Tag
提出 Conceptizer 训练框架，通过约 9K 个基于上下文同义词构建的概念词表，在 next-token prediction 之外增加意义层面的奖励信号，在 Llama 3、Qwen2.5 和 Falcon3（1B-8B）上提升了困惑度与下游知识/情感任务表现。
Proposes CSR, a framework that calibrates LLMs directly in semantic space using a novel semantic calibration reward, reducing ECE by up to 40% and improving AUROC by up to 31% over verbalized-confidence baselines across multiple datasets.
This paper proposes using reinforcement learning with semantic rewards (via GRPO) to expand LLMs to low-resource languages without the typical alignment tax of catastrophic forgetting, showing improved semantic quality and transferability over supervised fine-tuning.