liveBook · Manning

Drift

As explained in LLM Evaluation and Alignment, The Foundational Ideas

Distribution shift is a change between the data or behavior distribution on which a model was trained and the distribution encountered during use or optimization.

Distribution drift occurs when the data or behavior encountered during use differs from what a model or metric was trained or calibrated on. For COMET, changes in input data can degrade performance; for best-of-N sampling, selecting high-reward responses shifts outputs away from the base policy’s distribution. section 3.3.4 section 7.5.7

In production translation, a sudden drop in a low-cost proxy metric can signal drift or model degradation and prompt investigation. Best-of-N creates a different risk: a reward model calibrated on base-policy samples may not remain calibrated on the selected responses. section 3.4.3 section 7.5.7

How LLM Evaluation and Alignment, The Foundational Ideas covers it

How deep it goes
Distinguishes input-data changes from best-of-N policy shifts, but gives no general detection test or threshold.
Examples
Uses production translation metric drops and best-of-N candidate responses as drift-related examples.
What you can do after
Diagnose possible production drift by sampling expensive metrics or tracking low-cost proxy scores, then investigate sudden drops.
How it connects
Connects drift to COMET, BLEU, reward models, best-of-N sampling, and KL divergence.

Also explained in

Compare how 18 books explain drift →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of LLM Evaluation and Alignment, The Foundational Ideas. The section links point to the parts of the book it draws on.