Fixing inconsistent AI outputs in batches without slowing inference

Original: MarginGate: Sparse Margin-Triggered Verification for Batch-Invariant LLM Inference

Source: arxiv.org ↗