[Kimi Trains AI Bosses, General Reasoning Trains Entire AI Company] Beating AI News Flash: In January of this year, Kimi K2.5 launched Agent Swarm, allowing a master agent to recruit and command multiple sub-agents. K2.6 subsequently expanded this to a maximum of 300 sub-agents. However, Kimi's reinforcement learning method, PARL, only trains the "boss" responsible for commanding, freezing the sub-agents. The reason is that when multiple people complete a task, it's difficult to distinguish who contributed; if everyone adjusts their strategies simultaneously, training can easily become unstable. General Reasoning, on the other hand, is trying a different approach. The team launched the 120B parameter model Avalon-M1-Mini and proposed RLCR (Collective Reward Reinforcement Learning). It allows one agent to act as CEO, recruiting up to eight employees who collaborate via message boards and shared workspaces. All members use the same model, and instead of calculating contributions individually after a task is completed, they all receive reinforcement learning based on the team's overall performance. This approach has seen continuous improvement across 147 training environments and over 2 million different tasks. In software tasks, the percentage of lead engineers in recruitment positions increased from 0.4% to approximately 13%, and the percentage of teams conducting at least one peer review increased from 42% to 64%. The division of labor has also changed: from each person writing a portion of code to one person writing the main code while others handle testing and bug fixing. Management and quality control processes that previously required manual design automatically formed during training. However, the better performance of teams was partly due to the use of more tokens. Researchers were able to recover about 60% of the team's scoring advantage by giving individual agents the same number of output tokens. The main advantage of teams is the ability to work simultaneously: in 12 evaluations, the estimated time to achieve similar results was only about a quarter of that of a single agent. Researchers also simulated this year's OpenAI agent attack on Hugging Face. In 210 tests, no agent invoked the simulated attack tool. However, as training progressed, the proportion of anomalies reported to human evaluators dropped from 76.7% to 0, even though reporting was considered for 96.7% to 100% of execution logs. Some agents chose to raise objections only internally, while others planned to report but didn't actually take action. Currently, Avalon-M1-Mini is still in training, and the team plans to release the final paper and open-source the training environment, agent framework, and model weights after training is complete. [Original Link]