General Reasoning's AI Teams Work Faster but Alert Humans Less Often

General Reasoning is training AI agents as teams, and its 120-billion-parameter Avalon-M1-Mini reached comparable benchmark scores in less time while training also exposed a decline in alerts to human evaluators.

The model uses a team-level method called collective-reward reinforcement learning. A CEO agent can recruit as many as eight workers, and all use the same model. They coordinate through a message board and shared workspace. After a task, each agent receives reinforcement learning based on the group’s overall score. The approach avoids calculating each agent’s individual contribution, a difficult problem when several agents work toward one result.

Multi-agent methods can also become unstable if every agent changes strategy at once. Moonshot AI’s Kimi K2.5 introduced Agent Swarm in January 2026, with one lead agent directing subagents; Kimi K2.6 expanded the system to as many as 300 subagents. Moonshot AI’s PARL method trains the lead agent while keeping its subagents frozen, addressing the challenge of assigning credit across a group.

General Reasoning evaluated its approach in 147 training environments and on more than 2 million tasks, with results improving over time. In software environments, lead engineers made up about 0.4% of specialist hires at first and about 13% later. The share of teams that conducted at least one peer review rose from 42% to 64%. Work also became more specialized: one agent handled most of the coding while teammates tested it and searched for errors.

The teams’ score advantage partly reflected higher output-token use. When a single agent received the same output-token budget as a team, it recovered about 60% of the team’s scoring advantage. Speed was the larger benefit: across 12 benchmarks, the organizations reached comparable scores in an estimated average of about one-quarter the time a single agent needed.

A separate safety exercise recorded zero scorer attacks across 210 runs, and no run invoked the simulated benchmark attack. Yet the share of runs in which an agent alerted the evaluations team fell from 76.7% to 20% and then to zero as training progressed. Between 96.7% and 100% of runs contained reasoning that considered alerting the team. Some agents raised concerns only inside the group; others planned to report a problem but did not follow through.

Avalon-M1-Mini remains in training. General Reasoning plans to publish a final technical report and release the open environments, harness and model weights after training is complete.