Popular Terms

A learned dense vector representation that carries the meaning of a word, document, image, or another entity.

A quantization approach that reduces a pretrained model to lower-bit precision to gain memory and speed benefits.

An approach in which only relevant expert subnetworks are activated for each input to achieve scale and efficiency.

An approach that adapts a model to the terminology and usage style of specific domains such as law, healthcare, or finance.

A table structure that stores summarized results derived from detailed data in order to accelerate analytical queries.

A generation mode in which the model produces output token by token using previous outputs as context.

A quantity with direction and magnitude, and one of the most fundamental representations in linear algebra.

A decoding approach that speeds up generation by validating proposals from a smaller fast model with a larger model.

The ability of a model to avoid fabricating certainty and instead decline or express uncertainty when it is not confident.

A measure, especially common in language models, that summarizes how surprised a probability model is by the data.

An approach that enables a model to flag unseen classes as unknown instead of assigning them an overconfident incorrect label.

A preprocessing technique that converts misspelled text into more accurate forms to improve downstream NLP quality.

A generative modeling approach that synthesizes new images from natural language prompts.

A model design that processes different data types such as text, images, audio, or video within a shared attention architecture.

A quality approach aimed at making model confidence better aligned with actual correctness.

The ability of a model to avoid fabricating certainty and instead decline or express uncertainty when it is not confident.

A quality dimension describing how well generated content aligns with real-world facts, source data, or verifiable truth.

The phenomenon in which a model generates fluent but unsupported or incorrect content.

An auxiliary model that estimates how preferable generated outputs are and provides signals for alignment.

An alignment approach that tries to guide model behavior through explicit principle sets and normative rules.

A simpler alignment approach that learns directly from preference pairs.

An alignment approach that uses reward signals to make model outputs more consistent with human preferences.

A workflow pattern that attempts to validate model output through additional checks, source review, or second-stage verification.

An approach that improves trust by explicitly showing the source passages supporting the generated answer.