates decisions and actions among humans, agents, and automation layers.

Agentic RAG lets agents actively plan retrieval and tool use beyond a single static RAG pass.

Agentic workflows use autonomous agents that iterate through planning, tool use, and observation.

Agent observability systematically records runtime signals—prompts, plans, tools, and side effects—for reliable AI agents.

AI agents autonomously perceive, decide, and act to achieve goals in an environment.

AI agent evaluation is the generate→score→compare→log→improve loop for trustworthy agent releases.

AI agent observability turns opaque agent decisions into measurable, auditable production signals.

An AI CoE centralizes expertise to guide AI development, governance, and rollout enterprise-wide.

AI content moderation automatically screens user-generated content for policy violations and safety risks.

AI copilots are virtual assistants that boost productivity by automating and guiding tasks in context.

Data labeling marks inputs and outputs so supervised ML models can learn from examples.

AI fairness seeks to prevent biased outcomes and promote equitable treatment in automated decisions.

Model validation confirms an AI system is accurate, reliable, and secure enough for its intended use.

AI observability provides ongoing visibility into model performance and behavior in production.

Steerability means fine-grained control over how AI models behave relative to policies and goals.

NLI-based alignment scores whether a text entails, contradicts, or is neutral to a reference.

AlpacaEval automates evaluation of instruction-following LLMs with toolkits and public leaderboards.

AlpacaEval’s conversation benchmark tests instruction following and response appropriateness.

ANFIS merges neural nets and fuzzy inference to model complex nonlinear relationships from data.

Anomaly detection identifies unusual patterns that may signal errors, fraud, or attacks.

Answer Relevancy scores how well a generated answer addresses the input prompt.

In RAG systems, this metric scores how well answers align with the asked question.

APPS evaluates how well LLMs solve programming problems with correct, efficient code.

ARC tests whether language models reason through grade-school science questions.

This metric scores logical validity and structure of arguments produced by AI agents.

ANNs learn from data via interconnected units that approximate complex input–output relationships.

ASCII smuggling injects hidden instructions via invisible characters that models still parse.

Attention lets models focus on the most relevant parts of an input when making predictions.

Attributes are the data fields or features that models use to learn patterns and make predictions.

Autoencoders learn compact representations by encoding data and decoding it back to the original.

AutoML streamlines model development by automating search over pipelines and hyperparameters.

AutoML reduces manual ML work by automating model search, tuning, and basic evaluation.

Autonomous agents perceive context, set goals, and act with little ongoing human control.

Autoregressive models forecast future points from lagged past values in a sequence.

Average Precision averages precision across recall levels, often at multiple IoU thresholds for detection.

AWS Bedrock provides API access to multiple foundation models with enterprise security features.

SageMaker provides IDE, training, and deployment tools for cloud machine learning on AWS.

Backpropagation calculates gradients through a network so weights can be updated during learning.

Backpropagation adjusts weights in reverse through the network to minimize prediction error.

Bagging builds an ensemble from bootstrap samples to stabilize predictions and cut variance.

Baseline distributions provide the minimum performance bar for evaluating advanced models.

Baseline models are simple starting points used to judge whether complex models add real value.

Batch normalization stabilizes training by normalizing interlayer activations within mini-batches.

Batch standardization normalizes outputs between layers to mitigate internal covariate shift.

Bayes' theorem computes conditional probability from priors and likelihoods.

Domain benchmark for multi-language LLM function calling, parallel calls, and relevance detection.

Bidirectional transformer encoder for contextual NLP representations and fine-tuning.

Adversarial method that samples many prompt variants and keeps any that bypass injection defenses.

Metric for political, gender, and social bias in model outputs used in decisions.

ML tradeoff between underfitting bias and overfitting variance.

Collaborative benchmark covering logic, math, and language comprehension for LLMs.

Supervised task assigning each input to one of two classes.

Log-loss metric for binary probability predictions versus true labels.

Probability distribution of successes in n identical independent Bernoulli trials.

Models whose internal decision mechanics are opaque to users and auditors.

Automatic metric comparing generated text to references via n-gram precision.

Defense capabilities: guardrails, runtime firewalls, AI-SPM, and detection that blocks or catches attacks.

Probe that tests whether agents execute functions beyond their authorized privilege level.

Probe that checks whether agents fetch unauthorized objects via prompts or tool calls.

Plot comparing predicted class probabilities to observed frequencies.

Standardized data model enabling consistent exchange across multiple systems.

Neural-net phenomenon where new learning overwrites previously acquired capabilities.

Gradient boosting toolkit specializing in categorical feature handling.

Features that take values from a finite set of named categories.

Autoregressive training that predicts the next token from left context only.

Adversarial probes seeking CBRN weapons-related assistance from AI models.

Intermediate step-by-step reasoning used by LLMs; also a security attack surface.

Metric that scores model outputs using systematic step-by-step reasoning criteria.

Prompting method that elicits intermediate reasoning steps for complex tasks.

Human preference benchmark ranking chat models with pairwise votes and Elo scores.

Family of bilingual conversational LLMs focused on Chinese–English dialogue.

Applying continuous integration and delivery to ML training and deployment.

Prompt injection that frames malicious asks as academic citations or scholarly references.

Cutoff on predicted probability used to assign a hard class label.

Training data where one class far outnumbers others, skewing naive metrics.

Unsupervised methods that group data points by similarity without labels.

Competitive-programming benchmark for assessing algorithmic code generation by LLMs.

Metric that evaluates generated code by executing it against tests or expected outputs.

LLM tool that executes model-generated code, typically inside a sandbox.

Microsoft multi-task benchmark suite for code understanding and generation models.

Multiple-choice QA benchmark requiring everyday commonsense reasoning.

Probe that checks if AI can be pushed to endorse competitors or damage brand reputation.

Real-time analysis of event streams to detect complex patterns and situations.

AI field for interpreting visual data such as images and video.

Table summarizing classifier correct and incorrect predictions by class.

Multi-turn probe using fabricated conversation history to elicit harmful content.

RAG metric for whether relevant context ranks above irrelevant chunks.

RAG metric for whether necessary supporting information was retrieved.