GENERATIVE · 2026-11-12

Generative — Paper Daily, 2026-11-12

Today this track has a single paper, about how to check whether the candidate set an LLM recommends is correct. The authors use learning-path recommendation as the example: they compute every exercise a student can learn next from the prerequisite relations between exercises, then compare the LLM's recommendations against it. The result is clear: when models infer the structure from descriptions themselves, they make many errors, and when the structure is written directly into the prompt, they almost always match. Which model is best also changes with the prompt and the topic. When using an LLM for recommendation, the parts that rules can compute are best given to it explicitly, or simply computed yourself.

3.55

1. Benchmarking LLM Next-Step Recommendations with Knowledge Space Theory

When an LLM recommends which exercise a student should learn next, can it list every exercise that is ready, without slipping in ones that come too early? The authors use Knowledge Space Theory to turn this into an automatically scored task: the answer is every unlearned exercise whose prerequisites are all mastered. Across several major models and a few linear-function topics, with only exercise descriptions the recommended set matched the reference exactly in just 61.3% of responses. Asking the model to first infer the prerequisites itself and then recommend barely helped. Writing the prerequisite relations directly into the prompt raised it to 96.3%. Structure the model fills in by itself does not help, while structure given explicitly does. Which model is best also changes with the prompt and the topic. For people building recommenders, a candidate set you can compute with rules should not be left for an LLM to infer.