Compile by Training: turning natural-language specs into neural programs with 83.6% accuracy
_akhaliq · x · 2026-09-05
The paper "Compile by Training" proposes compiling natural-language specifications into small neural programs that run locally on a compact interpreter.
- Compilation takes about a minute.
- It hits 83.6% semantic accuracy on the FuzzyBench-Hard benchmark.
- Instead of emitting machine code, the compiler's output is a small neural network — training replaces traditional compilation.
If matured, this could let natural-language-defined applications run on-device with minimal dependencies, blurring the line between compilers and neural networks.
More from Research
- Procedural Generation and Modern ML Share the Same Ethos: Sampling Conditioned Distributions — keenanisalive · 2026-09-05
- John Schulman: this research is timely as CoT monitorability declines — johnschulman2 · 2026-09-05
- Princeton team says OpenAI's new architecture closely resembles its T2MLR paper — prfsanjeevarora · 2026-09-05
- Shared: article on recurrent-depth models and latent reasoning — ziv_ravid · 2026-09-05
- One policy controlling 4 different robot hands for in-hand manipulation rejected by CoRL — YuXiang_IRVL · 2026-09-05
- Orbis paper reframes video generation as a steerable, continuous visual process — rohanpaul_ai · 2026-09-05