Vision-Language-Action (VLA) models map visual observations and language instructions to low-level control actions, but their reliance on the current observation often limits temporal continuity in long-horizon manipulation. We propose ProRet-VLA, a temporal reasoning framework that links prospection and retrospection through predictive consistency. ProRet-VLA maintains compact episodic memories and predicts future observations from temporally enriched context. Rather than using future prediction only as an auxiliary target, we use its prediction error as a self-supervised deviation signal to train a lightweight Deviation Predictor. During inference, the predicted deviation conditions token-wise fusion between current context and retrieved memory, allowing memory usage to vary across decision stages without accessing future observations. Experiments on CALVIN ABC-D and LIBERO show that ProRet-VLA improves long-horizon task completion, achieving an average length of 4.50 on CALVIN and a 96.7% average success rate on LIBERO. Real-world experiments on two long-horizon manipulation tasks further indicate improved temporal progress tracking over strong VLA baselines.

Compared with relying solely on initial observations and language instructions, predicting goal images with generative models as high-level visual guidance can significantly enhance the robustness of Vision-Language-Action (VLA) models. However, most existing foundation models have not systematically incorporated goal image conditioning due to the high computational training cost. To this end, we propose Goal-Injected Fine-Tuning (GIFT), a lightweight and efficient fine-tuning framework that seamlessly integrates generated goal images into multiple representative pretrained VLA models. Our approach introduces goal image features into observations via a zero-initialized convolution which progressively grows parameters from zero and prevents harmful noise from disrupting the pretrained policy during fine-tuning. As training proceeds, goal information is gradually incorporated, enabling efficient goal understanding without disrupting model stability. We further introduce a refined image editing method to generate semantically and visually consistent goal images from initial observations and task instructions. Experiments show that goal-aware VLA models achieve substantial performance gains across tasks: with only a single epoch of fine-tuning, GIFT outperforms the base model by 6.0% and 13.4% on two SIMPLER settings, and by 4.7% on LIBERO, demonstrating both efficiency and effectiveness.