Definition
Plain language
Working backward from a real result to construct the question it would answer.
As stated in the literature
In Firefly-style synthetic data generation, the process of recording real tool-call trajectories first and then authoring the natural-language task whose ground-truth answer can be read off the observed outputs.
Also called: back-chained
Why it matters: It generates training data whose ground-truth answers are guaranteed to match real tool behavior, sidestepping the hallucination problem in synthetic tool-use data.
For example, instead of inventing a task and then trying to solve it, back-chaining records an agent calling real APIs and then writes 'what task would naturally produce this trace?' as the prompt.
Heard on the show
“… And there's a sentence in the paper that captures the whole thesis: this back-chaining construction makes label correctness a property of the data-generation process rather than a …”Episode 059 — Firefly's Inversion: Building Verified Tool-Call Training Data by Working Backward