LLM Humanities Exam Strategy: Data, CoT, Context, Multimodal

Analyze why large language models achieve high scores on humanities subjects (Chinese, history, geography) in the Gaokao. Break down the key factors: 1) Training data: high-quality, up-to-date, deduplicated corpus with high knowledge density and domain coverage; 2) Chain-of-thought (CoT): use multi-step internal reasoning before outputting structured answers, especially for complex questions; adaptively decide whether to reason based on question complexity (full CoT, no CoT, or adaptive CoT); 3) Long context: support large context windows (e.g., 256K) to read entire materials and questions without truncation, avoiding information loss; 4) Multimodal: directly process images (maps, charts, diagrams) without OCR, preserving visual information. Summarize that high performance relies on broad memory, reasoning before answering, strong image understanding, and full material ingestion.