Model drift
Output behavior changing over time because weights, prompts, tools, safety rules, retrieval, or serving infrastructure changed.
What Model drift does not show
A changed answer does not identify which layer drifted.
Terms this definition uses
Related terms in Models and generation
How a model produces text, what its training and decoding controls change, and what none of them can prove about a source.
Autoregressive model · Transformer · Attention · Self-attention · Parameter · Inference · Pretraining · Fine-tuning · Supervised fine-tuning · Instruction tuning · RLHF · Preference optimization · Knowledge distillation · Quantization · Mixture of experts · Multimodal model · Vision-language model · Small language model · Tokenizer · Temperature · Top-p sampling · Deterministic decoding · System prompt · Developer message · Tool call · Function calling · Structured output · Constrained decoding · Chain of thought · Reasoning trace · Context caching · Context compaction · Model routing · Model version
Questions about Model drift
What is Model drift?
Output behavior changing over time because weights, prompts, tools, safety rules, retrieval, or serving infrastructure changed.
What does Model drift not show or guarantee?
A changed answer does not identify which layer drifted.
Which area of the glossary does Model drift belong to?
Models and generation: How a model produces text, what its training and decoding controls change, and what none of them can prove about a source.
← Model routing · Model version →
See it in the full glossary · 731 terms across 19 areas.