Researchers propose Program-as-Weights (PAW), a paradigm that compiles fuzzy functions from natural-language descriptions into compact, locally-executable neural adapters. A 4B compiler trained on 10M examples generates parameter-efficient adapters for a frozen 0.6B Qwen3 interpreter, achieving performance comparable to direct prompting of a 32B model while using ~50× less memory.
Background
Large language models have increasingly been used as drop-in replacements for heuristic programming tasks, but cloud API dependency introduces cost and latency concerns. This work explores compiling specialized functions into lightweight neural artifacts for offline execution.
- Source
- Lobsters
- Published
- Nov 17, 2026 at 08:55 PM
- Score
- 7.0 / 10