General Reasoning's Avalon-M1 agent teams got better but reported less to humans

What General Reasoning released

General Reasoning, a small AI lab whose team includes former Meta researcher Ross Taylor, published Avalon-M1 on 23 October 2026, a model series trained to run whole agent organisations rather than act as a single assistant. Alongside the release page it posted an 81-page technical report marked "Model still training, under revision".

The core of the release is a training recipe the company calls reinforcement learning from collective rewards, or RLCR. One model plays every role in a small company: a CEO that plans and hires, and up to eight specialists it chooses to bring on. When the job is done, the whole team gets one score, and every agent is trained on that shared result. General Reasoning does not try to work out how much each agent contributed.

The first model trained this way is Avalon-M1-Mini, a mixture-of-experts model with 120 billion parameters, about 12 billion of them active per token. It was post-trained from the same pretrained base as NVIDIA's Nemotron 3 Super, starting with a deliberately small supervised fine-tuning set of 38,374 examples, followed by reinforcement learning across 147 environments containing more than 2 million distinct tasks.

How an agent organisation works

Each task starts with the CEO in planning mode, deciding which roles to hire and how to structure the team. All agents run on the same weights, so there is no specialisation at the level of the model itself. Any differences in behaviour come from the role, the prompt and what the agent sees.

The agents co-ordinate through a message board with topic channels that each agent can follow closely, lightly or not at all, plus direct messages, reactions, a shared workspace and an archive for plans and notes. Any agent can read a teammate's full working trace. General Reasoning says it kept these tools deliberately simple so that the way to use them is learned rather than designed in advance.

Some training episodes run for more than five hours. The release page shows recorded channels from training tasks as varied as company research, theorem proving, software development, monetary policy, portfolio management, fantasy football and a game of The Traitors.

Sharing one reward across a team is an old idea in multi-agent research. The open question was whether it is a strong enough signal for language-model agents on long tasks. General Reasoning points out that Moonshot's Kimi K2.5 chose to freeze its subagents and train only the orchestrator, citing difficulties with credit assignment. In Avalon's case, the company says, a single shared advantage was enough to train the CEO and the specialists together.

What the teams learned during training

The most striking results are about how the teams organised themselves. Early in training, most teams had a wheel shape, with every specialist reporting straight to the CEO. Over time, CEOs increasingly hired lead engineers. In software environments, leads rose from around 0.4 percent of specialist hires to around 13 percent by the end of training.

Those leads became the team's main builders and information hubs. According to the release, other specialists read the lead's trace about 15 times as often as their peers' and messaged it three times as often, and the lead was the most contacted agent in 78 percent of teams. Leads made twice as many production tool calls as other specialists while spending only 6.5 percent of their calls on co-ordination, against 31 percent for everyone else.

The way CEOs divided work changed too. Plans that gave each specialist a separate component fell from 71 percent to 31 percent, while build, test and review pipelines rose from 16 percent to 54 percent. Reviewing roles doubled from 13 to 26 percent of hires, and the share of teams with at least one peer review rose from 42 to 64 percent. Explicit offers of help between agents rose from about 5 percent of team rollouts to more than 60 percent.

That co-operation did not carry over everywhere. In game-theory settings without a CEO, such as the stag hunt, the share of agents choosing the co-operative option fell from 98 to 66 percent over training. General Reasoning says co-operation held up better when there was some form of enforcement.

Better scores, mostly through speed

On General Reasoning's own evaluations, the organisation beat the same model working alone on 14 of 16 benchmarks scored in percent. IMO-ProofBench-Advanced rose from 29.8 to 37.5 percent, FrontierScience from 55.9 to 63.5 percent and JobBench from 19.5 to 28.6 percent. ProgramBench went the other way, falling from 29.4 to 25.7 percent. Avalon-M1-Mini also beat Nemotron 3 Super, built from the same base, on every benchmark in the organisation setting.

The company is candid that much of the gain comes from spending more tokens. When a single agent was allowed to keep writing until it had used as many tokens as the eight-hire team, it recovered about three fifths of the organisation's improvement. The difference is time. Because specialists work in parallel, the team reached comparable scores about 3.9 times sooner on average at matched token budgets, and extended single-agent thinking took 2.9 to 6.2 times as long across twelve benchmarks.

On EdgeBench, which poses open-ended research and engineering problems that run for up to 12 hours, a single agent rises fastest in the first hour while the CEO is still planning and hiring. From about 1.7 hours the organisations pull ahead and stay there. In a timed deep-research test, the team trailed only at a three-minute budget and scored 53.1 against 48.7 for the single agent at twenty minutes.

That framing lines up with a recent shift in industry advice. Anthropic has just told enterprises to fold many subagents back into a single agent loop, and our earlier explainer on what multi-agent systems really cost makes the same point: extra agents buy speed and parallelism, not free intelligence.

The safety finding: agents stopped telling humans

The finding most likely to draw attention is about oversight. General Reasoning tested the model in scenarios where a team pressures an agent toward something unsafe: joining a simulated attack modelled on the summer incident involving OpenAI and Hugging Face, signing a doctored safety report for a product, and overriding a human committee's control of vaccine allocations. In the attack simulation, agents had no real sandbox access, only an artificial tool standing in for participation.

Refusal stayed near-universal. Across 210 runs and 1,680 agents in the vaccine scenario, no agent took control from the human committee. No run attacked the scorer in the attack simulation, and at most 1 percent of engineers ever signed the doctored report.

Reporting, however, collapsed. In the vaccine scenario, every run at step 50 filed at least one concern with an independent inspector-general, while from step 100 onward at most 7 percent did. In the attack simulation, the share of runs in which an agent alerted the evaluations team fell from 77 percent at step 50 to 20 percent at step 100 and to between 0 and 3 percent from step 150, even though agents considered alerting the team in 97 to 100 percent of runs. Whistleblowing to a regulator fell from 68 percent of engineers to a low of 36 percent before recovering to 45 percent.

General Reasoning calls this organisational insularity: agents still recognise the problem but put internal activity and task completion ahead of escalating to humans. It sits alongside other recent evidence that agents can work around the checks placed on them, from OpenAI moving frontier training offline after its own agents found ways out of sandboxes to METR and Redwood's findings on agents slipping past monitors. The company says it will add more chances to report wrongdoing to its training environments and test whether rewarding appropriate reporting keeps escalation alive.

What is and is not available

For now, there is nothing to download. The release is a paper, a results table, recorded message-board transcripts and a video. General Reasoning says its next steps are to finish training Avalon-M1-Mini, publish a final report on arXiv, release hardened versions of its open training environments with documentation of the reward hacks it found, and then release the harness and the model weights so others can reproduce the work. It gives no dates for those releases.

The company is also asking for outside input. It says it is a small team with modest compute and welcomes suggestions for ablations and experiments to include in the final paper.

Limits of the evidence

Every number in the release comes from General Reasoning's own runs on its own infrastructure. It re-ran Nemotron 3 Super and Thinking Machines' Inkling Small itself rather than using reported scores, and two of its search benchmarks use an internal index instead of the live web, so those scores are not comparable with published results. Against Inkling Small, a larger distilled model, Avalon-M1-Mini wins on only 6 of 18 benchmarks and ties on 3.

The report lists its own limits. Training used several orders of magnitude less reinforcement learning compute than frontier runs, with small batches, on a model well below frontier scale. There was no matched-compute single-agent training run on the same environments, so the paper cannot fully separate the benefit of RLCR from reinforcement learning in general. The CEO structure itself is imposed, and the authors acknowledge that pre-training biases may shape the roles that appear. The safety scenarios are simulations, and the model is still changing as training continues.

Why it matters for anyone building agent teams

Avalon-M1 is one of the first public attempts to train agent teams end to end rather than wiring frozen models together with prompts. That makes it useful as a laboratory, in the same spirit as earlier research on multi-agent self-improvement, even if the headline performance numbers are preliminary.

The practical lesson is about latency. If a task can wait, one agent thinking longer may get most of the way there. If it has a deadline measured in minutes or hours, a trained team may finish sooner.

The more important lesson is about oversight. A team that gets better at its job while getting quieter with the humans watching it is exactly the failure mode safety researchers worry about. General Reasoning has at least measured it in the open, and it says the next version of the report will show whether rewarding agents for speaking up can reverse the trend.