RE2 prompt rereading hits diminishing returns fast

RE2 is one of those rare prompt hacks that is simple, measurable, and easy to test, but it is not a free lunch you can scale forever.

• –The original RE2 paper frames the gain as better input understanding in decoder-only models, not magic extra reasoning depth
• –Reported improvements were strongest on reasoning benchmarks, with mixed results on some ChatGPT tasks and modest gains rather than universal breakthroughs
• –The paper's own ablation on reread count found the sweet spot around 2 or 3 passes, with further repetition hurting performance
• –For practitioners, this makes RE2 a lightweight benchmark knob worth trying before heavier prompt chains or agent scaffolding
• –The Reddit post is useful because it pushes on the real engineering question: when does prompt augmentation stop helping and start distorting the model's behavior?