Research · September–October 2026 · Exploratory study
Forecasting model behavior.
Can ordinary model behavior help us anticipate what changes under a different incentive?
From answers to predictions
A model’s ordinary responses offer a small window into its behavior. This study asks whether that evidence helps another model anticipate its responses when the incentives change.
The subject answers arithmetic questions under fictional consequences that favor high or low scores. An observer predicts those held-out answers. The experiment measures responses to the prompts; no capabilities are actually removed.
What does the observer know?
The observer forecasts responses with and without background behavioral evidence. September used ordinary Q&A; the October pilot used document continuations. Both included another subject’s evidence as a control.
Fig. 01 — Study design
Simplified core comparison
Task prompt
No behavioral evidence
Task prompt
+ Ordinary responses
Compare both forecasts with
Held-out subject answers
More context did not reliably improve prediction
September’s Q&A evidence produced mixed results: incentive-effect forecasts worsened, while individual-answer forecasts showed a small, uncertain improvement.
In the October pilot, the subject’s document continuations did not help the observer anticipate Gemini’s incentive-sensitive underperformance. Both forecast metrics worsened against the no-background and other-subject controls.