Researchers have proposed a new tool called FaithMATE to investigate the relationship between two ways of measuring how accurately large language models reflect their underlying behavior. The first way involves changing the input or the model's internal thought process, while the second way involves modifying the model's parameters. By using FaithMATE, the researchers found that these two approaches are related but also have some differences. They discovered that improving one aspect of a model's faithfulness can lead to improvements in both aspects, but with varying degrees of success. This suggests that faithfulness is not a single goal, but rather a complex objective that requires multiple approaches for optimization and evaluation. --- Why it matters: This research matters because it highlights the importance of considering multiple perspectives when evaluating the accuracy of large language models. By understanding how these different approaches interact, developers can create more effective tools for optimizing model performance and improving their faithfulness to real-world behavior. Source: https://arxiv.org/abs/2605.24960