COEVO: Co-Evolutionary Framework for Joint Functional Correctness and PPA Optimization in LLM-Based RTL Generation

2026-04-21 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, RTL Design & Synthesis · Depth: Expert, extended

Summary

COEVO is a co-evolutionary framework for Large Language Model (LLM)-based Register-Transfer Level (RTL) code generation that jointly optimizes functional correctness and Power, Performance, and Area (PPA) quality. Unlike existing methods that decouple these objectives, COEVO unifies them within a single evolutionary loop. It treats correctness as a continuous co-optimization dimension, enabled by an enhanced testbench providing fine-grained scoring and diagnostic feedback. An adaptive correctness gate with annealing allows partially correct but architecturally promising candidates to guide the search. To preserve PPA trade-offs, COEVO uses four-dimensional Pareto-based non-dominated sorting with configurable intra-level sorting, replacing scalar fitness. Evaluated on VerilogEval 2.0 and RTLLM 2.0, COEVO achieved 97.5% and 94.5% Pass@1 with GPT-5.4-mini, outperforming all agentic baselines and securing the best PPA on 43 out of 49 synthesizable RTLLM designs.

Key takeaway

For research scientists developing LLM-based hardware design tools, you should consider integrating co-evolutionary frameworks that unify correctness and PPA optimization. Decoupling these objectives or reducing PPA to a scalar fitness risks discarding architecturally superior, partially correct designs. Your approach should leverage continuous correctness scoring and multi-dimensional Pareto selection to explore the design space more effectively and achieve superior functional and PPA outcomes.

Key insights

Jointly optimizing functional correctness and PPA in RTL generation requires continuous co-evolution and multi-objective Pareto sorting.

Principles

• Treat correctness as a continuous optimization dimension.
• Preserve PPA trade-offs using multi-dimensional Pareto sorting.
• Allow partially correct, PPA-promising candidates to guide evolution.

Method

COEVO uses LLM-driven evolutionary operators, an enhanced testbench for continuous correctness scoring, and 4D Pareto-based non-dominated sorting with an adaptive correctness gate for survivor selection.

In practice

• Implement fine-grained testbench feedback for continuous scoring.
• Utilize adaptive gates to retain promising, partially correct designs.
• Apply Pareto-based sorting for multi-objective optimization.

Topics

• COEVO Framework
• LLM-based RTL Generation
• Co-evolutionary Optimization
• PPA Optimization
• Functional Correctness

Code references

• hping666/COEVO

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Hardware Engineer

Related on AIssential

• Mutation Without Variation: Convergence Dynamics in LLM-Driven Program Evolution — LLM-driven program mutation without selection pressure consistently converges to restricted structural forms, limiting diversity.
• CoEval: Ranking Language Models for Custom Tasks Without Labeled Data or Trustworthy Benchmarks — The CoEval framework enables reliable, label-free LLM ranking for custom tasks by synthesizing fresh benchmarks and using diverse judge ensembles.
• Self-Evolving Deep Research via Joint Generation and Evaluation — Co-evolving LLM generation and evaluation within a shared-parameter framework overcomes static evaluator limitations in deep research.
• Constrained Hybrid Metaheuristic: A Universal Framework for Continuous Optimisation — cHM is a universal, adaptive hybrid metaheuristic that dynamically combines diverse optimizers for robust continuous optimization.
• Towards Automated Knowledge Transfer in Evolutionary Multitasking via Large Language Models — LLMs can autonomously design efficient and effective knowledge transfer models for multi-task optimization.
• MetaEvo: A Meta-Optimization Framework for Experience-Driven Agent Evolution — MetaEvo enables LLM agents to continually improve reasoning by optimizing how they learn principles from experience, not just what they store.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.