Gong et al. (2025) — ARXIV-2506.00204
Key takeaways
- Standard FIM treats code as text and masks random character spans, which produces training examples that rarely correspond to an edit anyone makes. AST-FIM masks whole syntactic units — blocks, expressions, functions — at scale, so the middle span is a coherent structure.
- Real-FIM-Eval: a benchmark derived from more than 30,000 GitHub commits across 12 languages, so the evaluation is drawn from real edits rather than synthetic holes.
- On 1B and 8B models, AST-FIM outperforms random-character FIM by up to 5 points on standard FIM benchmarks and is most useful on the real-world editing task.
Standing in the anthology
The construction LIT-119 follows for its FIM data, alongside random splits kept for robustness. Not yet drawn into a practice of its own; the FIM practice in the record (SOTA-128) is about loss masking, which this paper does not address.