Beyond YOLO: Dynamic Re-Reading and Task-Aware Context Re-Initialization in Large Language Models

The prevailing paradigm of Large Language Models (LLMs) operates on a "You Only Look Once" (YOLO) principle: a single, unidirectional forward pass where historical tokens are compressed into hidden states and KV caches without a priori knowledge of the downstream generation task. While efficient, this one-shot prefilling yields a generalized, average-state representation that inherently struggles with highly specific, distant information retrieval. In this paper, we categorize generation tasks into two distinct classes based on context dependency: Macro-Semantic Abstraction (MSA) and Granular Distant-Context Retrieval (GDCR). We demonstrate that for GDCR tasks, standard attention mechanisms fail because early hidden states are generated blind to future queries. To resolve this, we propose a novel architectural mechanism: Dynamic Query-Aware Re-Prefilling. By introducing an auxiliary classifier gate alongside the standard language modeling head, our model learns to dynamically evaluate whether the local context window is sufficient for the current prediction. If insufficiency is detected, the model triggers a "re-read" action, prepending the recent query tokens to the absolute beginning of the context and re-initializing the KV cache. This mechanism transforms linear models from static, single-pass compressors into dynamic, task-aware readers, paving the way for significantly higher accuracy in extreme long-context reasoning and retrieval.