SFT or RL for Tool-Calling Agents? The Answer Depends on What You Need Them to Do
New Dialpad research separates the effects of training method, data, and model scale—and finds that the best default is not a single technique.
Every business should truly know its customers. For AI agents to help make that possible, they must do more than generate fluent answers. They need to reliably take the right action—selecting the correct tool, producing valid arguments, and working across the situations they encounter in production.
That creates a practical question for teams building tool-calling agents: should they use supervised fine-tuning (SFT), reinforcement learning (RL), or both? The answer is often presented as a choice of method. Our research suggests the more useful answer is a decision framework that also accounts for training data and the operating environment.
In SFT or RL for Tool-Calling Agents? A Controlled Study Across Data, Method, and Scale, accepted to the REALM workshop at EMNLP 2026, we compare SFT with LoRA, RL using Group Relative Policy Optimization (GRPO), and SFT followed by GRPO. We evaluate six Qwen3 models from 0.6B to 32B parameters, trained on three public tool-calling datasets and a mixture of all three.
The finding: SFT is the most reliable choice in distribution
Across the full 0.6B–32B range, SFT with LoRA was the strongest in-distribution post-training method. It achieved the highest accuracy in 15 of 18 experimental settings and improved on the zero-shot baseline in every one.
That matters because a tool-calling agent is only useful when it can reliably turn intent into execution. In this study, SFT with LoRA was the most dependable way to improve performance when the evaluation task resembled the training data.
We also found that LoRA outperformed full-parameter fine-tuning in our 4B comparison across all three datasets. This supports the hypothesis that parameter-efficient adaptation can better preserve an instruction-tuned model’s existing tool-use capabilities while teaching it new behavior.
RL has a role, but not as a universal upgrade
GRPO was more competitive when models were evaluated on a different dataset from the one they were trained on. It produced the highest result in 29 of 54 cross-dataset settings. But the average margin over SFT was less than one percentage point.
That distinction is important. RL may offer a modest transfer advantage when an agent must operate beyond its training distribution. But in this controlled study, it did not consistently outperform SFT, and combining SFT and GRPO did not provide a reliable additional benefit under a matched data budget.
The practical takeaway is that teams should decide which method to use based on the job: reliable in-distribution performance, cross-dataset transfer, or a balance of both. Evaluation should reflect that intended operating environment.
The training data can matter as much as the method
One of the clearest results had little to do with the optimization method. Training-corpus choice changed in-distribution accuracy by as much as 55 points.
The uniform mixture of the three datasets emerged as the strongest general deployment default. It remained close to the best specialist model on each individual target while delivering the strongest transfer results overall. By contrast, models trained on a narrower dataset could struggle when evaluated on a different tool-calling format.
For practitioners, the implication is direct: do not treat the training dataset as a commodity. Its interaction patterns—single- or multi-turn calls, parallel tool use, refusals, and ordinary dialogue—shape what an agent learns to do.
From training recipes to evidence-based agent design
Dialpad is the AI platform for customer experience. Our research is focused on building AI systems that can understand context and take the right action in the real world. That requires moving beyond broad claims about a training technique and testing the actual tradeoffs among data, method, scale, and deployment conditions.
This study is deliberately bounded: it evaluates one model family, public datasets, and tool-calling metrics. It does not include a compute-normalized comparison between SFT and GRPO, and further replication across model families will strengthen the conclusions. Still, it provides a useful evidence base for a question many agent builders face today.
The headline is straightforward: SFT with LoRA is the most reliable in-distribution choice in this study; RL may offer a modest transfer benefit; and diverse training data is essential to building agents that can perform beyond a narrow benchmark.
SFT or RL for Tool-Calling Agents? A Controlled Study Across Data, Method, and Scale was accepted to REALM at EMNLP 2026. The paper is authored by Xue-yong Fu and colleagues at Dialpad.