Large Language Models (LLMs) achieve strong reasoning performance by generating intermediate reasoning trajectories through autoregressive decoding. Existing approaches such as self-consistency, tree-based reasoning, and critical token analysis improve reasoning by exploring diverse trajectories or identifying influential tokens. In this work, we identify a distinct phenomenon in LLM reasoning: **critical positions** — locations where re-sampling continuations disproportionately influences the final outcome, independent of the specific token at that position. Motivated by this, we propose **Metacognitive Influence Tracing (MIT)**, the first unsupervised framework for identifying critical positions from a single reasoning trace using only attention structure. MIT models reasoning trajectories as influence graphs and estimates propagated influence through a diffusion process, enabling the discovery of structurally important reasoning positions without supervision or contrastive trajectories. Building on these insights, we further propose **Adaptive Critical Sampling (ACS)**, an intervention framework that selectively re-samples at critical positions instead of regenerating entire reasoning traces. Across six benchmark tasks and five models, ACS improves accuracy by an average of **7.78 points** over CoT while reducing computational cost by **59.83%** over Self-Consistency.