Minimal Computational Preconditions for Subjective Perspective in Artificial Agents

TL;DR

The paper addresses how to instantiate subjective perspective in artificial agents by grounding it in a minimal, phenomenologically motivated internal structure. It introduces a slow global latent $g_t$ that globally biases fast policy without being directly optimized for reward, tested in a reward-free regime-switch environment where $g_t$ exhibits direction-dependent hysteresis while policy behavior remains reactive. The approach combines a three-part architecture, a prediction-error learning objective with a slow-perspective regularizer, and a measurement protocol using $g$-score and switch trajectories to diagnose perspective-like structure. The results argue that hysteresis and temporal persistence in the slow latent constitute a computational signature of perspective-like subjectivity in machines, with discussion of links to latent world models and potential extensions to language-based agents for broader applicability.

Abstract

This study operationalizes subjective perspective in artificial agents by grounding it in a minimal, phenomenologically motivated internal structure. The perspective is implemented as a slowly evolving global latent state that modulates fast policy dynamics without being directly optimized for behavioral consequences. In a reward-free environment with regime shifts, this latent structure exhibits direction-dependent hysteresis, while policy-level behavior remains comparatively reactive. I argue that such hysteresis constitutes a measurable signature of perspective-like subjectivity in machine systems.

Figures (4)

• Figure 1: Overview of the proposed agent architecture. Note that the agent is endowed with an internal latent structure that mediates information processing. Among these latent variables, the global latent $g$ is designed to evolve gradually and to encode habitual, slowly varying regularities, thereby instantiating the agent's perspective. The action policy $\pi(a_t | z_t, g_t)$ is jointly conditioned on the fast-changing perceptual state $z_t$ and the slowly evolving perspective $g_t$, resulting in behavior that is both reactive to immediate inputs and biased by longer-term experiential structure.
• Figure 2: Visualization of the grid-world environment used in the experiments. The environment is rendered using a Pygame-based simulator and consists of three spatial zones with distinct observation noise, indicated by different background colors: red for $Z_0$, green for $Z_1$, and blue for $Z_2$. The white circle denotes the agent's current position. The overlay at the top displays diagnostic quantities during the simulation: the first row shows the total timestep, episode index, timestep in the current episode, zone label, agent's position, and selected action (5 actions available: move left / right / up / down / stay). The second row shows the total training loss, prediction loss (one-step world-model MSE), perspective smoothness regularization, policy entropy, and the L2 norm of the global latent state. No reward signal is present in the environment.
• Figure 3: Zone occupancy before and after training. Zone occupancy is measured as the mean fraction of timesteps spent in each zone, averaged across five random seeds, and compared between early training episodes (episodes 0-20) and late training episodes (episodes 180-200). Error bars indicate standard deviation across seeds. After training, the agent reliably concentrates its behavior in the most predictable zone $Z_2$.
• Figure 4: Switch-aligned IQR hysteresis trajectories. Median switch-aligned trajectories for the signed projection score of the global latent state $g_t$ (left) and policy entropy (right), following Regime A$\rightarrow$B and Regime B$\rightarrow$A transitions. The shaded area depicts the Inter-Quartile Range across seeds. The global latent score exhibits clear directional hysteresis and smooth, history-dependent adaptation, whereas policy entropy shows high variability and minimal sensitivity to transition direction.