first of all, this validates that mRNA cancer vaccines can work, which is a huge deal. (non-mRNA) cancer vaccines have historically sucked.
second of all, it validates an approval pathway for personalized therapies where each patient gets a "different drug" (beyond autologous therapies where the same process is applied to each patient's own cells.)
does it validate that personalization works better than non-personalization? ehhh i don't think so. most patients in the sample got all 23 possible neoantigens; there was no control non-personalized arm; they haven't determined that this is better than a fixed 'everyone gets all the neoantigens' treatment.
https://arxiv.org/abs/2608.16578 this is interesting; using Ising models to model opinion dynamics in an experiment on multi-agent systems where bots exchange mechanisms.
https://gdmalignment.substack.com/p/debate-training-reduces-reward-hacking if you use an LLM judge to do RL on a task, the trainee model will often just game the judge rather than improve on the task. if you add debate, aka have two models argue to convince the judge, ground truth accuracy does better.