Skip to content

Positioning against the 2025-2026 literature

A sweep of federated fine-tuning for software engineering, organization-scale adaptation and federated LoRA privacy, each record checked on arXiv. What it changes:

  • The nearest federated SE studies use synthetic organizations. Federated program repair (TOSEM, arXiv:2412.01072) partitions one dataset by Dirichlet draws over code features, runs six models, and reports no inferential test and no privacy attack; multi-task federated code review (arXiv:2412.15676) and federated code summarization (JSS 2026) likewise split public datasets. P4's organizations are real review cultures on permitted hosts, read over repeated partitions with a registered interval and power; RQ2 evaluates attacks and defences.
  • Organization-scale adaptation is measured without an ablation. Gemini for Google (arXiv:2605.16517) raised hunk acceptance 4.49% in a blind A/B test with 29,000 developers; its full text reports no analysis separating internal conventions from general ability, only a qualitative note on build rules and imports. That separation is what RQ1's decomposition does; cited in the Stage 1 background.
  • RQ2's attack family was uncited: source inference (Hu et al., TDSC 2024). FedAS-LoRA and LA-LoRA are placed without new cuts. A second model family is added as an exploratory item. All three are in ROADMAP Plan F.