intrinsic本质的，固有的，内在的extrinsic外在的，外部的auxiliary辅助的contradiction矛盾entailment蕴含leverage借助已有的…實現新（或更好）的目標consolidate巩固，整合anisotropy各向异性alleviate减轻，缓解sustainable可持续性的plausible貌似合理的. e.g. sampling more contextually plausible noises from its masked language modeling.frontier前沿领域substantially显著地，可观地，e.g., ELECTRA substantially outperforms MLM-based methodsdedicated致力于 be dedicated to sth.resemble类似于rationale逻辑依据trait特点promising前景看好的，富有前景的，e.g., promising resultsregarding对于dilemma窘境paradigm范例persuasive有说服力的perceive察觉practitioners从业人员facilitate促进ad-hoc专用post-hoc事后，e.g., they design post-hoc separate models to explain the results generated by “black box”arouse引起advancement发展distill蒸馏，提取perspective观点，方面peripheral外围的，次要的，e.g. 外围设备reveal揭示Medoid数据集中的质心（但必须是数据集中的一个存在的数据点）elucidate阐明elaborate阐述harnesse加固prominent重要的，显著的vague含糊的- In this work, we shed light on these issues.
discrepancy差异rationality合理性remedy纠正，弥补marginally略微地significantly显著地sophisticated精密的， 复杂的，高雅的inevitably不可避免地realize将…变为现实heterogeneous异质的homogeneous同质的on the horns of a dilemma陷入两难，faced with the choice between two equally unpalatable alternativescomputationally intractable计算上不可行的hinder妨碍sub-optimal次优（不够好的意思）compensate补偿abundant大量的mimic模范albeit虽然manifest表明Inductive归纳式de facto事实上的，e.g.,de facto standarddiscrepancy差异subtle巧妙的，微妙的transductive直推，直推学习：训练时是知道测试集的数据的inductive归纳，归纳学习，也就是普通的监督学习Seamless无缝的，连续不间断的，e.g., Seamless integration 无缝集成dubbed简称为burden负担，e.g., relieve the computation burdenorthogonal正交的，e.g., contribution of LiteGT is orthogonal with related workhinder妨碍，阻挡inherently内在地derive获得，得到，e.g., derive a method，得到一种方法- The space of all permutations, however, grows factorially in the size of the input set, thereby making the stated learning task intractable.
cardinality数学术语，集合的元素数量vantage point有利地点