SFT or RL for Tool-Calling Agents? A Controlled Study Across Data, Method, and Scale

What changed

A study evaluated SFT and RL across six Qwen3 models, finding SFT with LoRA strongest in-distribution while methods performed similarly on cross-dataset transfer.

Why you should care

Controlled evidence on agent training methods helps guide resource allocation between SFT and RL approaches.

Your move

Watch. Monitor further replication across other model families before altering production training pipelines.

Evidence

Preliminaryarxiv.org ↗

What to watch next

Independent replication of these findings on non-Qwen model families or larger parameter scales.

Event

research

Event date

Sep 17, 2026

Relevant to

General AI readers