A new study suggests that supervised fine-tuning remains the strongest default when a language-model agent will use tools in the same kind of tasks it saw during training. Researchers compared several training methods across six Qwen3 models ranging from 600 million to 32 billion parameters.
The experiments tested supervised fine-tuning with LoRA, reinforcement learning using Group Relative Policy Optimization, and a sequence that applied both. LoRA is a parameter-efficient method that changes a small set of added weights instead of retraining the entire model. On in-distribution tests, it was best in 15 of 18 settings across the full range of model sizes.
Reinforcement learning performed somewhat better when the training and test datasets differed, winning 29 of 54 cross-dataset settings. Its average advantage over supervised tuning was less than one point, however. Applying reinforcement learning after supervised tuning was rarely the strongest option in either comparison.
Mixing datasets produced consistently strong transfer while remaining close to specialized training performance, regardless of the adaptation method. The authors also found that LoRA beat full-parameter fine-tuning, which they attribute to better preservation of capabilities learned during pretraining. These are benchmark results on Qwen3 models, so production agents and other model families may behave differently.