LLM agents autonomously improve ABC logic synthesis tool
Researchers introduce a self-evolving framework where LLM agents autonomously rewrite and optimize the ABC logic synthesis system's source code, discovering new synthesis strategies that improve quality-of-results beyond human-designed heuristics.
Score breakdown
Developers and EDA researchers can leverage autonomous LLM-driven optimization to improve complex synthesis tools without manual heuristic design, enabling discovery of novel optimization strategies at production scale.
- 01The framework uses LLM-based agents to autonomously rewrite and evolve sub-components of ABC, a widely-adopted logic synthesis system
- 02The system operates on ABC's entire integrated codebase while preserving its single-binary execution model and command interface
- 03Each evolution cycle includes code modification proposals, binary compilation, correctness validation, and QoR evaluation on benchmarks including ISCAS 85/89/99, VTR, EPFL, and IWLS 2005
Cunxi Yu and Haoxing Ren introduce the first self-evolving logic synthesis framework that leverages LLM agents to autonomously improve the ABC logic synthesis system. The framework operates on ABC's entire integrated codebase while maintaining its single-binary execution model and command interface. The system bootstraps using existing open-source synthesis components covering flow tuning, logic minimization, and technology mapping, then deploys a team of LLM-based agents to iteratively rewrite and evolve specific sub-components following "programming guidance" prompts within a unified correctness and QoR-driven evaluation loop.
The results demonstrate that the framework can autonomously and progressively improve EDA tools at full million-line scale.
Each evolution cycle proposes code modifications, compiles the integrated binary, validates correctness, and evaluates quality-of-results on multi-suite benchmarks including ISCAS 85/89/99, VTR, EPFL, and IWLS 2005. Through continuous feedback, the system discovers optimizations beyond human-designed heuristics and learns new synthesis strategies that enhance QoR. The results demonstrate that the framework can autonomously and progressively improve EDA tools at full million-line scale.
Key facts
- 01The framework uses LLM-based agents to autonomously rewrite and evolve sub-components of ABC, a widely-adopted logic synthesis system
- 02The system operates on ABC's entire integrated codebase while preserving its single-binary execution model and command interface
- 03Each evolution cycle includes code modification proposals, binary compilation, correctness validation, and QoR evaluation on benchmarks including ISCAS 85/89/99, VTR, EPFL, and IWLS 2005
- 04The framework discovers optimizations and learns new synthesis strategies that enhance quality-of-results beyond human-designed heuristics
- 05The system demonstrates autonomous improvement of EDA tools at full million-line scale