Large-scale language models (LLMs) excel at understanding and generating natural language, but their vulnerability to factual errors limits their reliability in knowledge-intensive tasks. While decode-time strategies offer an efficient solution without training, existing methods process token- and layer-level signals separately, overlooking their joint dynamics. In this study, we present a token-aware, layer-localized contrastive decoding method that improves factual generation by aligning specific token types with their most influential transformer layers. Empirical attention analysis identifies two key patterns: punctuation tokens receive dominant attention in early layers, while conceptual tokens dominate semantic inference in intermediate layers. By selectively suppressing attention to these token types at this depth, we achieve controlled factual degradation and derive contrastive signals that guide final factual decoding. Our method requires no additional training or model modification, and we demonstrate through experiments that it consistently improves factuality across multiple LLMs and various benchmarks.
