SOSP 2026

Microsoft Research presents LLM-42 for deterministic AI inference

LLM-42 uses a decode-verify-rollback approach designed to make selected LLM inference requests reproducible without disabling dynamic batching.