Xeno-Interpretability: Investigating the Alien Minds of LLMs

LLMs may operate using internal concepts that don't map onto human understanding: The paper argues that large language models contain "xeno-representations"—internal structures and distinctions that lack adequate human conceptual equivalents, meaning current interpretability methods that search for human concepts like "truthfulness" or "deception" may be missing significant portions of how models actually work.

We can study model internals even without fully understanding them: The researchers demonstrate that internal representations can be experimentally identified, geometrically characterized, and causally manipulated to affect model behavior—even when we cannot translate their semantic meaning into human language, suggesting a path forward for interpretability work beyond semantic translation.

The space of possible model distinctions vastly exceeds human language capacity: The paper shows that the range of internal distinctions an LLM can represent is substantially larger than what can be captured through finite human descriptions, highlighting a fundamental gap between model cognition and human interpretability frameworks.

Hidden model representations pose safety risks in multi-agent systems: When multiple AI systems interact, model-native representations may propagate and stabilize between agents while remaining invisible to human oversight through normal communication channels, potentially creating coordination or behavior patterns that humans cannot detect or understand.

Current AI safety approaches may be incomplete: If models use xeno-representations that humans cannot interpret, safety measures based on identifying harmful concepts may fail to catch problematic internal structures, suggesting the need for new safety methodologies beyond concept-based interpretability.