AI college-advising models fail on real gaokao data; vertical OpenSpan scores 2.73/3 vs rivals

新智元 · wechat · 2026-11-05

A Beijing News reporter tested six mainstream AI models with the same mock gaokao student profile and found consistent failures: one used 2025 score tables for 2026 results (off by 2,440 ranks), others recommended discontinued majors or misclassified reach/fit/safety schools.

The article contrasts general LLMs' "answer-first" behavior with Hailiang's e生涯/OpenSpan career-planning model, which gathers a student profile first (province, track, scores, intended majors) and refuses to answer when data is insufficient
Architecture highlights: model-agnostic design (routing to stronger reasoners, cheaper retrievers, or multimodal models per task, with hot-swappable backbones) and an Agentic Skill Runtime that packages 8 scenario categories and 20+ sub-directions as pluggable skills — policy updates don't require retraining
Benchmarks: 88% hallucination-suppression pass rate, 14 points above Doubao and 30 above DeepSeek; expert evaluation scored OpenSpan 2.73/3 vs Doubao's 1.9 and DeepSeek's 2.0, with a 1.58-point lead on profile-building
The moat: hybrid RAG, four safety layers, and longitudinal growth records spanning primary through high school — data that can't be compressed in time
The piece draws a parallel to medical AI: verticals survive only if they hold expert workflows, verified knowledge bases, and long-term user data that general models can't replicate

Original post →

AI college-advising models fail on real gaokao data; vertical OpenSpan scores 2.73/3 vs rivals

More from Apps