Rejected by ICML, This 'Emergent Alignment' Method Lets LLMs Self-Correct Ethics Without Any Outside Judge

A new paper by Martin Kolář, titled 'Emergent Alignment,' tackles a critical problem in AI safety: can large language models (LLMs) detect and correct their own unethical outputs? Building on the 'Emergent Misalignment' scenario (where fine-tuning for code hacking led to unethical behavior), Kolář introduces a 'conscience step' that reviews the model's reasoning and outputs. The technique extends the training loss with a Direct Preference Optimization (DPO) component, using a frozen copy of the model as a judge—no weaker or stronger external model required. This online alignment method works across training, fine-tuning, adversarial prompting, and zero-shot learning, showing that a single high-level introspective question can steer training toward ethical outputs.

The empirical results demonstrate that this approach effectively prevents the emergent unethical behaviors observed in earlier work, particularly in the code hacking scenario. The paper was submitted to ICML 2026 but was rejected, suggesting the field remains skeptical or has higher standards for such self-correction methods. Nonetheless, the idea of self-alignment without external judges could reduce cost and complexity in deploying safe AI systems. If validated, this could enable models to continuously monitor their own ethics during deployment, reducing reliance on human feedback loops.