Verbosity

Verbosity is one of the most consistent complaints about AI model outputs: the model gives a long, padded response when a short, direct one would serve better. This often shows up as repeating the user’s question back to them before answering, adding lengthy preambles, summarizing what was just said, or including extensive caveats that weren’t asked for. Verbosity frequently stems from RLHF — human raters often perceive longer responses as more thorough and helpful, even when they aren’t, so models learn to pad. For behavior architects, combating verbosity means being explicit about length expectations in system prompts and training data, and building evaluations that specifically reward conciseness when the task calls for it.