Glossary · Term

premature exploitation

← all terms

Definition

Plain language

When an AI agent acts on what it thinks it knows before checking how the world actually works.

As stated in the literature

A characterized agent failure mode in which a model applies prior knowledge to a task before acquiring sufficient environment-specific information, often after task-only RL training has narrowed its exploratory behavior.

Why it matters: It explains why task-reward RL can make agents worse at unfamiliar environments by shrinking the exploration that uncovers local quirks.

For example, an agent uses its memorized SQL knowledge to query a custom database before checking which tables actually exist there.

Heard on the show

“They call it premature exploitation.”
Episode 052 — An Old Reinforcement Learning Tradeoff Sneaks Back Into LLM Agents

Mentioned in 1 episode

  1. 052
    An Old Reinforcement Learning Tradeoff Sneaks Back Into LLM Agents

Related terms