[throwaway202351]

meta: This is a paper that uses "landmarks" in text as a way to be able to scan an arbitrary part of the text, it is not necessarily a "landmark... transformer paper" as the post title may suggest.

[arijun → throwaway202351]

Thanks for that, I misunderstood in exactly that way.

@dang the HN title is confusing as is, can we change to the paper title or modify so it’s less confusing (and clickbaity)?

[bigyikes → throwaway202351]

Yeah, this got me. Especially because the title is already editorialized ("just dropped"), I assumed "landmarks" was further editorialization.

[anigbrowl]

The paper title is 'Landmark Attention: Random-Access Infinite Context Length for Transformers'. OP, please don't use marketing-speak to write submission titles on HN, it's an unhelpful distraction from the subject matter.

[jumploops]

> Our method uses a landmark token to represent each block of the input and trains the attention to use it for selecting relevant blocks

Landmark is used here to describe an aspect of this new work, not necessarily a description of this paper’s overall impact.