DeepRepro replaces upfront planning with execution-state-aware subplanning for paper-to-code reproduction

arXiv·medium signalXBlueskyLinkedInCopy link

Posted 27 August, DeepRepro targets the long-horizon case where an agent must build and keep a whole repository working, not emit one file. Existing systems plan once upfront then generate files sequentially, which drifts as dependencies, interfaces and execution feedback change during the run. DeepRepro instead converts the evolving repository state and runtime feedback into fine-grained implementation subplans continuously, adds repository-aware orchestration and a lightweight process-aware monitoring interface, and outperforms strong scientific and commercial code-agent baselines on PaperBench Code-Dev.