# Teacher‑free co‑evolution improves a 12B LLM for retail and airline tasks

> A Gemma‑4 12B agent learns policies from its own logs and gains up to 22 points on benchmark pass rates without a teacher model.

Oossa · 2026-11-13 · https://oossa.com/en/teacher-free-co-evolution-improves-a-12b-llm-for-retail-and-airline-tasks

Researchers led by Yujia Zhang released a paper on Nov 13 2026 describing a new training loop for a 12 billion‑parameter Gemma‑4 LLM used in customer‑service bots. The model writes its own policy rules from failed conversations and validates them before training a LoRA adapter. After two rounds, the agent lifted pass scores by 12.5‑19.5 points on retail and airline benchmarks, beating a reference version that hard‑codes the policy. The method narrows the performance gap to leading frontier models.

## The facts

- Paper published Fri Nov 13 2026
- Improvement: +12.5 retail, +19.5 airline pass points after two rounds

## Why it matters

Businesses can deploy smaller, open LLMs that self‑improve policy compliance without costly teacher models.

## Sources & references

1. [Knowledge in the Harness, Reliability in the Weights: Teacher-Free Co-Evolution of an LLM Agent](https://arxiv.org/abs/2611.15268) – arXiv, 2026-11-13

Last updated: 2026-11-13