Provenance: reference testbed run — Qwen3-4B backbone on the 150-instance curated BFCL demo
subset (90 train / 30 sel / 30 test, seed 42, 3 epochs = 90 generations). The paper's reported
numbers use Qwen3-8B on the full benchmarks; this replay illustrates the mechanism, not the paper's scores.
Replay data failed to load. This page needs static/replay_data.js next to it
(shipped in the repo under docs/static/). If you are viewing a copy of this file alone,
grab the full docs/ directory from the
GitHub repo.
gen 0 / 89
Keyboard: ← / → step through generations, space toggles play.
Best validation score on Ssel across generations
(dots: held-out test mean reward at epoch ends; dotted lines: epoch boundaries)
Why is the parent so often init? Diversity-aware selection retains every policy
that wins at least one selection instance — including the initial policy — and parents are sampled
proportional to instance wins, so init keeps being re-selected as long as it still wins
instances. Deep chains form when an evolved policy overtakes it (e.g. the deployed
g48-planner lineage).
lineage batch reward best val (Ssel) held-out test population