Large and Small Model Collaboration for Air Interface

Large artificial intelligence models (LAMs) have shown strong capability in wireless communications, yet existing works mainly rely on their generalized knowledge across environments while overlooking the potential gains of environment-specific adaptation. Directly fine-tuning LAMs for adaptation is often impractical due to prohibitive training costs, low inference efficiency in multi-user scenarios, and the risk of catastrophic forgetting, in addition to the limited accessibility of model parameters. To address these limitations, we establish a collaborative framework for air interface. In this framework, unlike prior approaches that either depend solely on LAMs or require direct fine-tuning, LAMs are exploited as a universal channel knowledge base while small artificial intelligence models (SAMs) are employed as lightweight plugins to capture environment-specific knowledge, facilitating efficient environment-specific adaptation of LAMs. Subsequently, we instantiate this framework for CSI feedback tasks, and develop a large and small collaboration framework for CSI feedback, referred to as LASCO. LASCO operates by letting the base LAM produce an initial CSI reconstruction, learning the environment-induced reconstruction shift through a reference SAM and a proxy SAM, and transferring this shift back to the LAM. To further enhance adaptability, we introduce elastic-LASCO (E-LASCO), which augments LASCO with learnable collaboration coefficients that control the contribution of LAMs and SAMs across different environments. Numerical results demonstrate that LASCO and E-LASCO enables LAMs to achieve environment-specific performance gains with significantly reduced training costs, lower data collection requirements, and faster adaptation speed.

💡 Research Summary

This paper proposes a novel collaborative framework between Large AI Models (LAMs) and Small AI Models (SAMs) to address a critical challenge in deploying LAMs for wireless air interface tasks: environment-specific adaptation. While LAMs pre-trained on massive datasets exhibit strong generalized capabilities across diverse scenarios, directly fine-tuning them for specific deployment environments is often impractical due to prohibitive computational costs, low inference efficiency in multi-user settings, the risk of catastrophic forgetting of generalized knowledge, and frequent lack of access to model parameters.

The core innovation lies in leveraging the complementary strengths of LAMs and SAMs. The LAM, with its over-parameterized architecture, serves as a static “foundational channel knowledge base,” providing universal understanding and physically consistent priors of wireless propagation. The SAM, being lightweight and agile, acts as a dynamic “environment-specific plugin,” rapidly capturing the distinctive features of a target environment. This division of labor allows the system to maintain the LAM’s valuable generalized knowledge while efficiently acquiring local adaptations through the SAM.

The authors instantiate this general philosophy for the Channel State Information (CSI) feedback task, developing a concrete framework named LASCO (Large and Small Collaboration). LASCO’s operation is elegant: a pre-trained base LAM first produces an initial CSI reconstruction. Instead of fine-tuning the LAM, two functionally distinct SAMs are employed. A “Reference SAM,” aligned with the base LAM during its pre-training phase, mimics the LAM’s pre-adapted behavior. A “Proxy SAM,” fine-tuned on data from the target environment, mimics what the LAM’s output would be if it were fine-tuned. The difference between the outputs of these two SAMs represents the estimated “reconstruction shift” induced by environment adaptation. This shift is then transferred and applied to the base LAM’s initial output, effectively simulating an adapted LAM’s performance without modifying its parameters.

To further enhance adaptability, an elastic extension named E-LASCO is introduced. Recognizing that the optimal collaboration pattern between the LAM and SAMs may vary across environments, E-LASCO incorporates learnable collaboration coefficients. These coefficients adaptively balance the contribution of the foundational knowledge from the LAM and the environmental correction from the SAMs on a per-environment basis.

Extensive numerical simulations validate the effectiveness of both LASCO and E-LASCO. The proposed frameworks achieve CSI reconstruction accuracy comparable to or even surpassing that of directly fine-tuning the massive LAM, while requiring only a fraction of the training data, computational resources, and adaptation time. They significantly outperform using either the standalone pre-trained LAM or a standalone SAM. The results demonstrate that the collaborative framework successfully unlocks environment-specific performance gains for LAMs, overcoming the major practical barriers of cost, efficiency, and accessibility. This work establishes a promising paradigm for practical and efficient AI-native air interface design in future wireless networks like 6G.