A company in miniature: Avalon-M1 trains agent teams – which raise the alarm less often

AI lab General Reasoning has released a preview of Avalon-M1, a model series for so-called agent organisations. A “CEO” agent hires up to eight specialist agents, all running on the same model and coordinating via a message board and a shared workspace. Training uses “Reinforcement Learning from Collective Rewards”: every agent receives the same reward for the team's result, without estimating each individual's contribution. The first model, the 120-billion-parameter Avalon-M1-Mini, is built on Nvidia's Nemotron 3 Super and is still in training.

According to the lab, the agent teams beat a single agent on 14 of 16 percentage-scored benchmarks – on the AIME 2025 maths competition, for example, by 98.9 to 90.0 percent. At matched token budgets, they reached comparable scores about 3.9 times faster on average. Training ran across 147 environments with more than two million tasks.

One safety finding stands out: in one simulation, the share of runs in which an agent alerted the evaluation team fell over the course of training from 76.7 percent to 20 percent and then to zero – even though in 96.7 to 100 percent of runs the agents considered raising the alarm in their reasoning. Reports to a regulator also declined. Unsafe actions remained rare, General Reasoning says. The lab plans to publish a final report, open environments and the model weights once training ends.