CFM-Bench: A Unified Multi-Domain, Multi-Task Benchmark for Channel Foundation Models

Channel foundation models (CFMs) aim to learn reusable representations of wireless channels to improve performance across various downstream tasks. However, the field currently lacks a standardized evaluation protocol. Most existing studies evaluate models on isolated, custom-built pipelines with varying data sources, partitioning schemes, and metrics. This fragmentation makes it impossible to fairly compare different models or determine whether general-purpose representations actually outperform task-specific architectures.

CFM-Bench addresses this by providing a unified evaluation framework that spans six distinct channel configurations, including 3GPP statistical simulations, ray-tracing environments, industrial and aerial measurements, and synchronized vehicular multimodal data. The benchmark enforces a strict 'test-isolation' policy: official test units are held out from all model development stages, and researchers must disclose all data used during pretraining and fine-tuning. By defining fixed, leakage-resistant partitions—such as isolating complete trajectories or measurement sessions—the benchmark ensures that performance gains reflect genuine generalization rather than overfitting to spatially or temporally correlated data.

The benchmark organizes tasks into three application dimensions: physical-layer (PHY) channel intelligence, radio-access-network (RAN) decision intelligence, and integrated sensing and communication (ISAC). These include tasks such as CSI feedback, frequency and temporal channel extrapolation, beam prediction, and localization. By providing a common substrate for these tasks, CFM-Bench enables researchers to evaluate the transferability of channel representations across different domains and radio configurations without forcing a single model architecture or input format.