Multi-Channel Conscious Binding (MCCB): Dynamic Concept Steering in Frozen Large Language Models

This record presents Multi-Channel Conscious Binding (MCCB), a lightweight and interpretable framework for continuous and compositional concept steering in frozen large language models. MCCB formulates internal model behavior in terms of multiple concept channels, represented as directional vectors in hidden-state space. During inference, these channels may be activated simultaneously and non-exclusively, with their strengths regulated by a calibrated gating mechanism based on contextual alignment. This design enables dynamic modulation of internal representational dynamics without modifying model parameters. The primary contribution of MCCB is the abstraction of concept binding as a controllable, continuous process rather than a discrete mode selection. This perspective supports both empirical investigation and theoretical analysis of internal control mechanisms in large neural networks. Depending on the version of this record, the archive may include: A conceptual and mathematical description of the MCCB framework Empirical evaluations using frozen transformer-based language models Reproducible experimental implementations, configurations, and generated results Where applicable, related records provide explicit links between conceptual descriptions and reference implementations using persistent identifiers. This separation is intentional, allowing future work to extend or reinterpret either the theoretical formulation or the experimental realization independently. MCCB is intended as a research foundation for further studies on controllable generation, interpretability, and compositional internal mechanisms in large language models.