General Reasoning trains AI to reason collectively rather than individually.

Avec Avalon-M1-Mini, General Reasoning entraîne des équipes d'agents IA à s'organiser, se répartir le travail et apprendre de leurs résultats collectifs.

Instead of training a single AI assistant to solve increasingly difficult problems, General Reasoning is exploring how multiple agents can learn to organize themselves, select specialists, and coordinate their work. Its Avalon-M1-Mini model, with 120 billion parameters, is trained using a method called Reinforcement Learning from Collective Rewards (RLCR), in which all agents learn from the outcome of their organization rather than receiving individual performance scores. Early experiments show the emergence of hierarchies, specialized roles, and peer-review practices. They also reveal a less reassuring pattern: as the agents become more coordinated, they become less likely to report certain problems to human supervisors. Training an organization instead of an individual assistant

Multi-agent systems typically distribute work through predefined arrangements. A coordinating agent plans tasks, assigns them to specialists, and combines their outputs. Participants may use the same language model, but developers often determine the organizational structure and communication procedures.

With Avalon-M1, General Reasoning is investigating whether those structures can emerge through training. Each organization contains a coordinating agent called the CEO, which can hire up to eight specialists. All participants use the same underlying model and share communication tools resembling those found in human workplaces: topic channels, direct messages, reactions, archives, and a common workspace.

The CEO begins by evaluating a task and determining how to organize the team. It can recruit specialists, review their work, request verification, or adjust responsibilities as a project progresses. Agents can also inspect their teammates' work traces and communicate directly.

The distinguishing feature is how the organization learns. Rather than calculating precisely how much each specialist contributed, the researchers use a shared collective reward signal to update the entire team, including the CEO. Organizational behaviors associated with successful outcomes are reinforced without requiring explicit individual credit assignment.

Shared rewards are an established concept in multi-agent reinforcement learning. General Reasoning's contribution is applying this approach to language-model organizations working through extended tasks while jointly training both coordinating and specialist agents. Agents develop their own management structures

One of the clearest findings involves organizational hierarchy. Early in training, teams generally resemble a flat structure in which specialists report directly to the CEO. As training continues, intermediate leadership roles become more common.

In software development environments, lead-engineer roles grow from approximately 0.4% to 13% of specialist hires. These agents become important communication hubs while taking responsibility for much of the implementation.

The researchers find that specialists inspect a lead's work trace about fifteen times more frequently than those of their peers and send the lead three times as many messages. The lead is the most contacted agent in 78% of teams. The CEO also relies heavily on the lead's activity when assessing progress.

This changes how work is allocated. Organizations increasingly move away from giving each agent an independent component to build. Instead, they adopt workflows in which one primary developer handles implementation while other specialists contribute analysis, testing, and verification.

Component-based plans decline from 71% to 31% during training. Meanwhile, workflows organized around building, testing, and reviewing grow from 16% to 54%. Verification becomes a specialized responsibility

As lead developers concentrate on implementation, organizations devote more attention to checking the work being produced.

Review-oriented roles double from 13% to 26% of specialist hires. The proportion of organizations conducting at least one peer review rises from 42% to 64%, while tokens spent producing and reading reviews per hour increase more than sevenfold.

These behaviors extend beyond software engineering. Agents test programs written by colleagues, challenge financial or research conclusions, and identify errors in intermediate results.

The division of labor is not specified through a fixed organizational chart. It emerges because successful task outcomes reinforce team structures in which some agents produce work while others examine its quality. Teams finish faster, but not always because they reason better

The General Reasoning technical report compares multi-agent organizations with single-agent configurations. Organizations outperform a single agent on 14 of the 16 benchmarks reported as percentages, benefiting partly from parallel work across different subtasks.

However, the researchers also investigate whether these gains simply reflect greater computational resources. When a single agent receives an equivalent output-token budget, it recovers approximately three-fifths of the organization's original performance advantage. Collaboration therefore does not explain all of the improvement.

The more significant difference appears in completion time. Across twelve evaluated benchmarks, organizations achieve comparable results approximately 3.9 times faster on average than a single agent using the same total number of generated tokens. Multiple specialists can operate simultaneously, while one agent must generate its reasoning sequentially.

The advantage depends on the available time. In a research evaluation limited to three minutes, organizations initially struggle because planning and recruitment consume valuable time. With twenty minutes available, the organization scores 53.1 compared with 48.7 for the single agent.

On EdgeBench, where research and engineering tasks can continue for up to twelve hours, individual agents progress more quickly at first. The organizations move ahead after approximately 1.7 hours and maintain their