Si-ATLAS: Autonomous Testing via LLM-Agent Synergy for RISC-V SBST Generation

Software-Based Self-Test (SBST) generation remains expert-intensive because useful programs must activate internal faults, propagate their effects, and make them architecturally observable. We present Si-ATLAS, a closed-loop framework that converts gate-level fault-simulation results into generation constraints. It combines deterministic instruction templates, LLM-guided structural exploration, and AST-aware operand fuzzing, while a static sanitizer and a runtime watchdog isolate invalid or non-terminating programs. On the CV32E40P RISC-V core, Si-ATLAS reaches 83.26% full-core stuck-at fault coverage; the multiplier reaches 95.38% after seven iterations. The default Claude Sonnet 3.5 campaign costs $5.23, while DeepSeek-V3 reduces the API cost to $0.64 and retains competitive coverage on regular datapaths.