Autonomous research loops hand the between-round call to one LLM: BRED reads only a champion and a full log, and beats published systems on four tasks
Synopsis
After each round of experiments, an autonomous research system must decide whether to refine its best solution or start a new direction, and existing designs fix that choice with rules, scores, or search trees. BRED gives this between-round decision to a single LLM call that keeps only the best verified solution (the Champion) and a per-round log, with no search structure around it. On the ETTh1, SST-2, and CIFAR-100 benchmark tasks its final Champion beats published systems and also beats InternAgent re-run with the same models; on speech-based depression detection under a leakage-free protocol, its macro-F1 on held-out seeds exceeds the strongest published system.
Interpretation
The between-round decision in an autonomous research loop, meaning whether to refine the current best solution or switch direction, which solution to build from, and what goal to set next, can now be made by a single LLM call without a search tree, scoring rule, or schedule wrapped around it. Earlier systems fixed this decision with a hard-coded policy, a UCT variant, a score rule that tightens with budget, or a schedule with a pre-set exploitation share; some let an LLM pick among nodes that already ran, but how much evidence the LLM needs to read had not been isolated. On the ETTh1, SST-2, and CIFAR-100 benchmark tasks, the final Champion's MAE and accuracy beat the published results and also beat InternAgent re-run with the same models.
The controller keeps only two objects per round: the Champion, the best verified solution stored with its code, metric, and a short mechanism summary, and a log with one entry per round holding the decision taken, the candidate results, and a lesson stating only measured facts. One LLM call reads both the Champion and the full log and returns an action, a goal-level instruction, optional hard constraints, and a rationale; the action is one of starting a new line from the baseline, extending the Champion with a new mechanism, refining the Champion without a new mechanism, or stopping. On ETTh1 and CIFAR-100 only what the decision step sees is varied, across no history, the last round only, and the full history; performance rises monotonically with the evidence shown, and the full controller finishes best on both tasks.
Perspective
The demonstrated scope is four tasks: ETTh1 time-series forecasting, SST-2 sentiment classification, CIFAR-100 image classification, and speech-based depression detection on DAIC-WOZ outside the benchmark; all four share one controller and prompt templates, differing only in task configuration and execution environment. The depression-detection result is measured under a leakage-free protocol on five seeds held out from selection, and the authors state the model is a research artifact for comparing controllers, not a screening or diagnostic instrument.
All runs use one controller LLM, and how its decision quality changes with other models is not measured; the ablation covers two tasks and shows the gain grows with context but does not decompose which part of the accumulated evidence contributes most. The lessons in the log are written by an LLM; they are restricted to facts, but a wrong lesson would still reach later decisions. Dolphin's numbers come from runs with other models and other hardware.
