通过对齐使大型语言模型成为更好的推理耠
原文中文，约200字，阅读约需1分钟。发表于：。Reasoning is essential for large language models, and the Alignment Fine-Tuning (AFT) paradigm addresses the Assessment Misalignment problem in fine-tuned models, improving their reasoning...
该研究提出了一种多模态 - CoT 框架，将语言和视觉信息结合起来进行答案推断。该框架在 ScienceQA 基准测试中准确度高达91.68％，比先前最先进的 LLM（GPT-3.5）提高了16个百分点，甚至超过了人类的表现。
