Researchers led by Waylon Li introduced GameSmith, a system that generates many two‑player zero‑sum games from fixed templates. Each game is checked by replay and a separate implementation, producing 26 new games in 29 families and 336 verified variants. The larger teacher model Qwen3‑32B self‑plays these games, then its knowledge is distilled into a smaller Qwen3‑4B‑Base model. On a 13‑benchmark suite, the distilled model’s mean pass@1 score rose from 45.5 to 62.5, and further self‑play with the CRAB estimator reached 64.9.
Why it matters
The improvement shows smaller language models can gain better reasoning ability by training on a larger variety of synthetic games.
