Mariana Meireles

The Biological Arrow of Time (Prokopenko et al., 2024) framework proposes that biological complexity increases through the emergence of tangled information hierarchies whose resolution drives evolutionary transitions. This process of sophistication ultimately gave rise to the architecture of the human brain. Central to such elaborate structure is the concept of synergy, that can be defined as information that only becomes available when multiple sources are considered jointly, revealing the integrated nature of the whole. In the human brain, a synergistic core in association cortices supports higher cognition and is a distinguishing factor relative to other primates (Luppi et al., 2022). We argue that an analogous process of sophistication is occurring in LLMs, where synergistic cores emerge through training and their ablation causes disproportionate damage to these systems' complex reasoning capabilities (Urbina-Rodriguez et al., 2026). In this work, we use emergent misalignment (Betley et al., 2025), where narrow fine-tuning induces broadly misaligned behaviour, as a probe into this synergistic organisation, investigating how the propagation of misalignment across seemingly unrelated domains reveals the underlying synergistic entanglement of concepts.