ProgramDistill: From Interactive Web Apps to Verifiable Reference-Guided SWE Tasks

Jeonghye Kim, Minseon Kim, Young Jin Kim, Matheus Pereira, Marc-Alexandre Côté, Alessandro Sordoni, Xingdi Yuan, Zhengyan Shi

Read on arXiv

Type: paper
Venue: arXiv
Year: 2026
Source: arxiv
Access: public
Language: en
Added: 2026-09-29
Verified: 2026-09-29

Summary

Turns interactive web apps into verifiable, reference-guided software-engineering tasks. The mine-craft-patch pipeline replays app behaviors to produce 1,975 replay-verified behaviors across 26 apps and 4,063 tasks with no human intervention.

Keywords

SWE tasks · agentic coding · distillation · web apps · verifiable tasks

Topics

SWE tasks, agentic coding, distillation, web apps, verifiable tasks

Research notes

Discovery: Shared in #random-papers as an arXiv link.
Method: mine-craft-patch: record and replay interactive web-app behaviors, verify them by replay, then distill reference-guided SWE tasks (cumulative-workflow restoration) from the verified behaviors.
Key findings: Reported cumulative-workflow success: GPT-6 Astra 49.2%, Claude Opus 5 28.8%. Partial-reconstruction success drops as restoration depth increases.
Overlaps in authorship with ROFT (arXiv:2609.35741); both are Microsoft Research agent-task papers from the same week.