SOMA, short for System One Metacognitive Awareness, pairs a reasoning model with a faster monitoring model. The monitor watches for described behaviours during reasoning instead of acting only as a gate before the main model starts.
Three layers connect detection to control
The scenario layer defines an objective and available actions in natural language. The awareness layer asks the System One model whether the relevant behaviour appears. The control layer accumulates those judgments, first reminding the reasoning model and taking over only if the behaviour persists.
This separates recognising a symptom from deciding how strongly to intervene. A single monitoring result does not have to trigger the most disruptive action immediately.
What the reported tests show
The authors share one System One model and one controller across tasks without additional training. They evaluate two reasoning models on overthinking and unanswerable questions.
For overthinking, the paper reports reasoning-token reductions of up to 36% while improving average accuracy. For unanswerable questions, abstention rises from 70.5% to 87.2%. These figures describe the studied conditions; they are not a promise that every deployment will save the same amount or refuse every unsupported answer.
What to measure in a trial
Franklin’s interpretation is that the monitor’s interventions need their own evaluation. Record when a reminder helps, when taking over prevents a failure and when the system interrupts useful work. Also account for the monitor’s calls when estimating the total cost of the combined workflow.
A reduction in the main model’s reasoning tokens is only one part of that assessment. A team should examine task completion, answer quality, abstention on answerable questions and the frequency of unnecessary intervention together.
The framework offers a way to make monitoring and control explicit, but its published results do not remove the need to define the behaviour being monitored carefully. Read the SOMA paper for the scenarios, controller and reported experiments.
Comments