Large language models compress centuries of discourse into fixed parameters, yet how they organize historical variation and whether this organization shapes generation remain unclear. We introduce the Time Travel Engine (TTE), a framework that compares geometric accounts of era-conditioned residual states and tests their behavioral accessibility under fixed prompts. Across six decoder language models in Chinese and English, broad eras repeatedly form ordered trajectories, although their exact spacing varies across model–language settings. Steering along these trajectories shifts both historical register and access to period-compatible knowledge, while adjacent anchors support local interpolation. Epistemic effects partly survive style removal, linking temporal steering to knowledge access beyond surface register. Across Chinese and English, partially alignable trajectories reveal corresponding chronological relations whose geometric realization remains language-specific. Together, ordered temporal geometry offers a testable account of how static language models organize diachronic variation, linking internal structure to historical register and access to period-compatible knowledge. Code and datasets: https://github.com/LilTiger/Time-Travel-Engine.

Language comprehension relies on integrating information across both local words and broader context. We propose a method to quantify the information integration window of large language models (LLMs) and examine how sentence and clause boundaries constrain this window. Specifically, LLMs are required to predict a target word based on either a local window (local prediction) or the full context (global prediction), and we use Jensen-Shannon (JS) divergence to measure the information loss from relying solely on the local window, termed the local-prediction deficit. Results show that integration windows of both humans and LLMs are strongly modulated by sentence boundaries, and predictions primarily rely on words within the same sentence or clause: The local-prediction deficit follows a power-law decay as the window length increases and drops sharply at the sentence boundary. This boundary effect is primarily attributed to linguistic structural markers, e.g., punctuation, rather than implicit syntactic or semantic cues. Together, these results indicate that LLMs rely on explicit structural cues to guide their information integration strategy.