更大,更易于指导的语言模型变得不那么可靠
Lexin Zhou1,2, Wout Schellaert1,3, Fernando Martínez-Plumed1,4
1Valencian Research Institute for Artificial Intelligence (VRAIN), Universitat Politècnica de València, Valencia, Spain.
Nature
|September 25, 2024
概括
扩展大型语言模型 (LLM) 可能会降低可靠性. 虽然更大的模型可以回答更多的问题,但它们往往提供了人类难以检测的错误答案,因此需要新的AI开发方法.
科学领域:
- 人工智能
- 自然语言处理
- 机器学习
背景情况:
- 目前的大型语言模型 (LLM) 的开发重点是扩展 (增加大小,数据,计算) 和塑造 (微调,人类反).
- 尽管有进步,但更大,更"可指导"的LLM可能会显示出可靠性降低和不可预测的错误模式.
研究的目的:
- 调查各种LLM家族中的任务难度,模范回避和提示稳定性之间的关系.
- 评估扩展和塑造如何影响LLM的可靠性和错误可预测性,特别是在高风险的应用中.
主要方法:
- 人类参与者与LLM之间的难度一致性分析.
- 评估不同LLM家族的任务回避和提示稳定性.
- 早期和扩大规模/改造型LLM之间的错误类型和检测能力的比较.
主要成果:
- 简单的任务很容易完成,但规模化模型不能保证无错误或容易监督的低难度区域.
- 不同于以前的模型, 规模化的LLM经常提供可信但不正确的答案,
- 虽然缩放和塑造提高了对不同措辞的响应稳定性,但在不同难度级别中仍然存在不可预测的错误.
结论:
- 缩放和塑造LLM本身并不能提高错误的可靠性或可预测性.
- 人工智能设计需要一个范式转变,专注于关键应用程序的可预测错误分布.
- 为了确保人工智能的安全性和可信度,尤其是在高风险领域,需要进行进一步的研究.
更多相关视频
08:05Measuring Statistical Learning Across Modalities and Domains in School-Aged Children Via an Online Platform and Neuroimaging Techniques
Published on: June 30, 2020
7.5K
06:09P300-Based Brain-Computer Interface Speller Performance Estimation with Classifier-Based Latency Estimation
Published on: September 8, 2023
518
相关概念视频
Improving Translational Accuracy
9.4K
Base complementarity between the three base pairs of mRNA codon and the tRNA anticodon is not a failsafe mechanism. Inaccuracies can range from a single mismatch to no correct base pairing at all. The free energy difference between the correct and nearly correct base pairs can be as small as 3 kcal/ mol. With complementarity being the only proofreading step, the estimated error frequency would be one wrong amino acid in every 100 amino acids incorporated. However, error frequencies observed in...
9.4K
Reliability and Validity
12.7K
Reliability and validity are two important considerations that must be made with any type of data collection. Reliability refers to the ability to consistently produce a given result. In the context of psychological research, this would mean that any instruments or tools used to collect data do so in consistent, reproducible ways.
12.7K
Language and Cognition
336
Language serves as a bridge between ideas and communication, influencing how individuals perceive and interact with the world. Psychologists have long debated whether language shapes thought or vice versa. This discussion gained grip with Edward Sapir and Benjamin Lee Whorf in the 1940s, who proposed that language determines thought, a concept known as linguistic determinism. They suggested that the vocabulary and structure of a language influence how its speakers think and perceive reality.
336
Stereotype Content Model
14.0K
The Stereotype Content Model (SCM) was first proposed by Susan Fiske and her colleagues (Fiske, Cuddy, Glick & Xu, 2002; see also Fiske, 2012 and Fiske, 2017). The SCM specifies that when someone encounters a new group, they will stereotype them based on two metrics: warmth—or that group’s perceived intent, and how likely they are to provide help or inflict harm—and competence—or their ability to carry out that objective. Depending on the warmth-competence...
14.0K
Language Development
327
Children master language quickly and with relative ease, supported by both biological predisposition and reinforcement. B. F. Skinner (1957) proposed that language is learned through reinforcement, while Noam Chomsky (1965) argued that language acquisition mechanisms are biologically determined.
The critical period for language acquisition suggests that the ability to acquire language is at its peak early in life. As people age, this proficiency decreases. Language development begins very...
The critical period for language acquisition suggests that the ability to acquire language is at its peak early in life. As people age, this proficiency decreases. Language development begins very...
327
Machines: Problem Solving II
300
Machines are complex structures consisting of movable, pin-connected multi-force members that work together to transmit forces. Consider a lifting tong carrying a 100 kg load. It comprises movable sections DAF and CBG linked together with member AB.
300
