RetryTrigger: intelligent inference duplication for enhancing LLM resilience to hardware transient faults

• RetryTrigger is a hardware-free, fault-aware inference methodology that predicts when duplicate inference is needed, eliminating reliance on specialized hardware or restrictive boundary settings. • It leverages runtime output features (maximum probability, top-k gaps, entropy, logits statistics, latency) to train a LightGBM meta-model that accurately detects and mitigates silent data corruptions. • Experiments across seven LLMs show SDC rate reductions up to 95.33% (average 92.97%) with a minimal performance overhead of 2.4012% (average 4.1167%). • RetryTrigger offers a superior reliability-efficiency trade-off compared to ABFT and FT2, making it suitable for safety-critical NLP applications such as medical diagnosis and legal document analysis.