Newest papers

SOMA: Continuous Monitoring and Intervention for System Two Reasoning with System One Models

Ruizhi Qian, Lei Shi, Tianyu Hu, Bo Yan, Song Wang, Yu Tian · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.26793 · cited by 0 works (OpenAlex) · open-access PDF

French

Dans les humains, la méta-cognition sépare la surveillance du contrôle: les signaux rapides et intuitifs gardent l'oeil pendant que le raisonnement lent, délibéré se déroule et aident à décider de continuer, de changer de cours ou d'arrêter. Les modèles de décision System One (S1) comme Jev sont un candidat naturel pour ce rôle: ils renvoient une probabilité calibré pour une question de langue naturelle sur un texte dans un appel rapide, mais ils ne sont utilisés que comme des portes avant le raisonnement. Nous introduisons SOMA (System One Metacognitive Awareness), dans lequel un modèle S1 surveille continuellement un modèle de raisonnement System Two (S2) et intervient lorsque nécessaire. SOMA a trois couches: une couche de scénario

Machine translation (Meta M2M100); the English original follows.

Abstract

Reasoning models often run into problems without noticing them, even when symptoms are visible in their own text. In humans, metacognition separates monitoring from control: fast, intuitive signals keep watch while slow, deliberate reasoning unfolds and help decide whether to continue, change course, or stop. System One (S1) decision models, such as Jev, are a natural candidate for this role: they return a calibrated probability for a natural-language question about a text in one fast call, yet they are used only as gates before reasoning. We introduce SOMA (System One Metacognitive Awareness), in which an S1 model continuously monitors a System Two (S2) reasoning model and intervenes when needed. SOMA has three layers: a scenario layer states each objective in natural language with predefined actions, an awareness layer lets S1 judge whether the described behavior appears, and a control layer accumulates these judgments, first reminds S2, and takes over only if the behavior persists. Sharing one S1 model and one controller across tasks, without training, SOMA addresses two different failures on two reasoning models: on overthinking, it cuts reasoning tokens by up to 36\% while improving average accuracy, and on unanswerable questions, it raises abstention from 70.5\% to 87.2\%.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Computer science, Metacognition, Soma, Control (management), Natural (archaeology), Artificial intelligence

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange