OmniEdu: Open Foundation Models for Learning and Teaching
Hao Liang, Qihan Lin, Meiyi Qiang, Linzhuang Sun, Hengyi Feng, Mingrui Chen, Sizhe Qiu, Wentao Zhang
cs.CL
2026-09-19
OmniEdu SFT 4B/9B/27B on 69,999 K-12 examples. The 27B hits 63.12% K12-Bench EM, 78.74% scaffold win rate, and 3.02 LongTutor Teaching.
Most educational language models pick a lane: they either solve exam problems or run tutoring chat. Their training mixes are usually sliced by source or subject, not by the behavior each example is supposed to teach. The typical failure mode is familiar. The model can get the answer, then cannot place the item in a curriculum, cannot name the misconception in a student's work, and blurts the solution on the first turn.
A usable K-12 assistant has to do four jobs at once: solve the problem, locate it in a curriculum, diagnose the learner, and pick an instructional move (ask, hint, reteach a prerequisite, or explain). OmniEdu, from Peking University, the University of the Chinese Academy of Sciences, and Zhongguancun Academy, turns those four jobs into the organizing principle of a post-training mixture, then fine-tunes 4B, 9B, and 27B models on the same corpus.
Supervision is grouped by capability, not by subject:
A six-stage pipeline cuts an education pool of about 1.34M examples down to 60,951:
9,048 general instruction examples are added (DataFlow-Instruct-10K 7,431, Tulu-3-SFT 1,495, MathV360K 122). The full mix is 69,999 examples and 15.96M supervised response tokens. Backbones are Qwen3.5-4B-Base, Qwen3.5-9B-Base, and Qwen3.8-27B, full-parameter SFT in LLaMA-Factory, learning rate 5e-6, 3 epochs, 32,768 context.
Curriculum scores rise at every scale against the matching base. OmniEdu-27B moves K12-Bench EM from 52.11% to 63.12% and F1 from 73.48% to 76.69%; MathFish from 83.54% to 85.89%; EDUMATH MaC (standards-aligned problem generation) from 70.60% to 86.95%, second to Kimi-K3 at 90.00%. Curriculum localization EM jumps from 29.85% to 48.77%; positive alignment on MathFish from 45.95% to 67.60%. The 4B EDUMATH MaC gain is 47.60% to 68.40%.
Problem solving:
| Model | GAOKAO full | EXAMS-V | MDK12 |
| OmniEdu-4B | 90.46% (base 88.98%) | 57.62% (44.69%) | 46.36% (35.43%) |
| OmniEdu-9B | 93.66% (92.94%) | 66.40% (63.00%) | 50.80% (44.50%) |
| OmniEdu-27B | 94.87% (91.55%) | 69.52% (68.65%) | 57.76% (46.04%) |
On MDK12, 27B open-question correctness goes from 62.08% to 78.77%; GAOKAO Chinese from 55.10% to 94.26%. Open educational baselines such as Confucius3-Math and EduChat collapse on several curriculum and multimodal exam splits, so those gaps are easy to over-read. Proprietary systems still lead on some exams: Claude-Opus-5 at 97.22% GAOKAO full, Kimi-K3 at 87.29% EXAMS-V.
Tutoring is where the mixture moves the needle. MathTutorBench scaffold win rate: 4B 20.42% to 75.79%, 9B 14.00% to 75.26%, 27B 57.16% to 78.74% (83.59% on the hard split). GPT-5.4 scores 6.32% on scaffold, the "just give the answer" failure. Claude-Opus-5 still leads at 87.89%. LongTutor evidence use for 27B goes from 36.80% to 78.20%, and the Teaching average of 3.02 is the best in the table. Knowledge-state diagnosis tops out at 54.04%. TutorBench barely moves: 27B 58.58% to 59.42%, behind Kimi-K3 at 63.65%.
General capability does not get spent to buy the education gains. IFEval prompt-strict for 27B goes 80.59% to 82.07%, GPQA Diamond 74.24% to 77.78%, MMMU-Pro 64.97% to 67.98%.
This is a reproducible open K-12 post-training recipe: data, weights, and code are public. For product teams the useful claim is not "more exam items." It is that solving, curriculum, diagnosis, and teaching are conflicting response policies, and they have to be budgeted and labeled separately, with system instructions that make the desired behavior explicit. At 70k examples and 16M tokens, a 4B or 9B open model is a realistic starting point. The scaffold jump is the tell: the base models already knew how to solve; they did not know when to stop short of the answer.
It is still data engineering, not a new architecture. TutorBench is almost flat and diagnosis sits at 54%. Teaching is not solved.
There is no dedicated limitations section. The holes are visible anyway. There is no capability ablation and no stage-wise pipeline ablation, so the four-way mix, the token budgets, and the 20 system prompts cannot be credited separately. The authors themselves call knowledge-state diagnosis comparatively hard. MathFish overall only rises 2.35 points, and some negative-alignment cells fall (27B ATC-neighbor negatives 88.31% to 79.07%). OmniEdu-4B GAOKAO Chinese drops from 90.89% to 59.30%; subject-level scores are too noisy to treat as product metrics. Training is SFT only, with no pedagogy-oriented RL or preference optimization. At inference the correct system instruction has to be chosen, so product orchestration is still required. Tutoring metrics lean on LLM-judge win rates; there is no classroom study with human teachers. Several open educational baselines fail so badly that the open-weight gap is inflated. The 27B backbone is Qwen3.8, while 4B and 9B are Qwen3.5.