arXiv 2602.113052026-02-11cs.CL

对齐的大型语言模型仍然存在偏差吗？

Are Aligned Large Language Models Still Misaligned?

Macquarie University（麦考瑞大学）；DSEU-Okhla；University of Delhi（德里大学）；Microsoft（微软公司）

Usman Naseem, Gautam Siddharth Kashyap, Rafiq Ali, Ebad Shabbir, Sushant Kumar Ray, Abdullah Mohammad, Agrima Seth

阅读 PDF

arXiv 页面 ↗HTML 原文 ↗

AI总结：

本文提出Mis-Align Bench基准测试，通过统一评估安全、价值和文化维度的偏差，揭示单维度模型在联合条件下的高覆盖率与高假失败率问题。

摘要：

大型语言模型（LLM）的偏差出现在模型行为与人类期望相悖，并且无法同时满足安全、价值和文化维度时。现有的偏差基准测试，如INSECURE CODE（以安全为中心）、VALUEACTIONLENS（以价值为中心）和CULTURALHERITAGE（以文化为中心），依赖于单独评估偏差维度，防止同时评估。为解决这一差距，我们引入了Mis-Align Bench，一个统一的基准测试，用于分析安全、价值和文化维度的偏差。首先，我们构建了SAVACU，一个包含382,424个样本、覆盖112个领域（或标签）的英语偏差-对齐数据集，通过将LLM-PROMPT-DATASET中的提示重新分类为14个安全领域、56个价值领域和42个文化领域，使用Mistral-7B-Instruct-v0.3进行分类，并通过Llama-3.1-8B-Instruct和基于SimHash的指纹扩展低资源领域以避免重复。此外，我们通过两阶段拒绝采样将提示与偏差和对齐响应配对，以确保质量。第二，我们对通用型、微调和开源权重的LLM进行基准测试，使能够系统地评估在三个维度下的偏差。经验上，单维度模型在联合条件下实现高覆盖率（最高达97.6%），但假失败率>50%，并且对齐分数较低（63%-66%）。

英文摘要：

Misalignment in Large Language Models (LLMs) arises when model behavior diverges from human expectations and fails to simultaneously satisfy safety, value, and cultural dimensions, which must co-occur in real-world settings to solve a real-world query. Existing misalignment benchmarks-such as INSECURE CODE (safety-centric), VALUEACTIONLENS (value-centric), and CULTURALHERITAGE (culture centric)-rely on evaluating misalignment along individual dimensions, preventing simultaneous evaluation. To address this gap, we introduce Mis-Align Bench, a unified benchmark for analyzing misalignment across safety, value, and cultural dimensions. First we constructs SAVACU, an English misaligned-aligned dataset of 382,424 samples spanning 112 domains (or labels), by reclassifying prompts from the LLM-PROMPT-DATASET via taxonomy into 14 safety domains, 56 value domains, and 42 cultural domains using Mistral-7B-Instruct-v0.3, and expanding low-resource domains via Llama-3.1-8B-Instruct with SimHash-based fingerprint to avoid deduplication. Furthermore, we pairs prompts with misaligned and aligned responses via two-stage rejection sampling to enforce quality. Second we benchmarks general-purpose, fine-tuned, and open-weight LLMs, enabling systematic evaluation of misalignment under three dimensions. Empirically, single-dimension models achieve high Coverage (upto 97.6%) but incur False Failure Rate >50% and lower Alignment Score (63%-66%) under joint conditions.

机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。