概率-熵校准：一种用于自适应微调的弹性指标
Probability-Entropy Calibration: An Elastic Indicator for Adaptive Fine-tuning
AI总结：
提出概率-熵校准信号（相对排名指标）进行token级重加权，以平衡预训练先验与下游对齐，在数学推理、分布外推理和代码生成任务上优于仅基于概率或熵的方法。
摘要：
Token级重加权是一种简单但有效的控制监督微调的机制，但常见的指标很大程度上是单维的：真实概率反映下游对齐，而token熵反映预训练先验引起的内在不确定性。忽略熵可能会将噪声或易替换的token误识别为学习关键，而忽略概率则无法反映目标特定的对齐。RankTuner引入了一种概率-熵校准信号，即相对排名指标，它比较真实token的排名与其在预测分布下的预期排名。逆指标作为token级的相对尺度用于重加权微调目标，将更新集中在真正未学习充分的token上，而不过度惩罚内在不确定的位置。在多个骨干网络上的实验表明，在数学推理基准上持续改进，在分布外推理上获得迁移增益，并且在代码生成性能上优于仅基于概率或熵的重加权基线。
英文摘要：
Token-level reweighting is a simple yet effective mechanism for controlling supervised fine-tuning, but common indicators are largely one-dimensional: the ground-truth probability reflects downstream alignment, while token entropy reflects intrinsic uncertainty induced by the pre-training prior. Ignoring entropy can misidentify noisy or easily replaceable tokens as learning-critical, while ignoring probability fails to reflect target-specific alignment. RankTuner introduces a probability--entropy calibration signal, the Relative Rank Indicator, which compares the rank of the ground-truth token with its expected rank under the prediction distribution. The inverse indicator is used as a token-wise Relative Scale to reweight the fine-tuning objective, focusing updates on truly under-learned tokens without over-penalizing intrinsically uncertain positions. Experiments on multiple backbones show consistent improvements on mathematical reasoning benchmarks, transfer gains on out-of-distribution reasoning, and pre code generation performance over probability-only or entropy-only reweighting baselines.
机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。
