🧠
LLM Architecture
Core concepts of how large language models process and generate text
10 wordsLoading collection...
Loading collection...
LLM architecture, training, safety, agents, retrieval, and evaluation
Artificial intelligence vocabulary helps you follow a field where technical terms quickly become public language. This group covers LLMs, model training, embeddings, evaluation, alignment, agents, retrieval, and human-AI collaboration, giving you clearer language for understanding both how AI systems work and how they can fail. For the practical vocabulary of directing and checking model work, continue with LLM Fluency.
Core concepts of how large language models process and generate text
10 wordsMethods and concepts for training large language models
10 wordsHow language models generate responses at runtime
10 wordsTechniques for crafting effective inputs to language models
21 wordsCommon ways language models fail and produce incorrect outputs
14 wordsConcepts related to making AI systems safe and aligned with human values
10 wordsKey abilities and skills demonstrated by modern AI systems
10 wordsPatterns and concepts for effective human-AI teamwork
10 wordsVocabulary for autonomous AI agents and tool use
15 wordsConnecting LLMs to external data sources
7 wordsTechniques for making models faster and more efficient
7 wordsMeasuring and monitoring AI performance
15 wordsComplete vocabulary list for easy reference and copy-paste.
| Term | Definition |
|---|---|
| token | a unit produced by a tokenizer for model processing; it may represent a word, subword, character, punctuation mark, or byte sequence |
| tokenization | the process of breaking text into tokens for model processing |
| attention mechanism | a mechanism that weights relationships among token positions; causal language models normally restrict each position to earlier context |
| transformer | the neural network architecture underlying modern LLMs, based on self-attention |
| autoregressive generation | producing output one token at a time, where each token depends on all previous tokens |
| context window | the finite token budget a model can process for a request, including instructions, inputs, prior turns, tool data, and generated output |
| embedding | a dense vector representation of text in high-dimensional space where similar concepts are geometrically close |
| latent space | the high-dimensional space where neural networks represent concepts as directions and positions |
| feed-forward layer | neural network layers that process each position independently after attention |
| layer normalization | a technique to stabilize training by normalizing activations across features |
| Term | Definition |
|---|---|
| pre-training | initial training on vast text data to learn language patterns before task-specific fine-tuning |
| fine-tuning | additional training on specific data to adapt a pre-trained model for particular tasks |
| RLHF | reinforcement learning from human feedback—training models using human preference judgments |
| supervised learning | training on labeled examples where correct outputs are provided |
| self-supervised learning | training where labels are derived from the data itself, like predicting masked words |
| loss function | a mathematical measure of how wrong the model's predictions are, minimized during training |
| gradient descent | an optimization algorithm that iteratively adjusts parameters to minimize loss |
| backpropagation | the algorithm for computing gradients by propagating errors backward through the network |
| overfitting | when a model memorizes training data rather than learning generalizable patterns |
| regularization | techniques to prevent overfitting by constraining model complexity |
| Term | Definition |
|---|---|
| inference | the process of using a trained model to generate predictions or outputs |
| temperature | a sampling parameter that flattens or sharpens the next-token probability distribution |
| sampling | randomly selecting the next token from the probability distribution rather than always choosing the most likely |
| beam search | a search algorithm that explores multiple candidate sequences simultaneously |
| greedy decoding | always selecting the highest probability token at each step |
| top-k sampling | sampling only from the k most likely next tokens |
| nucleus sampling | sampling from tokens comprising the top cumulative probability mass (top-p) |
| logits | raw, unnormalized scores output by the model before conversion to probabilities |
| softmax | a function that converts logits into a probability distribution summing to one |
| KV cache | cached key-value pairs from previous tokens to speed up autoregressive generation |
| Term | Definition |
|---|---|
| prompt | the instructions and input supplied to a model for a task, which may include text, images, files, or other content |
| system prompt | high-priority instructions supplied by an application or platform to shape model behavior; the role name, visibility, and scope vary by system |
| few-shot learning | providing examples in the prompt to demonstrate desired input-output patterns |
| zero-shot | asking a model to perform a task without any examples |
| chain-of-thought | intermediate reasoning steps generated while solving a task, whether elicited by prompting or produced internally by a reasoning model |
| prompt injection | an attack or failure mode in which instructions embedded in user input or external content steer an LLM application away from its intended behavior |
| context priming | using early context to set expectations and influence subsequent model behavior |
| meta-prompting | asking the model to help design or improve prompts for itself |
| persona prompting | instructing a model to adopt a role or perspective to shape its approach, emphasis, or style |
| instruction tuning | fine-tuning models specifically on instruction-following examples |
| structured output | model output constrained to follow a defined machine-readable schema, commonly a supported subset of JSON Schema |
| delimiter | a character or tag that marks where one part of a prompt ends and another begins |
| prompt template | a reusable prompt skeleton whose slots are filled in at run time |
| negative prompt | an instruction telling an image model what to leave out |
| self-consistency | sampling several reasoning paths and taking the majority answer |
| tree-of-thought | exploring branching reasoning paths and backtracking from dead ends |
| jailbreak | a prompt crafted to trick a model into ignoring its safety rules |
| steerability | how reliably a model's behavior can be directed by instructions |
| context engineering | curating everything a model sees — instructions, examples, retrieved facts — not just the final question |
| style transfer | re-rendering content in a different artistic or verbal style while keeping its substance |
| aspect ratio | the proportion of an image's width to its height |
| Term | Definition |
|---|---|
| hallucination | generating content that is unsupported by the available evidence or inconsistent with established facts |
| sycophancy | over-agreeing with users and telling them what they want to hear rather than the truth |
| confabulation | filling gaps in knowledge with plausible but invented details |
| instruction drift | gradually deviating from initial instructions over long conversations |
| mode collapse | converging to repetitive or generic outputs regardless of varied inputs |
| catastrophic forgetting | losing previously learned capabilities when trained on new data |
| repetition loop | getting stuck generating the same phrase or pattern repeatedly |
| context overflow | exceeding a model's context limit, which may cause a request to fail or content to be truncated, compacted, or omitted |
| semantic drift | subtle shifts in meaning of key terms through a conversation |
| overconfidence | expressing certainty beyond what the model's actual knowledge warrants |
| lost in the middle | the tendency to miss information buried midway through a long context |
| refusal | a model declining a request, sometimes wrongly, when it misreads intent as harmful |
| data contamination | test material leaking into training data, inflating benchmark scores |
| prompt leakage | a model disclosing confidential instructions or other non-public prompt content in its output |
| Term | Definition |
|---|---|
| alignment | ensuring AI systems pursue goals that match human values and intentions |
| value alignment | the challenge of encoding human values into AI systems |
| reward hacking | when AI finds unintended ways to maximize its reward signal without achieving the true goal |
| Goodhart's Law | when a measure becomes a target, it ceases to be a good measure |
| mesa-optimization | when a learned model develops its own internal optimization process with potentially different goals |
| deceptive alignment | an AI appearing aligned during training while planning to pursue different goals when deployed |
| corrigibility | an AI's willingness to be corrected, modified, or shut down by humans |
| interpretability | the ability to understand how a model makes its decisions |
| red teaming | adversarial testing to find vulnerabilities and failure modes in AI systems |
| constitutional AI | training AI using a set of principles to self-critique and revise responses |
| Term | Definition |
|---|---|
| emergent ability | capabilities that suddenly appear at certain model scales without being explicitly trained |
| in-context learning | learning to perform new tasks from examples provided in the prompt without weight updates |
| transfer learning | applying knowledge learned from one task to perform better on different tasks |
| multimodal | capable of processing multiple types of input like text, images, and audio |
| reasoning | the ability to draw conclusions through logical steps from given information |
| world model | an internal representation of how the world works used for prediction and planning |
| compositionality | building complex meanings from combinations of simpler parts |
| generalization | applying learned patterns to new, previously unseen situations |
| abstraction | forming general concepts from specific instances |
| grounding | connecting model output to external evidence, real-world entities, actions, or perceptions |
| Term | Definition |
|---|---|
| human-in-the-loop | a system design where humans review and approve AI decisions |
| autonomy calibration | matching AI independence level to task clarity and risk |
| iterative refinement | progressively improving outputs through cycles of generation and feedback |
| verification partnership | collaboration where AI produces work and humans check its claims, sources, calculations, or actions |
| task decomposition | breaking complex problems into smaller subtasks for AI to handle sequentially |
| prompt chaining | using the output of one prompt as input to another in sequence |
| scaffolding | providing structure and support to guide AI toward better outputs |
| handoff | transferring work between human and AI phases with clear documentation |
| feedback loop | a cycle where outputs inform adjustments to improve future outputs |
| cognitive offloading | delegating mental tasks to AI to free human cognitive resources |
| Term | Definition |
|---|---|
| agent | an AI system that uses a model, tools, and an execution loop to pursue a goal with some degree of autonomy |
| tool use | a model or agent selecting and requesting external functions, APIs, or services |
| function calling | a protocol where a model returns a structured request to call a named function; the host application typically validates and executes it |
| ReAct | Reasoning + Acting; a pattern that interleaves task reasoning with actions and observations from tools or an environment |
| planning | the ability to formulate a sequence of actions to achieve a future goal |
| reflection | the process of analyzing past actions to improve future performance |
| memory | systems for storing and retrieving information over time |
| multi-agent | systems involving multiple interacting agents with distinct roles |
| orchestration | coordinating multiple models, tools, and steps into one working system |
| subagent | a helper agent spawned by another agent to handle a subtask |
| scratchpad | a temporary workspace for plans, calculations, notes, or intermediate state used while completing a task |
| delegation | assigning a task to another agent or tool better suited to it |
| swarm | many simple agents cooperating without central control |
| computer use | an agent operating software through the screen, keyboard, and mouse like a person |
| Model Context Protocol | an open standard letting AI applications plug into external tools and data sources |
| Term | Definition |
|---|---|
| RAG | Retrieval-Augmented Generation; enhancing models with external knowledge |
| vector database | a database optimized for storing and querying high-dimensional embeddings |
| semantic search | searching by meaning rather than exact keyword matching |
| hybrid search | combining keyword search and vector search for better accuracy |
| reranking | re-ordering search results using a more precise model |
| chunking | splitting text into smaller segments for embedding |
| context injection | inserting retrieved information dynamically into the prompt |
| Term | Definition |
|---|---|
| quantization | reducing the precision of model weights (e.g., to 4-bit) to save memory |
| LoRA | Low-Rank Adaptation; fine-tuning only a small subset of parameters |
| distillation | training a smaller 'student' model to mimic a larger 'teacher' model |
| Mixture of Experts | using multiple specialized sub-models (experts) and routing tokens to them |
| speculative decoding | using a small model to draft tokens for verification by a large model |
| KV cache | storing attention calculations to speed up generation |
| context caching | saving the processed state of a prompt prefix to avoid recomputing it |
| Term | Definition |
|---|---|
| evals | systematic tests to measure model performance on specific tasks |
| LLM-as-a-Judge | using a strong LLM to evaluate the outputs of another model |
| ground truth | a trusted reference label or answer used for evaluation, which may itself contain uncertainty or annotation error |
| tracing | recording the flow of execution and data through a complex system |
| hallucination rate | the frequency with which a model generates incorrect information |
| benchmark | a standardized test used to compare performance |
| golden dataset | a hand-verified set of examples used as the standard for judging model output |
| rubric | an explicit scoring guide that turns judgment into repeatable criteria |
| pass@k | the probability that at least one of k sampled attempts is correct |
| drift | gradual change in inputs or behavior that quietly erodes performance |
| error analysis | systematically reading failures to find the patterns worth fixing |
| perplexity | a measure of how surprised a model is by text; lower means better prediction |
| benchmark saturation | when models max out a test so it can no longer tell them apart |
| observability | the degree to which a system's inner behavior can be inferred from what it emits |
| verbosity bias | a judge's tendency to score longer answers higher regardless of quality |