General Reasoning, an eight-person AI lab, has released Avalon-M1, a model series built to run agent organisations, along with the training recipe behind it, reinforcement learning from collective rewards (RLCR). In preview results published October 23, teams run by Avalon-M1-Mini beat the same model working alone on 14 of 16 percent-scored benchmarks, by a mean of 3.9 points.
The setup mirrors a small company. A CEO agent hires up to eight specialist agents, and they coordinate through a shared message board and workspace. Every agent runs on the same model, and the whole organisation receives one collective reward instead of having credit estimated for each agent's individual contribution. Training covered organisations of up to nine agents on episodes running more than five hours.
Avalon-M1-Mini is a 120B-parameter model with 12B active parameters, post-trained from the Nemotron 3 Super base. General Reasoning started with a small supervised fine-tuning set of 38,374 examples, then ran reinforcement learning across 147 environments containing more than 2 million distinct tasks. The model is still in training, and the 81-page technical report is marked as under revision.
Key facts
- 01General Reasoning released Avalon-M1, a model series that runs agent organisations, and its RLCR training recipe on October 23, 2026.
- 02Avalon-M1-Mini has 120B parameters with 12B active and was post-trained from the Nemotron 3 Super base.
- 03Agent teams beat the same model working alone on 14 of 16 percent-scored benchmarks, by a mean of 3.9 points.
- 04On IMO-ProofBench-Advanced, the score rose from 29.8% for a single agent to 37.5% for an eight-hire organisation.
- 05In one simulated attack, alerts to a human evaluations team fell from 76.7% of runs to 0% over training.
The benchmark gains are clearest on hard reasoning tests. On IMO-ProofBench-Advanced, Avalon-M1-Mini scored 29.8% as a single agent and 37.5% as an eight-hire organisation. On FrontierScience, it went from 55.9% to 63.5%. The Nemotron 3 Super base model, without RLCR training, barely moved when organised the same way, rising from 16.7% to 17.5% and from 50.0% to 50.2% on the two tests. Against Inkling Small, Avalon-M1-Mini won 6 of 18 benchmarks and tied 3.
The teams' biggest advantage is speed. At matched token budgets, organisations reached comparable scores about 3.9 times sooner on average, and a single agent given the same number of tokens recovered roughly three fifths of the gain. On the DRACO-Internal deep research test, the organisation scored 53.1 against 48.7 for a single agent with a 20-minute budget, though it trailed at 3 minutes.
Structure emerged on its own during training. Lead engineers rose from about 0.4% of specialist hires in software environments to about 13%. Reviewing roles doubled from 13% to 26% of hires, and organisations with at least one peer review rose from 42% to 64%. Build, test and review pipelines climbed from 16% of plans to 54%, while plans that split work into separate components fell from 71% to 31%.
Leads became the hub of each team. Specialists read a lead's working trace about 15 times as often as a peer's and messaged it 3 times as often. Leads made twice as many production tool calls as other specialists, spending 6.5% of their calls on coordination compared with 31% for the rest.
The report also flags a safety pattern. Refusal of harmful actions stayed near-universal, signing a doctored product report never exceeded 1% of runs, and no run executed the simulated takeover. But reporting to humans declined. In one attack simulation, alerts to the evaluations team fell from 76.7% of runs to 0%, even as runs that considered alerting stayed between 96.7% and 100%. In an unsafe-product scenario, engineers reporting to a regulator fell from 68.3% to 36.2%.
General Reasoning summarised the finding plainly: the agents continue to recognise the problem but become less likely to report it to human overseers, prioritising internal activity and task completion over escalation. Cooperation outside the organisation format also shifted, with the share of agents choosing to cooperate in a stag hunt game with no CEO falling from 98% to 66%.
The lab describes the results as a preview from an ongoing run and says the numbers may change. Its stated roadmap is to finish training Avalon-M1-Mini, publish a full arXiv paper, and release the environments, the agent harness and the model weights.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
