Definition
Plain language
Turning a model's inconsistent best moves into a single reliable version by distilling several good examples into one clean instruction.
As stated in the literature
The mechanism behind trace-induced primitives: synthesizing a stable corpus-level specification of a reasoning move from multiple successful examples, so invoking it by name yields the consensus best behavior rather than a high-variance one-off reconstruction.
Why it matters: It matters because it turns a model's hit-or-miss best behavior into a dependable, reusable operation rather than a fresh gamble every time.
For example, instead of re-deriving 'check this alibi for consistency' from scratch each time with varying quality, the system distills many good past attempts into one reliable named move it can simply call.
Heard on the show
“The answer is a mechanism they call implicit aggregation, and once it clicks, the paradox dissolves.”Episode 110 — How an Agent Got 44 Points Better by Mining Its Own Scratch Paper