Results for "output change"

Model-generated content that is fluent but unsupported by evidence or incorrect; mitigated by grounding and verification.

Raw model outputs before converting to probabilities; manipulated during decoding and calibration.

A single attention mechanism within multi-head attention.

One example included to guide output.

Temporary reasoning space (often hidden).

A formal privacy framework ensuring outputs do not reveal much about any single individual’s data contribution.

Enables external computation or lookup.

Training a smaller “student” model to mimic a larger “teacher,” often improving efficiency while retaining performance.

Reusing knowledge from a source task/domain to improve learning on a target task/domain, typically via pretrained models.

Training one model on multiple tasks simultaneously to improve generalization through shared structure.

A structured collection of examples used to train/evaluate models; quality, bias, and coverage often dominate outcomes.

The learned numeric values of a model adjusted during training to minimize a loss function.

The degree to which predicted probabilities match true frequencies (e.g., 0.8 means ~80% correct).

A proper scoring rule measuring squared error of predicted probabilities for binary outcomes.

Nonlinear functions enabling networks to approximate complex mappings; ReLU variants dominate modern DL.

Methods to set starting weights to preserve signal/gradient scales across layers.

Techniques that stabilize and speed training by normalizing activations; LayerNorm is common in Transformers.

Randomly zeroing activations during training to reduce co-adaptation and overfitting.

An RNN variant using gates to mitigate vanishing gradients and capture longer context.

Stepwise reasoning patterns that can improve multi-step tasks; often handled implicitly or summarized for safety/privacy.

Architecture that retrieves relevant documents (e.g., from a vector DB) and conditions generation on them to reduce hallucinations.

A preference-based training method optimizing policies directly from pairwise comparisons without explicit RL loops.

Model trained to predict human preferences (or utility) for candidate outputs; used in RLHF-style pipelines.

Converts logits to probabilities by exponentiation and normalization; common in classification and LMs.

Attacks that infer whether specific records were in training data, or reconstruct sensitive training examples.

Coordinating tools, models, and steps (retrieval, calls, validation) to deliver reliable end-to-end behavior.

Generating speech audio from text, with control over prosody, speaker identity, and style.

Neural networks can approximate any continuous function under certain conditions.

Allows gradients to bypass layers, enabling very deep networks.

The range of functions a model can represent.