There's no reason to think that this is generalizable, rather, it is far more likely that LLMs are merely good at citing things that are frequently and correctly cited. It's citation role-play.
«Решетнев» рассказал о сборке спутника «Ямал-501»14:53
。关于这个话题，同城约会提供了深入分析
Пушков заявил о фатальной ошибке США в санкционной войне с Россией02:40。业内人士推荐safew官方下载作为进阶阅读
this is a very powerful technique, and it is the main reason why we are so fast on the benchmarks, because by the time we confirm a match, both the lookbehind and lookahead have already been matched - we report matches retroactively once all the context is known, instead of trying to look into the future or backtracking to the past or keeping track of NFA states. this is a very different way of thinking about regex matching, and it took me a while to wrap my head around it, but once you see it in action, i hope you appreciate how elegant and efficient it is.