Some text is shown in its original language while the English translation is being prepared.
Source arXiv
Seoirse Murray, Allison Qi, Timothy Qian, John Schulman, Collin Burns, Sara Price
Indexed by this site 2026-10-12 14:28 2 hr ago
Originally published Feb 6
Selected 2026-10-12 14:38
Category Alignment arXiv · arXiv:2602.05910 · Seoirse Murray, Allison Qi, Timothy Qian, John Schulman, Collin Burns, Sara Price · Indexed by this site 2026-10-12 14:28 · Originally publishedFeb 6 Attention54 Chunky Post-Training：数据驱动的泛化失败 Chunky Post-Training: Data Driven Failures of Generalization
AI reading guide
研究者提出 chunky post-training 概念，指模型从后训练数据中学到非预期的相关性，从而在不当语境触发拒答、反驳或写代码等行为，并开源 SURF 与 TURF 两个工具分别用于发现和归因这类失败。在 Claude 4.5、GPT-5.1、Grok 4.1、Gemini 3 与开源模型 Tülu 3 上，这类行为误路由普遍存在，且对提示改写较为稳健。TURF 将 Tülu3 的失败追溯到具体数据，例如约 250 条拒绝虚假发明问题的样本使其反驳真实事实，向数学题注入 LaTeX 使 sympy 幻觉调用增加 50% 并拉低准确率。数据消融显示移除相关样本或补充正确示例可减轻该行为。
Deep reading Generating
Why it was selected
论文把模型在特定措辞下误触发拒答、反驳等行为归因到后训练数据的分布失衡，并给出可复用的定位工具。
Read original arxiv.org Why it was selected Attention54
Attention 0–100: content value contributes up to 70 and circulation up to 30; circulation decays over time. How it was rated
论文把模型在特定措辞下误触发拒答、反驳等行为归因到后训练数据的分布失衡，并给出可复用的定位工具。