指令微调如何塑造语言模型的选数偏好
原标题：3, 7, 14, 17, 42: Instruction Tuning Installs and Sharpens Number-Choice Heuristics in Language Models
AI 摘要
该研究提出一种无需采样的枚举方法，通过累加各整数拼写形式的 token 序列概率，精确计算语言模型在“随机选数”提示下的答案分布，消除了采样误差。作者评估了 18 个开放权重模型系列、5 个数值范围、29 种解码配置和两套推理栈，发现指令微调在全部 15 对可比检查点中都使分布更不均匀：微调模型会安装“回避整数”启发式并强化“回避端点”启发式，而基座模型并不回避整数。研究还发现微调模型仅在 1-10 范围内稳定偏好奇数，且没有跨模型共享的偏好整数，说明微调并未安装统一的“幸运数字”。
正文节选
3, 7, 14, 17, 42: Instruction Tuning Installs and Sharpens Number-Choice Heuristics in Language Models Abstract We measure, without sampling error, the probability a language model assigns to each integer from 1 to when asked for a random number. For a fixed prompt the model defines one probability vector, which we compute by summing over each integer’s scored spellings rather than estimate by sampling, removing the noise that limits earlier studies [1, 4]. We evaluate 18 open-weight model lines