Researchers have been trying to understand how language models improve their ethical judgments over time without being explicitly trained on new data. A recent study published on arXiv proposes that this 'intrinsic moral self-correction' is driven by the model's hidden representations shifting in a specific way when prompted with certain words or phrases. The study used six different large language models and four tasks related to morality, finding that these representation shifts align with certain 'steering vectors'. This alignment even occurs when the steering vectors are created from a completely different dataset. The researchers also found that adding these shifted representations to the model's activations can improve its behavior more effectively than using the original self-correction prompts or steering vectors. --- Why it matters: Understanding how language models refine their ethical judgments is crucial for developing trustworthy AI systems. This study provides new insights into the mechanisms driving intrinsic moral self-correction, which could inform the development of more robust and transparent AI models. Source: https://arxiv.org/abs/2505.11924