IME, the thumbs don't do anything. The only way to have a real affect on your news feed is to block channels.
For example, I follow the Chicago Sun-Times. And every day Apple News shows me the Sun-Times horoscope. And every day I give it a thumb's down. Every. Single. Day.
What a waste. It would be trivial (and effective!) to simply implement a bayesian "spam" filter that operates on the text of headlines to promote or demote headlines according to user rankings. This could be done completely on-device with utterly negligible power draw.
Yet 400 boolean data points is nowhere near enough data for any kind of learning system to get anywhere. And without the cloud to aggregate data from millions of other users, it'll never become great.
From personal experience with a HN headline scraping IRC bot I wrote a few years ago, 400 labelled headlines is more than sufficient to get pretty good results. The threshold of usefulness should be closer to a few dozen rankings.
For example, I follow the Chicago Sun-Times. And every day Apple News shows me the Sun-Times horoscope. And every day I give it a thumb's down. Every. Single. Day.
400+ thumbs later, it still shows up every day.