How It Works
Self-conditioning
When an AI model reads its own earlier outputs as part of its context while working through a long task, so that its mistakes feed back into its later steps. Small errors compound over many steps, which is one reason long-horizon tasks fail even when each single step looks easy.
Origin
A term from machine-learning research, given a precise modern meaning in the 2025 arXiv paper "The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs," which measures how self-conditioning amplifies errors over long tasks.