What is a fuzzy function?
Define a function in natural language, compile it, and call it from ordinary code.
Reading list
Neural compilation
and fuzzy functions
How can a description become a reusable neural program?
These papers study how to turn descriptions into reusable neural programs: from hypernetworks and generated adapters to Program-as-Weights and Compile by Training.
Programming, compilation, and evaluation
Neural compilation from natural-language specifications
Compare generated adapters, compressed instructions, and compilation through training.
How should neural functions be evaluated?
Test new tasks and noisy specifications, then measure compilation and execution costs.
Fuzzy-function programming
Program-as-Weights
First arXiv submission · July 2, 2026
NeurIPS 2026
Program-as-Weights: A Programming Paradigm for Fuzzy Functions
PAW proposes fuzzy-function programming: define a function in natural language, compile it into a reusable neural program, and execute it with a lightweight interpreter. The paper implements this compiler–interpreter interface with prefixes and LoRA adapters.
Weight-space programming and PAW → · Weight-space reading list → · Two-stage decision models: build once, run locally →
Research overview → · Paper on arXiv ↗ · Paper and citation →
Compilation through training
Compile by Training
First arXiv submission · September 3, 2026
EMNLP 2026 System Demonstrations
Compile by Training: Turning Natural-Language Specifications into Local Neural Functions
Teacher models generate examples from a function specification. Training on those examples refines the program produced by PAW’s fast compiler. The resulting function uses PAW’s program format and interpreter, with more work at compile time for higher accuracy.
How Program-as-Weights works
Generating LoRA Weights from Task Descriptions with Program-as-Weights
PAW's hypernetwork generates adapters from function definitions, then a small interpreter runs the saved programs.
Teacher-Generated Examples for Small Models: Program-as-Weights
How PAW learns from teacher examples to build reusable functions for a small language model.
Benchmarks for Natural-Language Function Compilation: Program-as-Weights
FuzzyBench tests whether PAW builds a function from its description and executes it on new inputs.
AI as Tool Builder: Yuntian Deng's Vision and Program-as-Weights
The June 2026 vision of pretraining AI to produce reusable neural programs, and its implementation in PAW.
Browse by topic
Yuntian Deng’s June 2026 essay, From AI as Problem Solver to AI as Tool Builder, describes the long-term vision behind PAW.
- Programs and reusable artifacts — Learn programs that can be saved and run on new inputs.
- Generated weights and adapters — Generate model weights from a task description, examples, or context.
- Context compression — Compress an instruction or context once and reuse the result across calls.
- Parameter-efficient building blocks — Adapt a frozen model with a small set of learned parameters.
- Compilation through training — Generate examples from a function specification and train a reusable neural program.