# GameSmith expands zero‑sum game training for LLMs

> The new pipeline creates hundreds of synthetic two‑player games, letting a 4B‑parameter model improve its reasoning score from 45.5 to 62.5.

Oossa · 2026-10-14 · https://oossa.com/en/gamesmith-expands-zero-sum-game-training-for-llms

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.

## The facts

- 26 new games in 29 families and 336 verified variants were generated
- Qwen3‑4B‑Base pass@1 mean improved from 45.5 to 62.5 after distillation

## Why it matters

The improvement shows smaller language models can gain better reasoning ability by training on a larger variety of synthetic games.

## Sources & references

1. [GameSmith: Scaling Zero-Sum Self-Play with Synthesised Games](https://arxiv.org/abs/2610.13995) – arXiv cs.LG, 2026-10-14

Last updated: 2026-10-14