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