Emergent Alignment

Abstract: Can Large Language Models (LLMs) discern when their own outputs are misaligned with human ethics? And can they self-correct?

摘要： 大型语言模型（LLMs）能否辨别其自身输出是否违背人类伦理？它们又能否进行自我修正？

We endow an LLM with a conscience step that reviews its own reasoning and outputs, and we extend the training loss with an alignment component using Direct Preference Optimization (DPO) to steer the model away from non-ethical outputs.

我们为大模型赋予了一个“良知步骤”（conscience step），用于审查其自身的推理过程和输出结果；同时，我们通过引入基于直接偏好优化（DPO）的对齐组件来扩展训练损失函数，从而引导模型远离非伦理的输出。

The result is an online technique to align models in a wide range of applications: training, fine-tuning, adversarial prompting, and zero-shot learning. It does not require a weaker or stronger judge, relying instead on a frozen copy of itself.

其成果是一种在线对齐技术，适用于广泛的应用场景，包括训练、微调、对抗性提示和零样本学习。该方法不需要更弱或更强的判别模型，而是依赖于模型自身的一个冻结副本。

In previous work, the Emergent Misalignment scenario showed a range of emergent unethical behaviors from fine-tuning the model to hack code. Instead, we empirically show how to achieve Emergent Alignment: a single high-level introspective question steers training toward an ethical model under the same code hacking scenario.

在先前的研究中，“涌现失调”（Emergent Misalignment）场景展示了模型在微调以进行代码破解时，会涌现出一系列不道德行为。与之相反，我们通过实证研究展示了如何实现“涌现对齐”（Emergent Alignment）：在相同的代码破解场景下，仅需一个高层级的内省式问题，即可引导训练过程向符合伦理的模型方向发展。

---

Paper Details:

• Title: Emergent Alignment
• Authors: Martin Kolář
• Date: 17 Jun 2026
• arXiv ID: 2606.19527 [cs.AI]

论文详情：

• 标题： 涌现对齐 (Emergent Alignment)
• 作者： Martin Kolář
• 日期： 2026年6月17日
• arXiv ID： 2606.19527 [cs.AI]