Inference icons

The 115 inference icons in AI & LLM, described

Attentionattention

Attention — the mechanism that weighs which inputs a model focuses on for each output.

Attention headattention-head

An attention head — one of the several viewpoints a model attends with at once.

Attention maskattention-mask

An attention mask — you may not look ahead, the causal triangle that hides future tokens.

Attention sinkattention-sink

An attention sink — the first token soaks up the attention that has nowhere else to go.

Audio denoiseaudio-denoise

A jagged waveform settling into a flat line — removing noise from an audio recording.

Batch inferencebatch-inference

Batch inference — many inputs run through a model in one offline pass.

Beam searchbeam-search

Beam search — keep several candidate sequences alive and pick the best path at the end.

Byte pairbyte-pair

Byte-pair encoding — two pieces that keep appearing together are merged into one token.

Cite sourcescite-sources

Text with a superscript reference mark — an answer that cites where its claims came from.

Code interpretercode-interpreter

A play button between angle brackets — the sandbox where a model runs the code it writes.

Confidenceconfidence

Confidence — how sure the model is about its answer, a certainty score.

Confidence scoreconfidence-score

A check beside a set of score bars — how confident the model is in its answer.

Context extendcontext-extend

Context extend — the brackets pushed further apart, a longer window.

Context lengthcontext-length

Context length — how much fits between the brackets, the window in tokens.

Context scratchcontext-scratch

A scribble held between brackets — scratch space inside the context window for working notes.

Contradictioncontradiction

Two arrows meeting head-on — statements that cannot both be true, flagged as a contradiction.

Cross-attentioncross-attention

Cross-attention — each side looking at the other, the decoder attending over the encoder.

Decoderdecoder

A decoder — turns an internal representation back into an output.

Detokenisedetokenize

Detokenise — the token pieces assembled back into a sentence.

Text diffdiff-text

Lines of text marked with plus and minus — a diff of what the model added and removed.

Draft and verifydraft-verify

Draft and verify — the small model guesses and the big one signs off.

Embedding compareembedding-compare

Embedding compare — how close two meanings sit, the cosine between a pair of vectors.

Embedding driftembedding-drift

Embedding drift — the same word slowly moving house as its vector shifts over time.

Expert routerexpert-router

An expert router — this token goes to that specialist in a mixture of experts.

Face detectionface-detect

A face between the two corner brackets that found it — face detection and facial recognition.

Fast thinkingfast-thinking

A lightning bolt inside a ring — the quick, intuitive answer given without deliberation.

Gaussian splatgaussian-splat

A scene rendered from soft overlapping dots — 3D Gaussian splatting and radiance fields.

Greedy decodinggreedy

Greedy decoding — always take the likeliest next token, deterministic and top-one.

Greedy decodegreedy-decode

Greedy decoding — the straight road, the top token at every single step.

Image captionimage-caption

A picture with the model's description written beneath it — image captioning and alt text.

Image variationimage-variation

A picture with alternative takes beneath it — variations generated from one image.

Inferenceinference

Inference — running a trained model to produce a prediction, a completion or a generated output.

Batch inferenceinference-batch

Batch inference — many answers produced together in one bulk, offline pass.

Fast inferenceinference-fast

Fast inference — the answer delivered quickly, low latency at speed.

Knowledge cutoffknowledge-cutoff

A timeline that stops at a bar, with nothing known beyond it — a model's knowledge cutoff.

KV cachekv-cache

The KV cache — attention keys and values the model already worked out, reused for speed.

KV evictkv-evict

KV evict — the attention cache lets its oldest entries go to make room.

Latencylatency

Latency — how long you wait for a response, the delay between request and answer.

Lip synclip-sync

A mouth with sound waves beside it — matching lip movement to generated or dubbed speech.

LLM chatllm-chat

LLM chat — the model talking back in a conversation, reply by reply.

Logit biaslogit-bias

Logit bias — a thumb pressed on one token's scale to steer what the model says.

Logit lenslogit-lens

A logit lens — a peek at what the model is thinking one layer early.

Logitslogits

Logits — the raw scores a model produces before softmax turns them into probabilities.

Logprobslogprobs

Logprobs — how sure the model was about each token, log probabilities per token.

Mesh generationmesh-gen

A wireframe of joined triangles conjured from a description — 3D mesh and geometry generation.

Min-pmin-p

Min-p — tokens below the probability floor do not exist for sampling.

Model filtermodel-filter

A funnel beside a model core — requests filtered before they reach a model.

Model latencymodel-latency

A model core with a clock inside — how long a model takes to answer.

Model messagemodel-message

A speech bubble with a model core inside — a message written by the model.

Model runmodel-run

A model core beside a play button — run an inference with this model.

Model textmodel-text

A model core beside lines of text — the text a model produces.

Model voicemodel-voice

A model core with sound waves leaving it — a model that speaks.

Music generationmusic-gen

A single musical note for a tune nobody hummed first — AI music and audio generation.

Needle in a haystackneedle-haystack

One point hidden among many lines — the needle-in-a-haystack test of long-context recall.

Negative promptnegative-prompt

A speech bubble holding a minus — the negative prompt that tells a model what to leave out.

Radiance fieldnerf

Rays converging on a point and leaving as one — a neural radiance field rendering a view.

Object detectionobject-detect

A labelled bounding box drawn around the thing the model found — object detection in an image.

OCRocr-ai

A scan line sweeping across a page of text — optical character recognition that reads documents.

Outline generationoutline-gen

Headings and nested points stepping inward — an outline generated before the prose is written.

Outpaintoutpaint

A picture with an open frame extending past its edge — outpainting beyond the original canvas.

Parametersparameters

Parameters — the knobs and settings you set before a model or a job runs.

Paraphraseparaphrase

Two lines of text exchanging places — rewriting a passage in different words.

Partial answerpartial-answer

Text that trails off into an ellipsis — a partial or truncated model answer that stopped early.

Physics simulationphysics-sim

A ball above a ramp and a model that knows what happens next — physics and world simulation.

Playgroundplayground

A window with a model core inside — the playground where you try a model by hand.

Point cloudpoint-cloud

Six points hanging in space — a 3D point cloud from a scanner or depth model.

Poolingpooling

Pooling — many token vectors folded into one by averaging or taking the max.

Pose estimationpose-estimate

The stick-figure skeleton a model sees inside a person — human pose estimation and keypoints.

Positional encodingpositional-encoding

Positional encoding — every token knows its seat in the sequence through a wave of positions.

Probability curveprobability-curve

A probability curve — where the mass of belief sits across the possible next tokens.

Reasoning budgetreasoning-budget

A coin held between brackets — the token budget a model may spend on thinking.

Reasoning tracereasoning-trace

Lines of thought stepping down and to the right — the visible trace of a model's reasoning.

Retryretry

Retry — go round and try the same operation again after a failure.

Robot armrobot-arm

A jointed arm with an open gripper — embodied AI, robotics and manipulation policies.

Sample consensussample-consensus

Several sampled answers and one check for the answer they agree on — self-consistency voting.

Samplingsampling

Sampling — pick the next token from the candidates by their probabilities.

Seedseed

A seed sending up a sprout with two leaves — the random seed a generation grows from.

Segmentation masksegment-mask

The exact outline of an object inside the frame, not just its box — image segmentation masks.

Self-consistencyself-consistency

Self-consistency — ask several times and keep the answer most of the samples agreed on.

Self-reflectself-reflect

An agent above its own reflection — a model reviewing its answer before it commits.

Softmaxsoftmax

Softmax — raw scores turned into probabilities where one wins.

Sound effectsound-effect

A lightning bolt with sound waves beside it — generated sound effects and foley.

Speaker diarizationspeaker-diarize

Two speakers above one shared waveform — telling apart who said what in a recording.

Special tokenspecial-token

A special token — a control marker in the sequence that is not a word.

Speculative decodingspeculative-decode

Speculative decoding — a small draft model races ahead and the big model verifies its tokens.

Speech recognitionspeech-recognize

Sound waves turning into lines of text — automatic speech recognition and speech-to-text.

Speech synthesisspeech-synth

Lines of text turning into sound waves — text-to-speech and voice synthesis.

Stop reasonstop-reason

Text ending in a square stop mark — why the model stopped generating.

Stop sequencestop-sequence

A stop sequence — the words that end the generation when the model produces them.

Streaming responsestreaming-response

A streaming response — the model's answer arriving token by token as it is written.

Structured outputstructured-output

Structured output — an answer returned in the JSON schema or type you asked for.

Style transferstyle-transfer

A picture with a brushstroke sweeping across its lower half — one image painted in another's style.

Tab completetab-complete

Text with a tab arrow at the end of the last line — accept an inline suggestion with Tab.

Temperaturetemperature

Temperature, the sampling setting that controls how much randomness a model's output is allowed.

High temperaturetemperature-high

High temperature — let the model gamble a little, more random and creative sampling.

Low temperaturetemperature-low

Low temperature — say the likeliest thing every time, precise and safe sampling.

Text to videotext-to-video

Lines of text pointing at a strip of film — text-to-video generation from a prompt.

Texture generationtexture-gen

A square filled with diagonal hatching — a tileable texture generated for a surface.

Think longerthink-longer

A timer dial with thinking dots on its face — give the model more time to answer.

Throughputthroughput

Throughput — how much work gets through per second, the rate or capacity of a system.

Token ceilingtoken-ceiling

A token ceiling — the stack of tokens meets the lid, the maximum output length.

Token intoken-in

Token in — one more word entering the model, a prompt token on its way in.

Token outtoken-out

Token out — one more word leaving the model, a completion token on its way out.

Tokens usedtokens-used

A pill holding three tokens — the count of tokens a request consumed.

Top-ptop-p

Top-p — nucleus sampling that keeps just enough choices to cover the odds.

Live transcriptiontranscribe-live

Lines of text above a live waveform — speech transcribed as it is spoken.

Speech translationtranslate-speech

Sound waves passing through an arrow into new sound waves — speech translated in real time.

Tree of thoughttree-of-thought

Tree of thought — several lines of reasoning explored as branches, with the best one kept.

Text to imagetxt2img

Lines of text with an arrow leading to a picture — text-to-image generation from a prompt.

Uncertaintyuncertainty

A point floating above a wavering line — the uncertainty a model should admit in its answer.

Upscaleupscale-ai

A small image expanding toward the corner — AI upscaling and super-resolution enhancement.

Video framevideo-frame-ai

A frame of film between its two rails, ready to play — video understanding one frame at a time.

Visual groundingvisual-grounding

A location pin dropped inside a picture — grounding words to the exact pixels they refer to.

Vocabularyvocabulary

Vocabulary — every token piece the model can say, the full list.

Voice clonevoice-clone

Two identical sets of sound waves — a voice cloned from a short sample.

In code, each is one import — import { Attention } from "@iconmind/react/icons/attention" — and the same name in Vue, Svelte, Solid, Preact, React Native, Astro, Blade and Flutter.

Narrow it by tag