论大语言模型适应性的局限：模型内化先验对标注任务性能的影响
On the Limits of LLM Adaptability: Impact of Model-Internalized Priors on Annotation Task Performance
AI总结：
通过毒性检测实验，研究大语言模型内化先验与指令交互的三个维度，发现近三分之二的零样本错误难以通过提示纠正，并引入定义特定熟悉度（DSF）指标，证明其与性能正相关，而文本记忆指标则无此关联。
摘要：
大语言模型（LLMs）越来越多地用于零样本标注和LLM-as-a-judge任务，但其可靠性取决于模型内化先验与用户提供指令的交互方式。我们研究了这种交互的三个维度：（1）LLM对数据和任务定义的熟悉程度如何影响性能；（2）提示中的额外信息能在多大程度上纠正零样本错误（“决策粘性”）；（3）模型对错误任务定义的敏感性。通过在多种数据集（涵盖社交媒体、游戏、新闻和论坛）上进行毒性检测实验，使用密集模型和混合专家模型，我们发现近三分之二的零样本错误难以纠正，提示纠正的总体挽救率（初始错误中被纠正的比例）仅为34.8%。高置信度错误尤其难以纠正。当给出错误定义时，LLM会遵循这些定义，同时保持与正确定义条件下相同的置信水平。关键的是，我们引入了定义特定熟悉度（DSF），它衡量模型内部概念与任务定义之间的一致性。在控制数据集层面的混杂因素后，DSF与模型性能呈正相关（偏相关系数r=+0.41），而三种不同的记忆指标（ROUGE-L、BERTScore和嵌入余弦相似度）均未显示正相关。这些发现揭示了基于提示的纠正在标注任务中的局限性，强调了定义对齐比文本级记忆更重要。
英文摘要：
Large Language Models (LLMs) are increasingly used for zero-shot annotation and LLM-as-a-judge tasks, yet their reliability hinges on how model-internalized priors interact with user-provided instructions. We investigate three dimensions of this interaction: (1) how an LLM's familiarity with data and task definitions relates to performance, (2) whether additional information in prompts can correct zero-shot errors ("decision stickiness"), and (3) model susceptibility to misaligned task definitions. We introduce Definition-Specific Familiarity (DSF), which measures alignment between a model's elicited concept and the target definition. Across nine LLMs and six toxicity datasets (five primary datasets plus an additional robustness dataset), DSF predicts annotation performance after controlling for dataset identity (partial $r=+0.41$). This association remains positive across all prompting conditions tested. In contrast, three common text-memorization metrics show no positive association. We show that prompting has limited corrective power: only 34.8% of zero-shot errors are corrected by additional instructions or examples, with high-confidence errors especially persistent. Misaligned definitions systematically shift predictions without reducing reported confidence, making confidence unreliable for detecting definition-policy mismatch. Together, these findings establish definition alignment as a practical model-selection criterion and show that better prompting alone cannot substitute for validating model-policy fit.
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。
