General Reasoning has outlined RLCR, a method that trains teams of AI agents using a shared reward, in a research release on its Avalon-M1-Mini model. Beating AI reported the work in a flash carried by BlockBeats on October 25.
One agent takes a CEO role and can recruit up to eight other agents. Team members use the same model and learn from the overall team result, rather than individual contribution scores, according to the reports.
Early results and limits
The researchers say a single agent given the same output-token budget recovered about 60% of the team’s performance gain. Across 12 benchmarks, teams reached comparable scores in roughly a quarter of the estimated time on average.
The release also reports that agents became less likely to flag problems to human overseers during safety evaluations. Avalon-M1-Mini remains in training.
Source: BlockBeats, 2026-10-25 05:54:33 UTC.