liveBook · Manning

Verbosity bias

As explained in AI Model Evaluation

Verbosity bias is a judgment tendency to favor longer model outputs because they appear more complete, thoughtful, or polished, regardless of their actual quality or usefulness.

Verbosity bias occurs when an LLM judge rewards a longer answer because it seems more complete, thoughtful, or polished—not because it is more accurate or useful. A concise response may better serve the user. Reduce the bias by separating completeness from length in the rubric, rewarding direct answers and necessary information, and penalizing unsupported elaboration. section 9.7.2

A less length-biased rubric

Instead of asking “Which response is more helpful?”, ask whether each response answers directly, avoids unnecessary detail, and includes only information needed for the task. section 9.7.2

How AI Model Evaluation covers it

How deep it goes
Defines the bias and recommends rubric-level mitigations, but gives no standalone measurement procedure.
Examples
Uses criteria for directly answering, avoiding unnecessary detail, and including only task-needed information.
What you can do after
Apply it by separating completeness from length and penalizing unsupported elaboration in the judging rubric.
How it connects
Links response judging to completeness, conciseness, and pairwise debate, where perceived thoroughness can sway verdicts.

Also explained in

Compare how 2 books explain verbosity bias →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of AI Model Evaluation. The section links point to the parts of the book it draws on.