Research map / Image generation / Autoregressive and latent generation
InCoder: A Generative Model for Code Infilling and Synthesis
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Causal masking (infilling) instead of left-to-right single generation | InCoder-6.7B · HumanEval single-line infilling | pass rate: 48.2 -> 69.0 | Helped |
| Causal masking (infilling) instead of left-to-right single generation | InCoder-6.7B · HumanEval multi-line infilling | pass rate: 24.9 -> 38.6 | Helped |
| Causal masking infilling with bidirectional context | InCoder-6.7B · CodeXGLUE return type prediction | accuracy: 12.0 -> 58.1 | Helped |
| Causal masking infilling with bidirectional context | InCoder-6.7B · CodeXGLUE variable renaming | accuracy: 18.4 -> 30.6 | Helped |
| Causal masking infilling for docstring generation (zero-shot) | InCoder-6.7B · CodeXGLUE docstring generation | BLEU: 16.05 -> 18.27 | Helped |
| Causal masking objective instead of standard LM objective (1.3B) (ablation) | InCoder-1.3B · HumanEval and MBPP left-to-right synthesis pass@1 | pass@1: HumanEval 8 vs 6; MBPP 10.9 vs 8.9 | No clear effect |
| Increasing model size from 1.3B to 6.7B (CM objective) (ablation) | InCoder CM models · HumanEval and MBPP pass@1 | pass@1: HumanEval 8 -> 15; MBPP 10.9 -> 19.4 | Helped |
| Adding StackOverflow data to Python training data (ablation) | InCoder-1.3B LM · HumanEval and MBPP pass@1 | pass@1: HumanEval 5 -> 9; MBPP 6.1 -> 9.8 | Helped |
| Training on multiple languages instead of Python-only plus StackOverflow (ablation) | InCoder-1.3B LM · HumanEval and MBPP pass@1 | pass@1: HumanEval 9 -> 6; MBPP 9.8 -> 8.9 | Hurt |
| Causal masking infilling vs code-davinci-001 templated prompting | InCoder-6.7B · HumanEval single-line infilling | pass rate: 69.0 vs 63.0 | Helped |
| Causal masking infilling vs supervised TypeWriter | InCoder-6.7B · TypeWriter OSS return type prediction | F1: 59.2 vs 48.3 | Helped |
| Causal masking infilling vs fine-tuned CodeT5 on docstrings | InCoder-6.7B · CodeXGLUE docstring generation | BLEU: 18.27 vs 20.36 | Hurt |
Closest papers by meaning
- ObscuraCoder: Powering Efficient Code LM Pre-Training Via Obfuscation Grounding (ICLR 2025 · similarity 0.87)
- Training Language Models on Synthetic Edit Sequences Improves Code Synthesis (ICLR 2025 · similarity 0.87)
- Magicoder: Empowering Code Generation with OSS-Instruct (ICML 2024 · similarity 0.87)
- LILO: Learning Interpretable Libraries by Compressing and Documenting Code (ICLR 2024 · similarity 0.86)
- NextCoder: Robust Adaptation of Code LMs to Diverse Code Edits (ICML 2025 · similarity 0.86)
- JanusCoder: Towards a Foundational Visual-Programmatic Interface for Code Intelligence (ICLR 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.