Coronal mass ejections (CMEs) pose a significant threat to critical infrastructure, making their accurate forecasting extremely important. While predicting CMEs is only the initial stage in forecasting their terrestrial impact, it offers valuable time for mitigation efforts. However, CME forecasting remains a persistent challenge. This study investigates the potential implications of coupling CME forecasts with flare predictions, a common approach in existing models. We employ a transformer-based architecture to develop two models: one for CME forecasting independent of flare occurrence (CME model) and another for predicting whether a >M class flare will be associated with a CME (flare-CME model), both using sequences of SHARP keywords and the same architecture. Consistent with previous research, neither model achieves a good performance when closely examined. However, the flare-CME model outperforms the CME model when evaluated on time periods containing >M flares, suggesting that each model relies on distinct precursors. We hypothesize that the CME model may focus on “easier” signatures, which become saturated during periods preceding >M flares, while the flare-CME model detects more subtle patterns.