{
  "id": "2609.04199",
  "title": "Compile by Training: Turning Natural-Language Specifications into Local Neural Functions",
  "authors": "Yuntian Deng, Pengyu Nie, Stuart Shieber",
  "first_public": "2026-09-03",
  "first_public_source": "https://arxiv.org/abs/2609.04199",
  "arxiv_first_submission": "2026-09-03",
  "date_label": "arXiv v1",
  "arxiv": "2609.04199",
  "paper_url": "https://arxiv.org/abs/2609.04199",
  "original_url": "https://arxiv.org/html/2609.04199v1",
  "venue": "EMNLP 2026 System Demonstrations",
  "venue_source": "https://arxiv.org/abs/2609.04199",
  "lens": [
    "neural-programs",
    "training-compilation"
  ],
  "related_terms": [
    "compile by training",
    "training-based compilation",
    "specification-specific optimization",
    "synthetic supervision",
    "local neural function",
    "FuzzyBench-Hard",
    "ProgramAsWeights"
  ],
  "summary": "Compiles a function description into a reusable local program by generating training examples and optimizing an adapter.",
  "contribution": "Makes task-specific training a software build step. Raises FuzzyBench-Hard semantic accuracy (LEM) from 22.4% to 83.6% and demonstrates composition in a website helper, avatar control, and English–Claudish translation.",
  "limitations": "Synthetic examples can inherit teacher errors. Cold compilation takes 50.9 seconds for one specification on a B300. FuzzyBench-Hard selects fast-compiler failures under exact match; the reported accuracy uses an LLM judge.",
  "program_form": "A LoRA adapter, runtime prompt, specification, and interpreter metadata in a PAW program.",
  "conditioning": "The specification guides teacher synthesis. PAW’s fast compiler supplies the initial adapter and runtime prompt.",
  "execution": "A frozen Qwen3-0.6B interpreter runs the saved function locally without further teacher calls.",
  "relationship": "Extends PAW with training at compile time while keeping its program format and runtime interface.",
  "relationship_source": "https://arxiv.org/html/2609.04199v1#S1",
  "project_url": "https://programasweights.com/playground?compiler=paw-ft-bs48",
  "demo_url": "https://programasweights.com/playground?compiler=paw-ft-bs48",
  "code_url": "https://github.com/programasweights/compile-by-training",
  "provenance": "September 2026 arXiv paper; its record lists EMNLP 2026 System Demonstrations.",
  "first_public_note": "The paper reports public-service measurements from May 2026.",
  "author_names": [
    "Yuntian Deng",
    "Pengyu Nie",
    "Stuart Shieber"
  ],
  "record_url": "https://programasweights.com/reading-list/papers/2609.04199/",
  "legacy_record_url": "https://programasweights.com/reading-list/papers/#paper-2609-04199",
  "citation_date": "2026-09-03",
  "json_url": "https://programasweights.com/reading-list/papers/2609.04199/record.json",
  "markdown_url": "https://programasweights.com/reading-list/papers/2609.04199/index.md",
  "bibtex_url": "https://programasweights.com/reading-list/papers/2609.04199/citation.bib",
  "csl_url": "https://programasweights.com/reading-list/papers/2609.04199/citation.csl.json",
  "method_comparisons": [
    {
      "title": "Fast prediction and training",
      "text": "PAW’s fast compiler predicts a program. Compile by Training starts from that prediction, generates examples, and trains an adapter. Both use the same runtime interface; the extra training improves semantic accuracy on FuzzyBench-Hard at a higher compilation cost.",
      "source_ids": [
        "2607.02512",
        "2609.04199"
      ],
      "source_urls": [
        "https://arxiv.org/html/2607.02512v1#S3",
        "https://arxiv.org/html/2609.04199v1#S3",
        "https://arxiv.org/html/2609.04199v1#S5"
      ],
      "source_labels": [
        "PAW compiler implementations",
        "Compile by Training method",
        "Compile by Training evaluation"
      ],
      "topic_id": "training-compilation",
      "topic_label": "Compilation through training",
      "topic_url": "https://programasweights.com/reading-list/guides/compilation/#training-compilation"
    }
  ]
}