Zero-shot Object Goal Navigation requires an embodied agent to locate a target object in an unseen environment, where the choice of which frontier to explore next is critical to both success and efficiency. Existing methods score frontiers from already-observed cues such as nearby objects or scene graphs, leaving the structure beyond each frontier implicit and offering no principled prior for choosing among rooms the agent has yet to see. In contrast, the hippocampal entorhinal system supports navigation through a predictive cognitive map that anticipates unseen space from partial observations. Inspired by this, we propose PredNav, which externalizes a hierarchical cognitive map with metric, instance-semantic, and topological layers, and casts frontier selection as room-level structured prediction over its topological layer. A language model infers the likely room type and target relevance behind each frontier, which is fused with CLIP visual evidence, an offline room category prior, and geodesic cost under phase-adaptive weights. A vision-language verifier then gates stop using direct visual evidence. Extensive experiments on MP3D and HM3D demonstrate the effectiveness of our method.

Tokenizers play a critical role in large language model studies. Despite recent advances, existing tokenizers fail to explicitly leverage historical tokenization results when making subsequent token decisions, nor do they selectively utilize such history based on contextual relevance. We propose SPEAK, a tokenizer that integrates spiking neurons to explicitly leverage historical tokenization results. Furthermore, we introduce an entropy-aware reset mechanism that selectively leverages history based on contextual relevance, which is determined by token-level entropy. High-entropy tokens are treated as contextual boundaries, whereas low-entropy tokens between consecutive such boundaries exhibit strong contextual relevance. Accordingly, we induce hard reset at high-entropy tokens to discard irrelevant historical tokenization results, and soft reset at low-entropy tokens to preserve and leverage relevant history. Experiments on 2 language models and 5 datasets spanning 16 languages demonstrate superior cross-lingual adaptability, with competitive performance and efficiency. Our code is publicly available at https://github.com/zju-bmi-lab/SPEAK.