What question did this study set out to answer?

This report presents a new version of the programming language Synoema, designed to improve code generation for large language models (LLMs).

April 13, 2026Open Access

Synoema v0.1.0-alpha.3: Design, Type-Constrained Decoding, and Empirical Evaluation of a BPE-Aligned Programming Language for LLM Code Generation

Key Points

This report presents a new version of the programming language Synoema, designed to improve code generation for large language models (LLMs).
Developed a BPE-aligned syntax to ensure single-token mapping for all operators.
Created a type-constrained decoding system with incremental Hindley-Milner inference.
Implemented structured diagnostics and error recovery in the LLM development pipeline.
Introduced language-integrated testing with type-driven property test generators.
Conducted an empirical study involving 10 models and over 1,200 LLM calls.
Achieved 91.4% token accuracy with QLoRA fine-tuning on AMD GPUs.
Demonstrated a 3.0× speedup over CPython in execution speed.
Validated five hypotheses related to prompt engineering and model scaling.

Abstract

Technical report presenting Synoema v0.1.0-alpha.3 — a statically-typed functional programming language designed for LLM code generation with BPE-aligned syntax. This is a new version of the initial report (v2.1), extending it with five contributions: (1) a BPE-aligned language design methodology where all 33 operators map to single BPE tokens, eliminating token misalignment during constrained decoding; (2) a type-constrained decoding system combining incremental Hindley-Milner inference with row polymorphism to produce sound token masks for LLM sampling; (3) an integrated LLM development pipeline with structured diagnostics, single-pass error recovery, and contract-aware API changelog; (4) language-integrated testing with type-driven property test generators derived from inferred type signatures; (5) a comprehensive empirical study with 10 models (1B–70B parameters), 1,200+ LLM calls, five validated hypotheses on prompt engineering and model scaling, and QLoRA fine-tuning on consumer AMD GPU achieving 91.4% token accuracy. The implementation comprises 33,200 lines of Rust, 1,217 tests, and a Cranelift JIT backend with 3.0× median speedup over CPython. This document constitutes a defensive disclosure of intellectual property; the Zenodo publication date establishes priority for all disclosed innovations.

Synoema v0.1.0-alpha.3: Design, Type-Constrained Decoding, and Empirical Evaluation of a BPE-Aligned Programming Language for LLM Code Generation

Key Points

Abstract

Cite This Study