A new paper on arXiv, CorrectRTL, targets the problem of generating register-transfer level (RTL) code from natural-language specifications using large language models. While LLMs have shown strong potential in this area, the authors focus on the self-correction paradigm, where a model iteratively reviews and fixes its own output. They note that this paradigm has already led to significant improvements in RTL generation quality.
However, the abstract states that two challenges still limit the effectiveness of self-correction in this domain. CorrectRTL is introduced as a method that boosts generation quality through automated self-correction combined with instruction decomposition. The paper's abstract does not specify the exact nature of the two challenges in the available excerpt, but the proposed approach is positioned as a way to overcome them.
The work contributes to the growing body of research on using LLMs for hardware design, particularly in automating the translation of high-level specifications into synthesizable RTL code. The full paper, available on arXiv, presumably details the decomposition strategy and the experimental results supporting the method's effectiveness.