Reproduce “Token Communication in the Era of Large Models: An Information Bottleneck-Based Approach” on OpenResearch

About this paper

Hao Wei, Wanli Ni, Wen Wang, Wenjun Xu, Dusit Niyato, Ping Zhang

Researchers propose UniToCom, a Unified Token Communication paradigm, which redefines communication units as tokens to align with generative AI models, leveraging a Generative Information Bottleneck (GenIB) for efficient tokenization and a Multimodal Large Language Model (MLLM) for unified multimodal understanding and generation. The framework consistently outperforms traditional and existing semantic communication methods across various tasks and channel conditions, demonstrating enhanced communication and computational efficiency through optimal token compression.

Key insights

• Aligning communication systems with the token-based intelligence of generative AI models offers a more natural, interpretable, and unified approach to multimodal processing than prior semantic communication methods.
• The σ-GenIB principle effectively addresses the variance collapse issue in autoregressive models, enabling robust and diverse latent representations for high-quality multimodal content generation.
• Moderate token compression significantly improves both communication efficiency (reduced bandwidth) and computational efficiency (faster MLLM processing and convergence) without degrading overall system performance.

Summary by alphaXiv.