LLMs writing RTL code are garbage at functional correctness. They'll generate something that compiles, but will it actually drive the right signals on a real chip? The answer has been "who knows"—until now.
EstRTL, a new framework from an anonymous academic team, adds a static functional score estimation loop to the RTL generation pipeline. Instead of just fine-tuning models or stacking retrieval tricks, they built a three-stage agent: Generation, Estimation, Correction. The key idea is an estimation agent that reads the generated RTL and scores it against a human-readable requirements spec before deciding what to do next.
The Static Score Changes the Game
Most prior work treats RTL generation as a translation problem—Verilog in, Verilog out—with no real check on semantics. EstRTL's estimation agent returns a quantitative score and a list of requirement comparisons. If the score is low, the code goes back for regeneration. If it's borderline, the correction agent takes over. If it passes, out it goes.
That static analysis means no simulation, no testbench orchestration, no runtime cost. It's a lightweight functional lint on steroids. The framework is model-agnostic and works with any LLM designed for RTL generation. Experiments show correctness gains of 3.2% to 9.0% across generic LLMs—no special fine-tuning required.
Real Transparency for AI Hardware Design
Beyond the numbers, the human-readable requirement comparisons change how engineers trust the output. You don't get a black-box RTL blob; you get a scorecard explaining what the generated code does and doesn't satisfy. That alone makes the framework useful for design review, even if the correction agent's edits need manual verification.
The authors open-sourced the code and experimental results at an anonymous repo (https://anonymous.4open.science/status/EstRTL-E200/). If you're working on LLM-assisted hardware design, this is the first tool I've seen that actually tries to verify intent—not just syntax.
Source: EstRTL: Functional Estimation Guided RTL Code Generation
Domain: arxiv.org
Comments load interactively on the live page.