Paper
Compile by Training: Turning Natural-Language Specifications into Local Neural Functions
First arXiv submission: · EMNLP 2026 System Demonstrations
The paper reports public-service measurements from May 2026. Submission history ↗
Compiles a function description into a reusable local program by generating training examples and optimizing an adapter.
Read paper ↗ · Code ↗ · Demo ↗
What it does
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.
- Conditioning
- The specification guides teacher synthesis. PAW’s fast compiler supplies the initial adapter and runtime prompt.
- Program form
- A LoRA adapter, runtime prompt, specification, and interpreter metadata in a PAW program.
- Execution
- A frozen Qwen3-0.6B interpreter runs the saved function locally without further teacher calls.
Related work
Extends PAW with training at compile time while keeping its program format and runtime interface.
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.
Related terms: compile by training, training-based compilation, specification-specific optimization, synthetic supervision, local neural function, FuzzyBench-Hard, ProgramAsWeights.
Related methods
Fast prediction and training
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.
PAW compiler implementations · Compile by Training method · Compile by Training evaluation
Cite this paper
ArXiv preprint
arXiv:2609.04199, initial submission 2026-09-03.