New QO-ISC semantic coding framework beats codecs with large AI models

Semantic communication promises to preserve meaning rather than raw bits, but current designs ignore user intent, rely on dataset-specific fine-tuning, and don't scale to large MIMO-OFDM systems. Sin-Yu Huang and Vincent W.S. Wong from the University of British Columbia address this with a generalized query-oriented image semantic coding (QO-ISC) framework. The transmitter extracts only features relevant to the user's query using a pretrained large AI model (LAM), and the receiver reconstructs the image from that semantic subset. This moves away from one-size-fits-all encoding and toward intent-driven transmission.

To make it work in real wireless systems, the authors also develop semantic-aware hybrid beamforming (SA-HBF), which prioritizes semantically important subcarriers and antennas in large-scale MIMO-OFDM. When tested on unseen object categories, QO-ISC outperforms traditional codecs and two state-of-the-art semantic coding schemes. The paper, accepted by IEEE Transactions on Communications and available on arXiv (2607.28276), shows that combining large AI models with semantic-aware physical-layer techniques can slash overhead while improving reconstruction accuracy for task-oriented users.