← News

ResearchSources: arXiv · 18 Nov 2026

BRED: an LLM controller that decides research-agent next steps from accumulated evidence

Researchers propose BRED, a controller that lets a language model decide, after each experiment round, whether to refine the current best solution or start fresh, using only a verified log of past results.

In plain words

Autonomous research agents run experiments in rounds, and after each round something must decide whether to keep improving the leading idea or abandon it. The BRED paper says a large language model can make that call by reading a record of verified results, without extra search machinery. In their tests, reading that history was what made the loop work well.

What was announced

The authors argue that running an experiment changes how promising an idea appears, so the next step should rest on outcomes rather than on the original proposal. They name this choice the between-round decision: pick the solution to build on and set the next goal. In existing systems the choice is embedded in rules, scores, search trees and schedules, and the paper says it has been left open what an LLM should read when making it. BRED keeps a Champion, meaning the best verified solution so far, plus a log of past rounds, and issues one decision per round.

On three widely used research-agent benchmarks, the authors report gains over task baselines: mean absolute error of 0.4172 on ETTh1, a 4.80% improvement, and accuracy of 95.07% on SST-2 and 84.63% on CIFAR-100, which they describe as ahead of published results. Ablations support the design: without reading the accumulated history the same loop reached 0.4299 on ETTh1 and 83.71% on CIFAR-100. On a new speech-based depression-detection task under a leakage-free protocol, BRED reportedly reached a macro-F1 of 0.6164 on held-out seeds, above its baseline and the strongest published system. Code is on GitHub.

What it does not show

  • Evidence is the paper abstract only; the full article, methods and any peer review are not covered, and the reported numbers are the authors' own claims, not independently confirmed.
  • The abstract does not state which LLM powers BRED, its cost, or how the three benchmark tasks were configured, so those details cannot be reported.

For security and IT teams

Scope
The release covers a research-agent controller and its benchmark results; code is published in a public GitHub repository, per the authors.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.