Do large language models truly understand curriculum structure?

The K12-KGraph paper separates two things that K-12 evaluation usually runs together. On C-Eval, CMMLU, GaokaoBench, and EduEval, frontier models "now rival, and sometimes surpass, top human students," which invites the conclusion that educational AI is close to solved. The authors argue that these benchmarks measure only "whether a model can answer an exam question, i.e., factual recall." What they leave untested is what the paper calls curriculum cognition: "the structured understanding of how knowledge is organized and visually presented." Its examples are prerequisite chains, concept taxonomies, experiment–concept links, and pedagogical sequencing, meaning why one topic must be learned before another and where in the textbook each idea lives. When they test for this directly with K12-Bench, the conclusion is that "current LLMs lack robust curriculum cognition despite strong factual recall," with a strong open-source model reaching 46 percent exact match.

The mechanism is a change of what gets represented. K12-KGraph is extracted from the official People's Education Press textbooks for mathematics, physics, chemistry, and biology across primary, middle, and high school. It has nine node types (Book, Chapter, Section, Concept, Skill, Experiment, Exercise, Figure, VisualElement) and fourteen relation types "spanning both curriculum structure and visual grounding." Because the graph encodes the relations between concepts as well as the concepts themselves, it can serve as the source for an evaluation of curriculum structure (K12-Bench) and for training data (K12-Train). The paper adds that curriculum cognition is "neither probed by current benchmarks nor explicitly taught by current instruction-tuning data," so the same gap shows up on the measurement side and the training side.

The training results echo what the vault already holds about structure beating volume. With a matched budget of about 2,300 samples, K12-Train-Text outperforms equally sized subsets of eight mainstream SFT corpora on GaokaoBench and EduEval, which parallels Can organizing knowledge structures beat raw training data volume?. It also extends Can knowledge graphs teach models deep domain expertise? from medicine to school curricula and adds a multimodal result: K12-Train-Full beats larger general-purpose datasets on three multimodal educational benchmarks and beats both the text-only and image-only variants, so textual and visual supervision are described as complementary. As a distant parallel, Do foundation models learn world models or task-specific shortcuts? also finds that accuracy on a task's own test does not show that the underlying structure was learned. That paper is about orbital mechanics, and the K12 excerpt does not test for heuristics.

The excerpt is silent on several points a reader would want. It does not say what K12-Bench tasks look like, how exact match is scored, which model produced the 46 percent, or how humans or other models score on the same items. It also does not report whether K12-Train improves K12-Bench itself. The stated training gains are on GaokaoBench, EduEval, and three multimodal benchmarks, which are exam-style measures. So the sample-efficiency result shows KG-guided data helping on existing benchmarks, and the claim that it teaches curriculum cognition is inferred rather than shown in the passages available. The supportable reading is narrower: a recall-based score is not evidence of curriculum structure, and a curriculum-aligned graph offers one way to test for it and to build training data.