A new study examines how large language models (LLMs) assign credit to individual steps in a task. The researchers found that none of the current methods for assigning credit can identify which steps are causally important better than chance. They also discovered that existing evaluations, which focus on step correctness, do not align with the actual causal contribution of each step. The study suggests that implicit credit is often tied to the policy's fluency rather than its actual impact, and that conditioning on outcomes does not provide additional information about causal relationships. The authors propose a new method for assigning credit based on confidence-only routing, which can recover pivotal steps at chance level but reduces judge cost by 13.1% per turn. --- Why it matters: This study matters to AI researchers because it highlights the limitations of current methods for assigning credit in LLMs and suggests that implicit credit is often tied to fluency rather than actual impact. Understanding how to accurately assign credit is crucial for developing more effective and transparent AI systems. Source: https://arxiv.org/abs/2608.19760