Skip to content

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.

By Margalla News Desk

Published

1 min read

Microsoft Research has presented LLM-42, a scheduling-based technique for deterministic large-language-model inference. The system generates candidate tokens on a fast path and verifies them under a fixed-shape schedule, rolling back when mismatches occur. Researchers say the approach allows determinism to be applied selectively, which can be useful for evaluation, auditing and continuous-integration workloads.

Article: https://margallanews.com/story/microsoft-llm42-deterministic-inference-verified-speculation-2026-09-29