Researchers at various institutions have created a benchmark to test the performance of large language models (LLMs) in temporal legal reasoning. They found that LLMs often apply the most recently enacted law, regardless of when the relevant facts occurred. This bias is not due to an inability to understand laws' temporal scope or lack of knowledge about historical statutes. Instead, reinforcement-learning-shaped explicit reasoning may contribute to this issue by reducing the diversity of reasoning paths and causing models to converge on applying current laws. The study's findings suggest that improving general reasoning ability can actually worsen performance in temporally grounded legal reasoning. --- Why it matters: This matters because it highlights a potential flaw in LLMs' ability to apply laws correctly, which is crucial for tasks like legal judgment prediction. Engineers and researchers need to address this issue to improve the reliability of AI systems in legal contexts. Source: https://arxiv.org/abs/2608.14610