Aim for insights over minor improvements
Currently, there are too many research papers showing minor "state-of-the-art" benchmark improvements for LLMs. These are not interesting.
Instead, focus on papers that yield new insights on models.
Linus @thesephist