--:--
今日 PV 0
独立 IP 0
在线 0
今日实时07/25LIVE
0今日访问
真人访问050%
搜索蜘蛛050%
当前在线0now
0PV 浏览量
0独立 IP
0蜘蛛抓取
24 小时访问趋势暂无数据
小伍的游乐场大数据

K12-KGraph:对齐课标的教育大模型知识图谱与基准(北大)| Paper: A Curriculum-Aligned KG for Educational LLMs

燎原Ai Build阅读 0留言 0

【中文】论文速递|K12-KGraph:对齐课标的教育大模型知识图谱与基准(北京大学)

北京大学从官方 K-12 数学、物理、化学、生物教材中构建课标对齐知识图谱 K12-KGraph(9 类节点、14 类关系,含视觉锚定),并配套 23,640 题多选基准 K12-Bench(5 个任务族)与 7,335 条 SFT 语料 K12-Train。评测显示头部模型差距明显:Gemini-3-Flash 精确匹配率仅 57%。在同等样本预算下,K12-Train 文本语料优于 8 个主流指令微调语料,图文联合监督进一步提升效果,说明理解课程知识结构仍是大模型短板。

【EN】Paper Brief | K12-KGraph: A Curriculum-Aligned Knowledge Graph for Educational LLMs (PKU)

Peking University constructs K12-KGraph, a curriculum-aligned knowledge graph extracted from official Chinese K-12 textbooks in math, physics, chemistry, and biology (nine node types, fourteen relation types, with visual grounding), plus K12-Bench, a 23,640-question multi-select benchmark across five task families, and the 7,335-sample K12-Train SFT corpus. Leading models show clear gaps: Gemini-3-Flash reaches only 57% exact match. At equal sample budgets K12-Train-Text beats eight mainstream instruction-tuning corpora, and combined text-visual supervision helps further, evidence that understanding curriculum structure remains a weak spot for LLMs.

来源 Source:

https://arxiv.org/abs/2605.09635

https://huggingface.co/papers/2605.09635

留言

NO REGISTRATION · 临时网名 + 邮箱即可开聊

无需注册 · 邮箱仅用于回复通知,绝不公开 · 广告与机器人会被蜜罐直接吞掉

K12-KGraph:对齐课标的教育大模型知识图谱与基准(北大)| Paper: A Curriculum-Aligned KG for Educational LLMs | 小伍的游乐场