Workflows vs Agents for Code Translation
Compares structured workflows versus agentic approaches for MATLAB-to-HDL translation, showing that agentic methods with the Model Context Protocol increase simulation reach rates by over 20 percentage points on mid-sized models.
Abstract
This paper compares two LLM-driven approaches for syntax repair in MATLAB-to-HDL translation. The researchers evaluated a structured, expert-designed flow versus an agentic method using the Model Context Protocol across three model sizes (Qwen 8B/30B/235B) on 42 signal-processing functions. The agentic approach achieved superior results on mid-sized models, increasing the simulation reach rate by over 20 percentage points on the 30B model. The authors attribute gains to short prompts, aggressive context management, and conditional tool use, noting that agent-based frameworks are most effective at compensating for the capacity limits of small and mid-sized models.
Related Research
UniPar: A Unified LLM-Based Framework for Parallel and Accelerated Code Translation in HPC
Introduces UniPar, an evaluation framework for assessing how large language models translate between parallel programming languages, achieving 69% compilation success and 33% functional correctness through fine-tuning and compiler-guided repair.
Counting Without Running: Evaluating LLMs' Reasoning About Code Complexity
Introduces gpuFLOPBench, a benchmark containing 577 CUDA kernels to evaluate whether language models can predict floating-point operation counts without execution, revealing limitations in understanding hardware-specific performance details.