Program-as-Weights: A Programming Paradigm for Fuzzy Functions

Imagine you're trying to automate some programming tasks, like sorting through logs or fixing JSON errors. Traditionally, you might write specific rules or scripts for each task, but that can get messy and hard to maintain. Plus, if the rules don't cover every edge case, your system can fail or produce incorrect results. This is what's called the brittleness of rule-based systems.

Now, many people turn to large language models to handle these tasks, but that comes with its own set of problems. Using these models can be expensive, and they often require an internet connection, which isn't always feasible. This is where the idea of fuzzy-function programming comes in. Instead of relying on a large model for every single task, you can define what you want in plain language, and a smaller, local model compiles that into a compact, reusable function.

The authors introduce a specific implementation called Program-as-Weights (PAW). With PAW, a relatively small compiler takes your natural language description and creates a lightweight interpreter that can execute the function efficiently. The results show that this approach can match the performance of a much larger model while using far less memory and running faster on standard hardware. This means that developers can create and deploy functions quickly and cheaply, making it easier to build robust systems without the overhead of large models.

The approach introduces a new paradigm for programming tasks using natural language specifications.

The claims are supported by experimental results comparing performance and efficiency against existing models.