Glossary

~220 terms, one crisp line each, with links to the page that explains them properly. Organized alphabetically within thematic groups for browsability; use Ctrl-F for lookup.

Jump to: A–C · D–F · G–L · M–P · Q–S · T–Z

---

A–C

Term	Definition
Ablation	removing a component to measure its contribution.
Activation checkpointing	recompute activations in the backward pass instead of storing them; ~33% more compute for \(O(\sqrt L)\) memory. → NN refresher
Active parameters	parameters used per token in an MoE — distinct from total parameters. → MoE
AdaLN-Zero	conditioning via LayerNorm scale/shift, residual branch zero-initialized; used in DiT.
AdamW	Adam with decoupled weight decay. The standard optimizer. → Optimization
ALiBi	positional method adding a linear distance penalty to attention scores. → Positional encoding
Alignment	making a model behave as intended: helpful, honest, harmless. → Alignment
Alignment tax	capability loss on benchmarks caused by alignment training.
ANN	approximate nearest neighbour search. → Vector search
Attention	weighted average of values, weights from query–key similarity. → Attention
Attention sink	a token (usually the first) that absorbs excess attention mass because softmax must sum to 1.
Autoregressive	\(p(x)=\prod_i p(x_i\mid x_{<i})\). → AR models
AWQ	activation-aware 4-bit weight quantization. → Efficiency
Backpropagation	chain rule with memoization; backward ≈ 2× forward cost.
Base model	a pretrained model before instruction tuning or alignment.
Batch size (critical)	the point past which larger batches stop reducing step count proportionally.
Beam search	keep the \(B\) best partial sequences; good for constrained tasks, bad for open generation.
BERTScore	text similarity via contextual embedding matching. → Metrics
BF16	bfloat16: FP32's exponent range with 7 mantissa bits. The training default.
Bias (statistical)	systematic error; in generative AI, also representational harm. → Societal impact
BitNet	models trained natively with ternary \(\{-1,0,1\}\) weights.
BLEU	n-gram precision metric for translation. Cannot see meaning.
BM25	classic sparse lexical retrieval scoring function. → Vector search
BPC / BPB	bits per character / per byte; the tokenizer-independent way to compare LMs.
BPE	Byte-Pair Encoding: greedily merge the most frequent adjacent pair. → Tokenization
Bradley–Terry	model turning pairwise preferences into a scalar score; the basis of reward models.
C2PA	signed content-provenance standard.
Calibration	whether stated confidence matches actual accuracy. RLHF degrades it.
Capacity factor	per-expert token buffer size in MoE; overflow tokens are dropped.
Catastrophic forgetting	losing prior capability during fine-tuning.
Causal mask	prevents attending to future positions.
CFG	classifier-free guidance: \(\epsilon(\varnothing) + w(\epsilon(c)-\epsilon(\varnothing))\). → Latent diffusion
Chain-of-thought (CoT)	intermediate reasoning tokens; converts sequence length into computational depth. → Reasoning
Chinchilla	the compute-optimal scaling result: ~20 tokens per parameter. → Scaling laws
CLIP	contrastively trained joint image–text embedding model. → Multimodal
Consistency model	maps any point on a diffusion trajectory directly to its endpoint → 1-step generation.
Constitutional AI	alignment using AI feedback guided by explicit written principles.
Constrained decoding	mask invalid tokens to guarantee grammar conformance. → Structured output
Contamination	benchmark test data leaking into training data. → Benchmarks
Context window	maximum sequence length the model can process.
Contextual retrieval	prepending an LLM-written situating sentence to each chunk before embedding.
Contrastive learning	pull positives together, push negatives apart (InfoNCE).
ControlNet	spatial conditioning via a trainable copy of the encoder with zero-init connections.
Cross-attention	Q from one stream, K/V from another; the universal conditioning mechanism.
Cross-encoder	scores a (query, document) pair jointly; accurate, not precomputable.
Cross-entropy	\(-\sum p\log q\); the standard training loss. → Info theory
Curse of dimensionality	data requirements grow exponentially with dimension.

D–F

Term	Definition
DDIM	deterministic diffusion sampler; a discretization of the probability-flow ODE.
DDPM	denoising diffusion probabilistic model. → Diffusion
Decoder-only	the modern LLM architecture: causal attention, one homogeneous stack.
Decoding	turning a probability distribution into tokens. → Inference
Deduplication	removing duplicate training documents; improves loss and reduces memorization.
Dequantization	adding noise to discrete data so a continuous density model is well-defined.
Diffusion model	learns to reverse a gradual noising process.
Distillation	training a small student to match a large teacher.
DiT	Diffusion Transformer — a Transformer backbone replacing the U-Net.
DoRA	LoRA variant decomposing weights into magnitude and direction.
DPO	Direct Preference Optimization: alignment without a reward model. → Alignment
Dropout	random unit zeroing; mostly absent from modern pretraining.
EBM	energy-based model: \(p(x)\propto e^{-E(x)}\). → EBM
ECE	expected calibration error.
Elo	rating system used by preference arenas; 100 points ≈ 64% win rate.
ELBO	evidence lower bound: \(\log p(x) = \text{ELBO} + \text{KL}(q\|p(z\mid x))\).
Embedding	a vector representation where semantic similarity is geometric proximity.
Emergence	capabilities appearing abruptly with scale; partly a metric artifact.
Encoder-only	bidirectional architecture (BERT); good at understanding, can't generate.
Entropy	\(-\sum p\log p\); the irreducible number of bits needed to describe a distribution.
Exposure bias	train/inference mismatch: teacher forcing vs the model's own outputs.
Expert parallelism	distributing MoE experts across devices; requires all-to-all communication.
Faithfulness	(a) whether generated claims are supported by context; (b) whether stated reasoning reflects actual computation.
Few-shot	including examples in the prompt.
FID	Fréchet Inception Distance; the standard image-generation metric, with real caveats.
FlashAttention	exact attention with \(O(T)\) memory via tiling and online softmax.
Flow matching	regress onto a straight-line velocity field; the modern diffusion alternative. → Flow matching
FP8	8-bit float; used on recent hardware for training and inference.
Free bits	a KL floor per latent dimension; prevents posterior collapse.
FSDP	Fully Sharded Data Parallel — PyTorch's ZeRO-3.

G–L

Term	Definition
GAN	generator vs discriminator minimax game. → GAN
GELU	\(x\Phi(x)\); smooth activation used in BERT and GPT-2/3.
Goodhart's law	when a measure becomes a target, it ceases to be a good measure.
GPTQ	4-bit weight quantization with Hessian-based error compensation.
GQA	grouped-query attention; fewer K/V heads than Q heads → 8× smaller KV cache.
Gradient clipping	cap the global gradient norm (typically 1.0). Essential.
GraphRAG	RAG over an entity/relation graph; handles corpus-wide questions.
GRPO	Group Relative Policy Optimization; PPO without a value model.
Guidance scale	the \(w\) in CFG; trades diversity for prompt fidelity.
Hallucination	fluent, confident, false output. Structural, not a mere bug. → Safety
Hard negative	a plausible-looking but incorrect training negative; what makes embeddings good.
HNSW	hierarchical navigable small-world graph; the quality-default ANN index.
HyDE	embed a hypothetical answer rather than the query, to search in document-space.
In-context learning	learning from examples in the prompt, without weight updates.
Induction head	attention circuit implementing "[A][B]…[A]→[B]"; linked to in-context learning.
InfoNCE	the contrastive loss; bounds mutual information by \(\log N\) (batch size).
Instruction tuning	supervised fine-tuning on (instruction, response) pairs.
IVF-PQ	inverted file + product quantization; ~32× index compression.
Jailbreak	eliciting behaviour the model was trained to refuse.
JSD	Jensen–Shannon divergence; the symmetric, bounded relative of KL.
KL divergence	\(\sum p\log(p/q)\); excess bits from using the wrong model. Asymmetric.
KV cache	stored keys and values for generated tokens; the real inference memory constraint.
Langevin dynamics	noisy gradient ascent on log-density; samples using only the score.
Latent diffusion	diffusion in a VAE's compressed latent space. → Latent diffusion
LayerNorm	normalize across features per token.
Lethal trifecta	private data + untrusted content + external communication ⇒ exfiltration risk.
LLM-as-judge	using a strong model to evaluate outputs; ~80–85% human agreement if debiased.
LoRA	\(W' = W_0 + \frac{\alpha}{r}BA\); low-rank fine-tuning that merges at inference. → PEFT
Lost in the middle	information in the middle of a long context is used poorly.

M–P

Term	Definition
Mamba	selective state-space model; linear-time training, \(O(1)\) inference.
Manifold hypothesis	real data occupies a low-dimensional manifold in its ambient space.
Matryoshka embedding	truncatable embeddings — the first \(k\) dimensions are still valid.
Memorization	verbatim reproduction of training data; scales with duplicates and model size.
MFU	model FLOPs utilization; 40–50% is good for large runs.
MLA	multi-head latent attention; compresses the KV cache into a shared low-rank latent.
Mode collapse	a generator producing limited variety; the classic GAN failure.
MoE	mixture of experts: route each token to top-\(k\) of \(E\) expert MLPs. → MoE
Monte Carlo	estimating expectations by sampling.
MQA	multi-query attention: one shared K/V head. Max cache savings, some quality loss.
Mutual information	\(I(X;Y)\): bits that knowing \(Y\) saves in describing \(X\).
nDCG	position-weighted, graded retrieval quality metric.
NF4	4-bit NormalFloat; quantile-spaced levels matched to a normal distribution.
Normalizing flow	invertible map with a tractable Jacobian → exact likelihood. → Flows
Nucleus sampling	top-\(p\): keep the smallest token set with cumulative probability \(\ge p\).
ODE (probability flow)	the deterministic ODE with the same marginals as a diffusion SDE.
ORM / PRM	outcome / process reward model — final-answer vs per-step supervision.
Over-refusal	refusing benign requests; must be measured alongside attack success.
PagedAttention	OS-style paging for the KV cache; cuts waste from ~70% to <4%.
pass@k	probability that at least one of \(k\) samples passes; use the unbiased estimator.
PEFT	parameter-efficient fine-tuning.
Perplexity	\(\exp(\text{cross-entropy})\); not comparable across tokenizers.
PII	personally identifiable information.
Positional encoding	injecting order information; attention alone is permutation-equivariant.
Posterior collapse	the VAE latent carrying no information (\(I(x;z)\to 0\)).
PPO	Proximal Policy Optimization; the RL algorithm in classic RLHF.
Pre-norm	normalization inside the residual branch; enables training at depth.
Prefill	processing the prompt; compute-bound, unlike decode.
Prompt caching	caching KV states for a reused prefix; large cost and latency win.
Prompt injection	untrusted content acting as instructions. No complete solution exists. → Security

Q–S

Term	Definition
QK-norm	normalizing Q and K before the dot product; prevents attention-logit blowup.
QLoRA	LoRA on a 4-bit (NF4) frozen base; a 65B fine-tune on one GPU.
Quantization	representing weights/activations in fewer bits. 4-bit is the sweet spot.
RAG	retrieval-augmented generation. → RAG
ReAct	interleaved reasoning and tool-use loop.
Rectified flow	flow matching with straight-line interpolation; SD3 and Flux use it.
Reflow	re-couple \((x_0,x_1)\) pairs to straighten flow trajectories → 1–2 step generation.
Reparameterization trick	\(z=\mu+\sigma\epsilon\): moves randomness off the gradient path.
Reranker	a cross-encoder rescoring retrieved candidates; +5–15 nDCG.
Residual connection	\(y = x + F(x)\); makes identity the default and preserves gradients.
Residual stream	the additive \((B,T,d)\) bus that sublayers read from and write to.
Reward hacking	optimizing the proxy objective while violating its intent.
Reward model	learns human preference from pairwise comparisons; 65–75% accurate.
RLHF	reinforcement learning from human feedback.
RLVR	RL from verifiable rewards (math, code); how reasoning models are trained.
RMSNorm	LayerNorm without mean subtraction; ~10% faster, no quality loss.
RoPE	rotary position embedding; rotate Q and K so similarity depends only on distance.
ROUGE	n-gram recall metric for summarization; rewards copying.
RRF	reciprocal rank fusion; combines rankers without score calibration.
SAE	sparse autoencoder; decomposes activations into interpretable features.
Scaling law	loss as a power law in parameters, data and compute.
Score	\(\nabla_x\log p(x)\); independent of the normalizing constant.
Score matching	learning the score without knowing the density.
SDE	stochastic differential equation; the continuous-time view of diffusion.
Self-consistency	sample \(n\) chains, take the majority answer; also a confidence signal.
SFT	supervised fine-tuning.
Slerp	spherical interpolation; the correct way to interpolate high-dimensional latents.
Sliding window attention	attend only to the last \(w\) tokens; \(L\times w\) indirect reach.
SmoothQuant	migrate activation outliers into the weights so INT8 works.
Softmax	\(e^{z_i}/\sum_j e^{z_j}\); shift-invariant, temperature-controlled.
Speculative decoding	draft with a small model, verify with a large one; exact output distribution.
SSM	state-space model; linear recurrence, parallelizable by prefix scan.
Superposition	representing more features than dimensions using near-orthogonal directions.
SwiGLU	gated activation \(\text{Swish}(xW_1)\odot(xW_2)\); the modern LLM default.
Sycophancy	agreeing with the user against the evidence; a direct consequence of RLHF.

T–Z

Term	Definition
Teacher forcing	feeding ground-truth prefixes during training; enables parallel training.
Temperature	softmax scaling; 0 = greedy, ∞ = uniform.
Tensor parallelism	splitting a layer's weights across devices; must stay intra-node.
Test-time compute	inference-time compute as a second scaling axis.
Token	the discrete unit a model operates on; ~4 characters in English.
Tokenizer	text ↔ integers. Cause of arithmetic and character-counting failures.
Top-k / top-p	truncated sampling; top-\(p\) adapts to model confidence, top-\(k\) doesn't.
Transformer	attention + MLP blocks with residuals. → Transformer
Trilemma (generative)	quality, speed, coverage — pick two.
U-Net	encoder–decoder with skip connections; the classic diffusion backbone.
Unfaithful CoT	stated reasoning that doesn't reflect the actual computation.
VAE	variational autoencoder; ELBO = reconstruction − KL. → VAE
Vanishing gradient	exponential gradient decay through depth/time; what killed RNNs.
v-prediction	predicting \(v=\sqrt{\bar\alpha}\epsilon - \sqrt{1-\bar\alpha}x_0\); better-behaved at extremes.
Vector database	store + ANN index for embeddings.
ViT	Vision Transformer; treats 16×16 patches as tokens.
VQ-VAE	VAE with a discrete codebook latent; enables Transformers over images/audio/video.
Warmup	ramping the learning rate from 0; needed because early Adam variance estimates are unreliable.
Wasserstein distance	earth-mover's distance; smooth even when supports don't overlap.
Watermarking	statistical signal embedded in generated content; removable by paraphrase.
Weight decay	L2 penalty on weights; use AdamW's decoupled form, matrices only.
Weight tying	sharing the embedding and output projection matrices.
WSD schedule	warmup–stable–decay; resumable alternative to cosine.
YaRN	RoPE context-extension method combining interpolation with NTK-aware scaling.
Zero-shot	performing a task with no examples.
ZeRO	sharding optimizer states, gradients and parameters across data-parallel ranks.
z-loss	auxiliary loss on \(\log Z\); stabilizes training and helps quantization.

---

Next → Math cheatsheet