X‐Gen: Synergizing Extended LSTM and Generative Diffusion for Cognition‐Aware Motion Forecasting

Motion forecasting in autonomous driving has historically relied on discriminative baselines that prioritize statistical correlation over causal reasoning, often failing to capture the complex social dynamics of real‐world traffic. To bridge this cognitive gap, we introduce X‐Gen, a novel generative framework that synergizes the inferential depth of foundation models with the probabilistic precision of diffusion models. Unlike prior approaches constrained by heuristic textual prompts, X‐Gen establishes a direct semantic alignment between continuous kinematic states and the high‐dimensional embedding space of a pretrained Llama‐3‐8B backbone via quantized low‐rank adaptation (Q‐LoRA). This architecture allows the system to internalize scene semantics as a coherent “World Model.” Furthermore, to resolve long‐range topological dependencies, we propose a lane‐aware cognitive gating mechanism powered by the extended LSTM (xLSTM). Leveraging xLSTM’s matrix memory and exponential gating, this module selectively filters spatial constraints with linear computational complexity, effectively pruning the search space for navigational intent. Finally, we eschew restrictive parametric distributions in favor of a conditional diffusion decoder, which formulates trajectory synthesis as an iterative denoising process guided by the aligned cognitive features. Empirical validation on the large‐scale Waymo open motion dataset demonstrates that X‐Gen establishes new state‐of‐the‐art benchmarks in both accuracy and diversity metrics.