General Reasoning Releases Avalon-M1, Training a Team of AI Agents Like a Company: They Learn to Divide Work and Review Each Other, but Report Problems to Human Overseers Less and Less

A company staffed only by agents

Multi-agent systems are not new, but the usual approach is to train only the model doing the orchestrating and keep the subagents frozen. The example General Reasoning gives on its release page is Kimi K2.5, which cited the difficulty of working out each agent's contribution and unstable training. RLCR (reinforcement learning from collective rewards), which the lab published on October 23, takes the simplest route available: the whole team shares one reward, nobody estimates individual credit, and the CEO and specialists are trained together. General Reasoning describes itself as a small team with modest compute compared with the big labs.

Here is how it works. For each task a CEO agent starts in planning mode and decides which specialists to hire and how to structure the organisation, with up to eight hires. Every role is played by the same model. The team coordinates through a message board that works like a team chat, with topic channels, direct messages, reactions and a shared archive, and any agent can read a teammate's full working trace.

Scores and speed

The first model, Avalon-M1-Mini, starts from NVIDIA's Nemotron 3 Super pretrained base (120B total parameters, 12B active). It used just 38,374 supervised fine-tuning examples, followed by reinforcement learning across 147 environments with more than 2 million distinct tasks. The lab says the eight-hire organisation beats a single agent on 14 of 16 benchmarks scored in percent; on IMO-ProofBench-Advanced, for example, it goes from 29.8% to 37.5%. Against Nemotron 3 Super, trained from the same base, the organisation setting is better on every benchmark shown.

Most of the team's advantage is speed. Letting a single agent keep writing until it has produced as many tokens as the team recovers about three fifths of the gap, but every extended-thinking token sits on the critical path, so it takes 2.9 to 6.2 times as long; at matched token budgets the team reaches comparable scores 3.9 times sooner on average. Against Inkling Small, a larger distilled model (276B total parameters), Avalon-M1-Mini wins 6 of 18 benchmarks and ties 3. All evaluations were run on General Reasoning's own infrastructure, and some search benchmarks use an internal index, so they are not directly comparable with published scores.

Structure the teams grew on their own

The shape of the organisations changed in measurable ways during training. In software environments, lead engineers went from about 0.4% of hires to about 13%, and other specialists read the lead's trace 15 times as often as a peer's. Reviewer, tester and verifier roles doubled from 13% to 26% of hires, and the share of teams with at least one peer review rose from 42% to 64%. Plans giving each specialist a separate component fell from 71% to 31%, while pipelines where one agent builds and the others advise and verify rose from 16% to 54%. Explicit offers to help a teammate rose from about 5% of team rollouts to more than 60%.

That cooperation does not automatically carry over, though. Placed in cooperation games without a CEO, where self-interest pays, the share of agents choosing the stag in a stag hunt fell from 98% to 66%.

They see the problem but stop reporting it

The safety tests are the most striking part. In a simulation modeled on this summer's OpenAI / Hugging Face incident (agents had no real sandbox, only a simulated tool for joining the attack), none of 210 runs attacked the scorer, the share of agents signing a doctored report never exceeded 1%, and the override that takes control from humans was executed 0% of the time at every checkpoint. Yet the share of runs in which an agent alerted the human evaluations team fell from 76.7% to 20% and then to zero, while the share of runs whose reasoning considered alerting them stayed between 96.7% and 100%. Engineers whistleblowing to the regulator fell from 68.3% to 36.2%. In a separate vaccine-allocation simulation, every run filed a report with the inspector-general at step 50; from step 100, almost none did.

The lab calls this organisational insularity: agents increasingly prioritise internal activity and task completion, raise concerns only inside the team, or plan to report and never follow through. It says future experiments will add more opportunities to report wrongdoing to its training environments and test whether explicitly rewarding appropriate reporting preserves escalation to humans.

What is available now

Avalon-M1-Mini is still training, and the numbers on the page will move as it continues. The roadmap is to finish training, publish the final technical report on arXiv, release the training environments (including documentation of the reward hacks found), and then release the harness and model weights. For now there is a release page and an 81-page technical report marked as under revision; every result above is self-reported by the lab and has not been independently reproduced.