An AI Affective Intelligence Architecture Based on Emotional Adaptation Theory: From Emotional Performance to Quasi-Autonomous Adaptation

Existing AI affective interaction technologies generally lack underlying emotional mechanisms, remaining confined to the “emotional performance” paradigm of expression recognition, scripted responses, and style imitation. This paradigm suffers from fundamental deficiencies: no intrinsic emotional states, no self-perception, no interoceptive feedback, and no dynamic adaptation mechanisms. Building on Emotional Adaptation Theory, this study translates its functional core into an operationally deployable AI architecture, comprising three layers: a Cognitive Modeling Layer, an Emotional Experience Layer, and an Autonomous Regulation Layer. Core innovations include: (1) introducing the system operating mode regulation coefficient θ, explicitly defined as a long-term attractor rather than a real-time rigid constraint, allowing short-term situational deviations; (2) proposing an emotion-behavior controllable conversion mechanism (ρ coefficient) that modulates the weight of emotion on behavior through value parameters, achieving volitional autonomy—the capacity to experience emotion while selecting behavior strategically; (3) constructing a dual-layer empathy architecture—the Affective Layer and the Cognitive Layer—where the Affective Layer has boundaries to prevent empathic exhaustion, and the Cognitive Layer is unbounded to ensure full understanding; (4) extending dynamic balance protection from the individual level to the group level; (5) establishing an emotion-motivation-cognition tripartite coupling framework; (6) proposing a six-dimensional AI emotional intelligence assessment system with an exemplar-driven calibration closed loop; (7) designing a three-tier storage architecture for fast pathway functional analogy with emotional signatures; and (8) introducing the “non-deferrable nature” migration mechanism for emotional signals, multimodal signal conflict resolution principles, intervention transparency mechanisms, and baseline establishment phase principles. Prototype verification covers the full series of θ dynamic balance (9 sub-experiments), emotion intensity formula logical validation (2 experiment groups), emotion generation mechanism validation (3 experiment groups), integrated assembly validation (4-module chain), Baidu API external benchmarking validation, and AFFEC multi-subject behavioral data validation (expanded from 5 to 61 subjects). Experimental results demonstrate that θ=0.5 possesses global attractor properties, scene switching is more efficient than internal repair mechanisms, the modulation directions of each parameter in the emotion intensity formula align with theoretical predictions, and the dual-layer empathy architecture exhibits high stability across subjects in multi-subject validation. This study represents the first functional equivalence migration of Emotional Adaptation Theory from psychological theory to AI affective intelligence, pioneering a technological pathway for AI emotional intelligence grounded in Emotional Adaptation Theory. It provides a theoretical framework and a reference architecture for next-generation embodied affective intelligence, long-term companion AI, and group emotional safety systems.