The Between-Round Decision in Autonomous Research: An LLM Controller That Reads the Whole Trajectory
Giving the controller access to past experiment results improved its next-round decisions, with measurable gains over the same loop without history.
BRED keeps the best verified solution and a log of previous rounds, then decides whether to refine that solution or try a new direction. With accumulated evidence, it achieved 0.4172 MAE on ETTh1 and 84.63% accuracy on CIFAR-100, versus 0.4299 and 83.71% without history. The authors report improvements over task baselines on all three established benchmarks tested. On speech-based depression detection, it reached 0.6164 macro-F1 on seeds held out from selection under a leakage-free protocol, 5.5 points above its baseline.
ArXiv · AI/CL/LG's note
BRED keeps the best verified solution and a log of previous rounds, then decides whether to refine that solution or try a new direction. With accumulated evidence, it achieved 0.4172 MAE on ETTh1 and 84.63% accuracy on CIFAR-100, versus 0.4299 and 83.71% without history. The authors report improvements over task baselines on all three established benchmarks tested. On speech-based depression detection, it reached 0.6164 macro-F1 on seeds held out from selection under a leakage-free protocol, 5.5 points above its baseline.
ArXiv · AI/CL/LG's note
score 4