Consciousness, a central element of human cognition, has been studied with multiple scientific approaches spanning neuroscience, psychology, artificial intelligence and robotics. Unfortunately, poor integration between these fields limits a full and clear understanding of consciousness. Here we contribute to improving this integration by proposing, within a neurocomputational framework, the `Goal-Aligning Representations Internal Manipulation' (GARIM) theory of consciousness. The central idea of the GARIM theory is that consciousness supports the active manipulation of goal-relevant internal representations (e.g., world states, objects, and action sequences), making them more aligned with the goals pursued. These manipulations allow the conscious agent to internally produce the knowledge it lacks to cope with novel conditions and goals, increasing the flexibility of goal-directed behaviour. The manipulation of representations is supported by four neuro-functional macro-systems (hierarchical perceptual working memories, abstract working memory, internal manipulator, motivational systems) that operate through a set of computational manipulation operations (abstraction, specification, decomposition, composition). The theory also presents the concept of `GARIM agency', proposing that subjective conscious experience derives from the ability of agents to generate and control a vivid internally simulated reality. Furthermore, the theory highlights the criticalities of the experimental investigation of consciousness, suggesting a new approach to testing consciousness in biological and artificial agents. Finally, the GARIM theory can benefit technological fields such as machine learning and autonomous robotics (e.g., the manipulation processes proposed by the theory could be linked to the operations performed by systems based on transformers).

翻译：意识作为人类认知的核心要素，已通过神经科学、心理学、人工智能和机器人学等多个科学方法展开研究。然而，学科间整合不足限制了我们对意识的完整清晰认知。本文通过提出神经计算框架下的“目标对齐表征内部操控”（GARIM）意识理论，致力于改善这种整合。GARIM理论的核心观点在于：意识支持对目标相关内部表征（如世界状态、物体及行动序列）的主动操控，使其更契合所追求的目标。这些操控使意识主体能够内部生成应对新情境和新目标所需的知识，从而提升目标导向行为的灵活性。表征的操控由四个神经功能宏观系统（层级感知工作记忆、抽象工作记忆、内部操控器、动机系统）支撑，通过一组计算操控操作（抽象化、特化、分解、组合）实现。该理论还提出了“GARIM主体性”概念，指出主观意识体验源于主体生成并控制生动内部模拟现实的能力。此外，理论揭示了意识实验研究的关键难点，提出了在生物与人工主体中测试意识的新方法。最终，GARIM理论可为机器学习与自主机器人等科技领域（例如，该理论提出的操控过程可与基于Transformer的系统所执行的操作建立关联）提供启示。