LLMs encode social role granularity along dominant geometric axis

A recent study reveals that large language models (LLMs) represent the complexity of social roles—from individual to institutional—along a significant geometric axis. Researchers introduced a Granularity Axis based on the contrast between mean hidden states of macro- and micro-roles. In Qwen3-8B, this axis corresponds to the principal component (PC1) of the role representation space with a cosine similarity of 0.972, explaining 52.6% of the variance. The research team identified 75 social roles across five levels of granularity and gathered 91,200 responses conditioned on these roles through various prompts. The findings indicate that role projections increase steadily along the axis, underscoring granularity as the key organizing factor for prompted social roles in LLMs. The paper can be found on arXiv with identifier 2605.06196.