Classifier

The paper treats conditional samples as the positive class and the particles in the diffusion process as unlabeled data, training a lightweight classifier online to estimate the probability that each particle belongs to an undesirable condition.

When the classifier call fails, the message is treated as in-scope and handed to the answering model, rather than staying silent.

Defenses against prompt injection (classifiers, instruction hierarchies, source labels) are without exception probabilistic; the model may be bypassed by novel phrasing, and classifiers may misjudge legitimate input.

A classifier maps input to one of a finite set of predefined categories with a fixed, closed candidate space; a generative language model predicts the next token from the entire vocabulary at each step, with an open candidate space, building output autoregressively.

Detectors are responsible for locating targets in images or videos and outputting bounding boxes and confidence scores, while classifiers determine the species or category of the target.

Locally-run classifiers can operate on ordinary laptops, servers, or even edge devices without relying on GPU resources or cloud GPU instances.

Classification tasks such as sentiment analysis, spam filtering, intent recognition, and content tagging do not require the complex reasoning capabilities of generative large models, making lightweight classifiers more efficient and controllable.

The classifier's default decision threshold systematically inflates salience estimates at the corpus level.