Definition
Plain language
A speed trick where multiple AI runs that share the same opening context only compute that opening once.
As stated in the literature
In Forge, an exact (non-approximate) optimization that computes shared prompt prefixes once across many sampled rollouts and branches into divergent continuations, yielding up to ~40x training speedup.
Why it matters: Reusing shared prefixes turns wasted recomputation into pure speedup, dramatically lowering RL training cost without changing the math.
For example, 32 rollouts that share the same 4,000-token system prompt run that prompt through the model once instead of 32 times.