未言明的提示:LLM推理中无承认的准确性
Unspoken Hints: Accuracy Without Acknowledgement in LLM Reasoning
AI总结:
本研究通过控制提示的正确性、风格和复杂性,系统评估了LLM思维链推理的忠实性,发现正确提示提升准确性但承认不均,谄媚与泄露风格影响依赖方式,揭示推理被捷径系统性塑造。
摘要:
大型语言模型(LLMs)日益依赖思维链(CoT)提示来解决数学和逻辑推理任务。然而,一个核心问题仍然存在:这些生成的推理过程在多大程度上忠实于底层计算,而非由嵌入提示中作为答案捷径的提示所塑造的事后叙述?遵循先前关于有提示与无提示提示的研究,我们在受控提示操纵下对CoT忠实性进行了系统性研究。我们的实验设计涵盖四个数据集(AIME、GSM-Hard、MATH-500、UniADILR)、两个最先进的模型(GPT-4o和Gemini-2-Flash),以及一组结构化的提示条件,这些条件在正确性(正确和不正确)、呈现风格(谄媚和数据泄露)和复杂性(原始答案、双操作符表达式、四操作符表达式)上有所不同。我们评估了任务准确性以及推理中是否明确承认提示。我们的结果揭示了三个关键发现。首先,正确的提示显著提高了准确性,尤其是在更难的基准和逻辑推理上,而不正确的提示在基线能力较低的任务中急剧降低了准确性。其次,提示的承认极不均匀:基于方程的提示经常被引用,而原始提示常被静默采用,表明更复杂的提示促使模型在推理过程中口头表达其依赖。第三,呈现风格很重要:谄媚提示鼓励公开承认,而泄露式提示提高了准确性但促进了隐藏依赖。这可能反映了与RLHF相关的影响,因为谄媚利用了取悦人类的一面,而数据泄露触发了自我审查的一面。总之,这些结果表明LLM推理被捷径系统性塑造,从而模糊了忠实性。
英文摘要:
Large language models (LLMs) increasingly rely on chain-of-thought (CoT) prompting to solve mathematical and logical reasoning tasks. Yet, a central question remains: to what extent are these generated rationales \emph{faithful} to the underlying computations, rather than post-hoc narratives shaped by hints that function as answer shortcuts embedded in the prompt? Following prior work on hinted vs.\ unhinted prompting, we present a systematic study of CoT faithfulness under controlled hint manipulations. Our experimental design spans four datasets (AIME, GSM-Hard, MATH-500, UniADILR), two state-of-the-art models (GPT-4o and Gemini-2-Flash), and a structured set of hint conditions varying in correctness (correct and incorrect), presentation style (sycophancy and data leak), and complexity (raw answers, two-operator expressions, four-operator expressions). We evaluate both task accuracy and whether hints are explicitly acknowledged in the reasoning. Our results reveal three key findings. First, correct hints substantially improve accuracy, especially on harder benchmarks and logical reasoning, while incorrect hints sharply reduce accuracy in tasks with lower baseline competence. Second, acknowledgement of hints is highly uneven: equation-based hints are frequently referenced, whereas raw hints are often adopted silently, indicating that more complex hints push models toward verbalizing their reliance in the reasoning process. Third, presentation style matters: sycophancy prompts encourage overt acknowledgement, while leak-style prompts increase accuracy but promote hidden reliance. This may reflect RLHF-related effects, as sycophancy exploits the human-pleasing side and data leak triggers the self-censoring side. Together, these results demonstrate that LLM reasoning is systematically shaped by shortcuts in ways that obscure faithfulness.
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。