A new paper on arXiv offers a process-centered synthesis of large language model tutoring. The authors note that while LLMs can produce fluent tutoring dialogue, whether that dialogue actually helps students learn remains an open question. They argue that prior evaluations have focused too much on surface-level conversational quality.

The proposed synthesis follows the full tutoring loop, starting with how the model infers the learner's state and continuing through the actions it takes in response. By examining each stage of this process, the paper aims to identify where LLM tutors succeed or fail at supporting genuine learning.

Because this is a single synthesis paper, there are no conflicting sources to compare. Its contribution is primarily conceptual: it offers a framework for future empirical work on LLM tutoring, rather than new experimental results.