Learn the Interface, Not the Implementation: Output Representation Dominates Training Regime in LLM-Based RTL Generation

Abstract

Large-language-model register-transfer-level (RTL) generation has concentrated on how to train the model: fine-tuning, continued pre-training, and reinforcement learning from a verifier. We argue this is the second-order question. On a bounded, formally specifiable task of generating single-error-correcting, double-error-detecting (SECDED) ECC memory wrappers, we hold the task, data, and correctness oracle fixed and vary two axes independently: the training regime (seven settings) and the output representation (four levels, from fully expanded flat Verilog to an eight-field configuration expanded by a deterministic generator). Because a reference generator exists, correctness is decided by formal equivalence over a design space of approximately ten million points, rather than sampled from a hand-written testbench. Two results follow. First, representation dominates: every representation close to raw RTL reaches a pass rate of at most 0.17, while the configuration reaches 0.83 with no training at all. Once the model emits the configuration, every training regime achieves a pass rate between 0.89 and 1.00. Second, raw RTL hits a hard output-length limit: a 512-bit design requires 44,237 tokens of flat Verilog compared with just 69 tokens for the configuration. Consequently, raw generation becomes infeasible above 32-bit data words rather than merely inaccurate. A model trained from scratch with fewer than 50 million parameters, combined with the verified generator, achieves 0.89 formal equivalence, compared with 1.00 for a 32-billion-parameter fine-tuned model, while using approximately 1600 times less inference energy per correct design. Inference energy also dominates amortized training energy. The lesson is to learn the interface, not the implementation.