Peking University releases K12-KGraph, a 23,640-question benchmark for educational LLMs
PekingUniversity · hf · 2026-07-24
Peking University researchers introduce K12-KGraph, a curriculum-aligned knowledge graph built from official People’s Education Press textbooks in math, physics, chemistry, and biology across primary, middle, and high school.
From it they derive:
- K12-Bench: 23,640 multi-select questions across five task families — Ground, Prereq, Neighbor, Evidence, and Locate
- K12-Train: 7,335 supervised samples, including 2,267 text-only QA pairs and 5,068 multimodal VQA pairs
The paper argues that existing education benchmarks mostly test exam-style recall, while real educational capability also requires “curriculum cognition” — understanding prerequisite chains, concept taxonomies, experiment-concept links, sequencing, and visual grounding.
Results:
- Gemini-3-Flash gets only 57% exact match on K12-Bench
- Gemma-4-31B-IT gets 46%
- Prereq and Neighbor are the hardest tasks
- Under a matched 2,300-sample budget, K12-Train-Text beats equally sized subsets of eight mainstream instruction-tuning corpora on GaokaoBench and EduEval
- For vision-language models, K12-Train-Full performs best on Gaokao-MM, MDK12-medium, and K12Vista among compared configurations, and outperforms both text-only and multimodal-only variants
The authors release the graph, benchmark, training data, and the full construction pipeline.
More from Multimodal
- Codex generates production-friendly concept art for a voxel-world prototype — Dimillian · 2026-07-24
- FLUX 3 unifies image, video, audio and action prediction in one model — robrombach · 2026-07-24
- A Seedance 2.5 prompt library is being built now to survive the next model cycle — Practical_Low29 · 2026-07-24
- A new Opus build appears to render a highly detailed voxel pagoda in one shot — cedric_chee · 2026-07-24
- Swiss AI releases Apertus v1.5 with 70B and 8B open multimodal models — jacek2023 · 2026-07-24
- Wan2.2 INT8 Convrot is slower than Q8 GGUF on a 3090 Ti — Fun-Class3451 · 2026-07-24