Research
PILOT Adds a Supervisor That Can Abort a Running Worker, Gaining 9.8 Points on Terminal-Bench 2.0 While Cutting Output Tokens 42.9%
arXiv 2608.26530 argues self-improvement should happen during a run, not after it, and builds a supervisor-worker harness with two mechanisms: live steering (a separate supervisor redirects or kills the active worker mid-execution) and live self-evolution (distilling procedures and failure modes into reusable skills and memory as they appear). PILOT ranks first in five of six configurations across two frozen backbones and three benchmarks, beating counterpart harnesses by up to 9.8 points on Terminal-Bench 2.0 and gaining 14.6 points with GLM-5.1. Mean output tokens fall 42.9-47.4% while successful evaluations per million output tokens rise 110-134%.
↳ Follow the thread