Papers I've read and am thinking about
I don't do formal academic research, but I read papers when something comes up in practice that I can't explain from first principles. My reading clusters around three areas.
PagedAttention and the follow-on work on prefix caching are the most directly relevant to what I've been building. The core insight — that KV cache is just paged virtual memory for attention — is one of those ideas that seems obvious in retrospect but clearly wasn't. Most of what Helios does is a working-through of these papers rather than original research.
MLIR, XLA, and the broader question of how much of “optimize the compute graph” is solved by compiler passes vs. still requiring hand-tuning. My tentative view is that compilers help a lot for standard architectures and much less for anything novel, but I'm still forming opinions. Operator fusion and memory planning are the parts I understand best.
Adjacent to security work I've done. The gap between the threat models in most robustness papers and real attack surfaces is large enough that I'm skeptical of a lot of the benchmarks — but the underlying question of how to reason about model behavior under distribution shift is genuinely interesting and probably unsolved.