This paper proposes a context-in-context learning (ICL) theory using an LLM-based transformer to address the low throughput problem in dynamic channel environments with the binary exponential backoff scheme used in WiFi 7. This approach overcomes the limitations of existing model-based approaches that assume a fixed node density. It trains the transformer using collision threshold data examples and query collision cases as inputs to generate predicted contention window thresholds (CWTs). An efficient algorithm is developed to ensure optimal CWT prediction within a limited training time. Considering the difficulty of obtaining perfect data in real-world environments, we present an extended model that allows for erroneous data inputs. NS-3 simulation results demonstrate that our approach achieves faster convergence and near-optimal throughput than existing model-based and DRL-based approaches.
