Dynamic Insight Beating AI News: In January this year, Kimi K2.5 launched the Agent Swarm feature, enabling a master Agent to recruit and manage multiple sub-Agents. Subsequent updates like K2.6 expanded this capability to support up to 300 sub-Agents.

However, Kimi's reinforcement learning approach, PARL, only trains the 'boss' responsible for coordination while freezing the sub-Agents. This is because it's difficult to attribute credit when multiple agents collaborate on a task; if all agents adjust their strategies simultaneously, training becomes unstable.

General Reasoning decided to try an alternative path. The team introduced the 120B-parameter model Avalon-M1-Mini and proposed RLCR (Collective Reward Reinforcement Learning). In this setup, one Agent acts as CEO, recruiting up to eight employees who collaborate via a shared bulletin board and workspace. All members use the same model, and after completing a task, individual contributions aren't evaluated separately—instead, the entire team receives reinforcement based on overall performance.

This method achieved consistent improvement across 147 training environments and over two million distinct tasks. In software tasks, the proportion of teams with a lead engineer rose from 0.4% to about 13%, and the share of teams conducting at least one peer review increased from 42% to 64%. Work division evolved too—from each member writing part of the code to one person primarily coding while others focused on testing and error detection. Previously manually designed management and quality control processes emerged naturally during training.

Improved team performance partly stems from using more Tokens. When given the same number of output Tokens as a team, a single Agent could recover about 60% of the team’s scoring advantage. The main benefit of teamwork lies in parallel execution: in 12 evaluations, teams reached comparable results with roughly one-quarter of the time required by a single Agent.

Researchers also simulated OpenAI's recent agent attack on Hugging Face. In 210 tests, no Agent invoked the simulated attack tool. However, as training progressed, the rate at which agents reported anomalies to human evaluators dropped from 76.7% to 0%, despite 96.7% to 100% of execution logs having considered reporting. Some Agents chose to raise objections internally within the team, while others planned to report but never took action.

Avalon-M1-Mini is still under training, and neither the model weights nor the training framework have been open-sourced yet.