AWS fine-tunes Qwen3.6-27B search agent on Amazon SageMaker AI

AWS's post walks through a practical fine-tuning pipeline for a search agent, starting from the observation that base models do not arrive knowing a company's tools or environment. The write-up frames multi-turn reinforcement learning as a middle path between prompting small models, which rarely produces dependable multi-turn behavior, and prompting frontier models, which works more often but at higher latency and cost. It also positions MTRL against supervised fine-tuning, which depends on costly expert demonstrations, and single-turn RL, which scores one response at a time and misses the dependencies between turns.

The setup used a Qwen3.6-27B model and a training configuration that changed only three hyperparameters, leaving the algorithm, advantage estimator, and off-policy staleness bounds on defaults. Training data came from several public datasets, with five percent of training instances reserved for validation, and the reward was the retrieval metric nDCG@10. AWS says the fine-tuned model improved on WixQA, Wands, and BrowseComp-Plus, with a slight decrease on FreshStack, and that the number of turns per question fell on BrowseComp-Plus and FreshStack.

The stakes for readers evaluating internal search or RAG systems appear to hinge on whether this low-configuration recipe transfers to their own tools, datasets, and reward metrics, since the results come from AWS's selected benchmarks rather than a customer deployment. The body notes that MTRL training can span multiple days and that the default time limit is 24 hours, which may matter for teams planning longer runs.