Time Travel Engine: A Shared Latent Chronological Manifold Enables Historical Navigation in Large Language Models

TL;DR

The paper identifies a continuous latent temporal geometry in LLMs and introduces the Time Travel Engine (TTE) to map, traverse, and control this chronology. By combining static era anchors with dynamic manifold trajectories, TTE enables fluid, era-consistent linguistic and cognitive shifts while preserving core reasoning and restricting future knowledge. The authors demonstrate cross-lingual topological universality between Chinese and English temporal subspaces and show that temporal—and not merely stylistic—controls can be disentangled and robust under cross-language transfer. This work bridges historical linguistics and mechanistic interpretability, offering a principled framework for diachronic navigation, epistemic boundary enforcement, and cross-lingual temporal transfer in large language models.

Abstract

Time functions as a fundamental dimension of human cognition, yet the mechanisms by which Large Language Models (LLMs) encode chronological progression remain opaque. We demonstrate that temporal information in their latent space is organized not as discrete clusters but as a continuous, traversable geometry. We introduce the Time Travel Engine (TTE), an interpretability-driven framework that projects diachronic linguistic patterns onto a shared chronological manifold. Unlike surface-level prompting, TTE directly modulates latent representations to induce coherent stylistic, lexical, and conceptual shifts aligned with target eras. By parameterizing diachronic evolution as a continuous manifold within the residual stream, TTE enables fluid navigation through period-specific "zeitgeists" while restricting access to future knowledge. Furthermore, experiments across diverse architectures reveal topological isomorphism between the temporal subspaces of Chinese and English-indicating that distinct languages share a universal geometric logic of historical evolution. These findings bridge historical linguistics with mechanistic interpretability, offering a novel paradigm for controlling temporal reasoning in neural networks.

Figures (5)

• Figure 1: The Time Travel Engine (TTE). This framework navigates a traversable latent chronological manifold within the model. By integrating manifold-derived temporal control signals $\mathbf{z}(t)$ into the residual stream, the mechanism dynamically modulates cognitive states and linguistic registers. This approach facilitates fluid navigation to specific historical eras (e.g., Old, Middle, Early Modern [E. Mod]) and highlights a topological isomorphism between Chinese and English temporal subspaces.
• Figure 2: Isomap projection of the latent chronological manifold extracted from Qwen2.5-14B-Base. Each data point represents the activation centroid of a self-generated CAA sample, color-coded by era. The curves illustrate the continuous geometric progression from Old to Modern periods across both Chinese and English subspaces.
• Figure 3: Perplexity matrices for temporal signal modulation. Rows indicate the steered temporal signal; columns indicate the test corpus era.
• Figure 4: Epistemic detection visualizing knowledge boundary integrity. The unsteered Baseline (Modern State) shows minimal leakage.
• Figure 5: Visualization of the entanglement vector fields. The arrows depict the shift in performance metrics when orthogonalizing the time vector to remove stylistic features ($\mathbf{v}_{time} \to \mathbf{v}_{cog}$).