BugGen 2.0: Rethinking the Roles of LLMs for Scalable RTL Bug Synthesis
BugGen 2.0 met every requested bug quota while using 6.8–15.4 times fewer tokens than its predecessor in completed same-model comparisons.
Deterministic algorithms select and validate mutation targets, leaving the LLM to rewrite small sections of RTL hardware code. Across four OpenTitan IP blocks and FabScalar with three LLMs, the generated scenarios compiled, caused at least two test failures each, and covered 14–16 of the 16 configured mutation types. The system also generates alternative bugs at accepted locations and estimates acceptance from a simulated subset, skipping 63.6% of AES and 33.9% of FabScalar variant simulations. Full simulation confirmed acceptance rates of 77.7% and 95.2%, respectively, with a 1.09% estimation error on the directly comparable FabScalar set.
ArXiv · AI/CL/LG's note
Deterministic algorithms select and validate mutation targets, leaving the LLM to rewrite small sections of RTL hardware code. Across four OpenTitan IP blocks and FabScalar with three LLMs, the generated scenarios compiled, caused at least two test failures each, and covered 14–16 of the 16 configured mutation types. The system also generates alternative bugs at accepted locations and estimates acceptance from a simulated subset, skipping 63.6% of AES and 33.9% of FabScalar variant simulations. Full simulation confirmed acceptance rates of 77.7% and 95.2%, respectively, with a 1.09% estimation error on the directly comparable FabScalar set.
ArXiv · AI/CL/LG's note
score 4