Meta's AutoInteract trains coding agents with simulated users who push back
Meta reports Qwen3.5-4B rose from 6.0% to 20.6% on SWE-Together after training on conversations where users clarify, revise and correct.
By RuntimeWire Staff · Published · Updated
Primary source: RuntimeWire original reporting
Why it matters
Coding agents are usually trained and graded on fully specified problems, but real users clarify and change requirements as they see output. AutoInteract's reported gains suggest interaction itself can be synthesized as training data while keeping executable tests as the reward. At 27B, the faster configuration required slightly more corrections.
Meta researchers have built AutoInteract, a framework that trains coding agents on conversations with simulated users who clarify vague requests, change requirements midway and correct mistakes. In Meta's research blog, the researchers say the usual post-training recipe uses fully specified tasks with an automatic verifier and no human in the loop, while real users often start with a goal that is not fully articulated and revise it after seeing output.…
When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.