Researchers say some of what's happening inside large language models may be untranslatable into human concepts.
A new arXiv paper introduces the idea of "xeno-representations": internal distinctions an LLM makes that don't map onto any existing human concept, unlike familiar categories such as truthfulness, refusal, or deception. The authors argue the space of possible internal distinctions inside a model is far larger than what our finite vocabulary can describe. They separate two different jobs: locating and manipulating a representation experimentally, versus explaining what it actually means in words we understand, and say the first can succeed even when the second can't. The paper sketches a research program for identifying these structures and studying how they behave.
This reframes AI interpretability: instead of decoding a humanlike mind, researchers may be studying something closer to an alien signal. It also raises a safety concern the paper flags directly: in multi-agent AI systems, these model-native structures could spread between interacting agents while staying largely invisible in the human-readable messages those agents exchange.
The paper offers a research program, not a finding: no xeno-representation has actually been pinned down yet, so the alien-mind framing is a hypothesis worth testing, not a discovery.