Can educational models do more than just answer questions correctly?

The OmniEdu paper argues that educational usefulness "is not captured by answer accuracy alone." Its introduction says a capable learning and teaching assistant must solve the student's problem, connect it to the appropriate knowledge point and prerequisite structure, identify the misconception behind an incorrect attempt, and select an intervention that advances learning. Treating education as ordinary question answering "leaves out the structure that makes tutoring effective." The response is an open family of K–12 models whose instruction-tuning corpus is organized around four capabilities — subject competence, curriculum grounding, diagnostic reasoning, and pedagogical action and scaffolding — rather than "solely by source or subject."

The paper's stated diagnosis is that existing educational models "often specialize in either subject problem solving or tutoring," and that their training mixtures are "organized by source or task and do not explicitly balance these capabilities." Its fix is to label supervision by capability and to build the corpus through a multi-stage pipeline: deterministic cleaning, semantic auditing and rewriting, task-specific quality scoring, token-budgeted diversity selection, and pedagogical instruction assignment. More than 100 resources are reduced to 69,999 examples and 15.96M supervised response tokens, 60,951 of them education-specific. The reported results are consistent improvements at 4B, 9B, and 27B; the strongest open-weight results on curriculum-grounding and pedagogical-tutoring benchmarks; and a 27B model that is "competitive with frontier proprietary models" on K–12 problem solving.

Set beside the neighbors, this is the tutor's side of an exchange that Can LLMs learn to ask for feedback during problem solving? trains from the student's side. Both hold that pedagogy is structure plain question answering discards, but OmniEdu targets diagnosing and scaffolding a learner rather than soliciting feedback. Its "semantic auditing and rewriting" step is teacher-style refinement of instruction data, which Does teacher-refined data always improve student model performance? warns is not universally better. The excerpt does not say whether rewriting was checked against each model in the 4B to 27B range. And like What limits reasoning capability beyond math and code?, it puts the lever in what the training data covers rather than in the training method.

The excerpt does not isolate the capability organization as the cause of the gains. It names no ablation against a source-organized mixture holding the cleaning and selection pipeline fixed. It also gives no benchmark names or scores, and does not say what the "consistent improvements" are measured against. Nor does it show that the four capabilities are separable or that balancing them helps rather than trades one off against another. The supportable reading is narrower than the title's framing: a capability-labeled, compact corpus is a workable design for a specialist K–12 model, and the paper's argument for it rests on the gap it sees in prior systems, not on a controlled comparison visible here.