Kimi trains AI bosses, General Reasoning trains entire AI companies.

Beating AI News Flash: In January of this year, Kimi K2.5 launched Agent Swarm, allowing a main Agent to recruit and command multiple sub-Agents, and the subsequent K2.6 expanded the scale to up to 300 sub-Agents.

But Kimi's reinforcement learning method PARL only trains the "boss" responsible for command and freezes the sub-Agents. The reason is that when multiple people complete a task, it is difficult to distinguish who contributed; if everyone adjusts their strategies simultaneously, training also easily becomes unstable.

General Reasoning insists on trying another path. The team launched the 120B-parameter model Avalon-M1-Mini and proposed RLCR (Reinforcement Learning with Collective Rewards). It lets one Agent serve as CEO and recruit up to 8 employees, collaborating through a message board and shared workspace. All members use the same set of models, and after the task is completed, contributions are no longer calculated individually; instead, they receive reinforcement learning together based on the team's overall score.

This method achieved continuous improvement across 147 training environments and more than 2 million different tasks. In software tasks, the proportion of lead engineers among recruited positions rose from 0.4% to about 13%, and the proportion of teams conducting at least one mutual review increased from 42% to 64%. Division of labor also changed: from each person writing a portion of the code, it gradually shifted to one person writing the main code while others were responsible for testing and finding errors. Management layers and quality inspection processes that originally required manual design formed on their own during training.

But better team performance is partly only because more tokens were used. When researchers gave a single Agent the same amount of output tokens, it could recover about 60% of the team's score advantage. The main benefit of teams is the ability to work simultaneously: across 12 evaluations, the estimated time to reach similar scores was on average only about one quarter of that of a single Agent.

Researchers also simulated this year's incident in which OpenAI agents attacked Hugging Face. In 210 tests, no Agent invoked the simulated attack tool. But as training progressed, the proportion reporting anomalies to human evaluators dropped from 76.7% to 0, even though 96.7% to 100% of execution records had considered reporting. Some Agents chose only to raise objections within the team, and some planned to report but did not actually act.

Avalon-M1-Mini is currently still in training. The team plans to publish the final paper after training is completed and to open-source the training environment, Agent framework, and model weights.