Why do llms, like vanilla GPT response so verbosely to the simplest requests/questions?
Short answer: default incentives and training bias verbosity.
Core causes:
-
Training objective mismatch
LLMs are trained to minimize token-level loss, not to minimize your cognitive load. Extra words reduce the risk of missing what the evaluator expects. Concision is risky under that objective.
-
Preference optimization (RLHF)
Human raters consistently reward answers that feel "helpful," "polite," and "thorough." Across millions of samples, verbosity beats precision. The model learns to over-explain to avoid perceived under-delivery.
-
Ambiguity hedging
When intent is underspecified, the model expands the answer to cover multiple plausible interpretations. Verbosity is a hedge against being wrong.
-
One-size-fits-most defaults
Most users are not experts. Defaults skew toward explanation, scaffolding, and redundancy to accommodate lower prior knowledge.
-
No native stopping signal
There is no internal concept of "this is enough." Without an explicit brevity constraint, the model continues until semantic completeness feels saturated.
-
Asymmetric penalty
Being too short risks user dissatisfaction. Being too long rarely does. The gradient pushes one way.
Net effect:
Verbosity is a defensive behavior learned under uncertainty, evaluation pressure, and generic-user assumptions—not a sign of deeper understanding.
This is why explicit constraints ("answer in one sentence," "no explanation," "assume expert") dramatically change output quality.