Paper Lantern lifts coding-agent tuning 3.2%

This feels less like a prompt hack and more like search-space expansion: Paper Lantern didn’t just give the agent more context, it gave it access to newer and more specialized optimization moves. The interesting part is that the gain came from better decisions under compute pressure, not from brute-force trial and error.

• –The paper-backed run surfaced post-cutoff ideas like AdaGC, WSD cooldown, REX, and the sqrt batch-scaling rule, then applied them correctly on the first attempt.
• –The no-paper run stayed inside the standard ML playbook and diverged when batch size was cut without the matching learning-rate adjustment.
• –Not every paper idea worked; DyT and SeeDNorm were rejected as architecture mismatches, which is a good sign that the system is synthesizing rather than blindly copying.
• –The evidence is still early: one run per condition, a tiny 7M-parameter model, and some of the gain may come from extra reasoning time spent on each technique.