幻觉,单一事实和错误校准:一个经验调查
Miranda Muqing Miao1, Michael Kearns1
1Department of Computer and Information Science, University of Pennsylvania, Philadelphia, PA 19104.
概括
大型语言模型 (LLM) 幻觉事实,受单事实率和错误校准的影响. 选择性上加权将幻觉减少40%,同时保持准确性,挑战复制.
科学领域:
- 自然语言处理自然语言处理.
- 机器学习理论机器学习理论
- 统计学学习 统计学学习
背景情况:
- 大型语言模型 (LLM) 呈现幻觉事实,这一现象理论上与单事实率和模型误校准有关.
- 现有的研究表明,基于这些因素,幻觉的统计下限是基于这些因素.
- 要了解这种关系,需要在不同的模型架构中进行实证验证.
研究的目的:
- 在经典的n-gram和变压器模型中实证地研究单事实率,误校准和幻觉之间的关系.
- 为了开发和验证经验上的幻觉.
- 探索减轻幻觉的技术,而不会影响模型的准确性.
主要方法:
- 从帕雷托分布生成的训练数据来控制单事实率.
- 使用双智的库尔巴克-莱布勒 (KL) 分歧来得出经验幻觉的结合.
- 通过训练实例的战略重复引入和评估选择性上加权技术.
主要成果:
- 在单事实率和幻觉水平之间确立了正相关性.
- 经验幻觉边界证明了实际的可行性.
- 选择性加权减少了高达40%的幻觉,同时保持了准确性.
- 标准训练提高了准确性,但没有有效地减少持久幻觉.
结论:
- 在LLMs中的幻觉经验上与单事实率和模型校准错误有关.
- 选择性加权提供了一种可行的策略来缓解幻觉,挑战普遍的复制.
- 在标准培训目标中,提高准确性和减少幻觉之间存在一个权衡.
相关概念视频
Cause and Effect
While variables are sometimes correlated because one does cause the other, it could also be that some other factor, a confounding variable, is actually causing the systematic movement in our variables of interest. For instance, as sales in ice cream increase, so does the overall rate of crime. Is it possible that indulging in your favorite flavor of ice cream could send you on a crime spree? Or, after committing crime do you think you might decide to treat yourself to a cone?
Gestalt Principles of Perception
Gestalt principles provide a framework for understanding how humans perceive objects as unified wholes within their context. These principles are essential in explaining the cognitive processes that make sense of complex visual stimuli by organizing them into coherent groups. One fundamental principle is proximity, which posits that objects located close to each other are perceived as a collective group. For instance, when dots are positioned near one another, the visual system interprets them...
Extrasensory Perception
Extrasensory perception, or ESP, suggests the ability to perceive events beyond the conventional senses of sight, hearing, and touch. Parapsychologists, who research ESP and related psychic phenomena, categorize ESP into three main types: precognition, telepathy, and clairvoyance.
Precognition involves foreseeing future events, such as predicting an accident before it happens. An example of precognition could be someone dreaming about a specific event, like a car crash, which then occurs...
Precognition involves foreseeing future events, such as predicting an accident before it happens. An example of precognition could be someone dreaming about a specific event, like a car crash, which then occurs...
False Memories
False memories represent a cognitive distortion in which individuals recall events that did not happen, or remember them in an altered form. This phenomenon highlights the brain's constructive nature in processing and recalling memories, emphasizing that memory is not a perfect representation of past events but rather a dynamic reconstruction influenced by various factors.
One primary source of false memories is misattribution, where individuals incorrectly associate external information with...
One primary source of false memories is misattribution, where individuals incorrectly associate external information with...
Magical Thinking
Magical thinking encompasses the belief in assumptions that defy logical reasoning yet appear intuitively convincing. It is a common psychological phenomenon that persists across various cultural and individual contexts. While these assumptions contradict empirical evidence and scientific laws, they often serve meaningful psychological roles in promoting emotional resilience and a sense of control, especially under stress or uncertainty.Thought-Action Fusion and the Law of SimilarityA key...


