Definition
Plain language
A training method that rewards a model for actually doing what it says it does, by checking its stated principles against its real behavior.
As stated in the literature
Self-Consistency Training with Reinforcement Learning, an MIT framework that scores agreement between a model's meta-level self-description and its object-level behavior generated in separate contexts, then uses RL to align them via explanation training (transparency) or behavior training (alignment), tunable along a single dial.
Why it matters: It pushes models to practice what they preach, making their self-descriptions trustworthy guides to how they'll actually act.
For example, a model that claims it always refuses violent requests is checked against its actual responses, and rewarded when its stated rule and its real behavior line up.
Heard on the show
“It's called "Self-CTRL: Self-Consistency Training with Reinforcement Learning," out of MIT CSAIL, and it went up on arXiv on June sixteenth, twenty-twenty-six; we're recording two days later.”Episode 152 — Training a Model to Mean What It Says, And Why That Isn't the Same as Being Good