---
id: 20261107-T0-03
title: "BlockLock：固定批几何让LLM推理结果可复现"
title_en: "BlockLock Enables Reproducible LLM Inference via Fixed Batch Geometry"
url: https://ai.daily.yangsir.net/daily/20261107-T0-03
issue_date: 2026-11-07
publish_date: 2026-11-06T05:00:00.000Z
category: research
source_name: "arXiv cs.AI"
source_url: https://arxiv.org/abs/2611.06977
---

# BlockLock：固定批几何让LLM推理结果可复现

arXiv新论文提出BlockLock，解决LLM推理的可复现性问题：同一请求在不同批次组合下应产生相同输出。现有方案要么用批不变核约束每个算子，要么隔离prefill阶段，代价较高。BlockLock通过固定批几何结构实现可复现，避免对全部算子施加限制。该方案对需要审计、调试和一致性验证的生产环境有直接价值，开发者可在不牺牲吞吐的前提下获得确定性输出。

## English Version

**BlockLock Enables Reproducible LLM Inference via Fixed Batch Geometry**

A new arXiv paper introduces BlockLock, a method for reproducible LLM inference where a request yields identical output regardless of which other requests share its batch. Existing approaches either constrain every operator with batch-invariant kernels or isolate prefill, both costly. BlockLock achieves reproducibility through fixed batch geometry, avoiding blanket operator constraints. The work matters for production settings requiring auditing, debugging, and consistency checks, offering deterministic outputs without necessarily sacrificing throughput.

---

**来源**：[arXiv cs.AI](https://arxiv.org/abs/2611.06977)

**详情页**：https://ai.daily.yangsir.net/daily/20261107-T0-03

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*