safety
A Fast Monitor Interrupts the Slow Reasoner
SOMA watches a reasoning trace, warns first and takes over only when a named failure persists.
Summary
SOMA watches a reasoning trace, warns first and takes over only when a named failure persists.
SOMA pairs a fast decision model with a slower reasoner through scenarios, continuous judgments and predefined interventions. Without task-specific training, the shared monitor cut reasoning tokens by as much as 36 percent while improving average accuracy on overthinking tests, and raised abstention on unanswerable questions from 70.5 to 87.2 percent. It depends on visible text and calibrated monitors; it is not oversight of hidden reasoning.
Why it matters
SOMA watches a reasoning trace, warns first and takes over only when a named failure persists.
Limits and context
- It depends on visible text and calibrated monitors; it is not oversight of hidden reasoning.
Key claims
SOMA watches a reasoning trace, warns first and takes over only when a named failure persists.
Qualification: It depends on visible text and calibrated monitors; it is not oversight of hidden reasoning.
Evidence: source-2026-10-20-019
Sources
- arXiv preprint 2610.26793, original v1arXiv · primary research
Corrections
No corrections have been recorded for this story.