The Granularity Axis: A Micro-to-Macro Latent Direction for Social Roles in Language Models
:::info Stub — Full Engineering Breakdown Coming This paper was featured on Hugging Face Daily Papers on 2026-05-07 with 5 upvotes. A full breakdown with production viability rating, implementation notes, and honest limitations is being written. Subscribe to AI Letters → :::
| Authors | Chonghan Qin et al. |
| Year | 2026 |
| HF Upvotes | 5 |
| arXiv | 2605.06196 |
| Download | |
| HF Page | View on Hugging Face |
Abstract
Large language models (LLMs) are routinely prompted to take on social roles ranging from individuals to institutions, yet it remains unclear whether their internal representations encode the granularity of such roles, from micro-level individual experience to macro-level organizational, institutional, or national reasoning. We show that they do. We define a contrast-based Granularity Axis as the difference between mean macro- and micro-role hidden states. In Qwen3-8B, this axis aligns with the principal axis (PC1) of the role representation space at cosine 0.972 and accounts for 52.6% of its variance, indicating that granularity is the dominant geometric axis organizing prompted social roles. We construct 75 social roles across five granularity levels and collect 91,200 role-conditioned responses over shared questions and prompt variants, then extract role-level hidden states and project them onto the axis. Role projections increase monotonically across all five levels, remain stable across layers, prompt variants, endpoint definitions, held-out splits, and score-filtered subsets, and transfer to Llama-3.1-8B-Instruct. The axis is also causally relevant: activation steering along it shifts response granularity in the predicted direction, with Llama moving from 2.00 to 3.17 on a five-point macro scale under positive steering on prompts that admit local responses. The two models differ in controllability, suggesting that steering depends on each model's default operating regime. Overall, our findings suggest that social role granularity is not merely a stylistic surface feature, but a structured, ordered, and causally manipulable latent direction in role-conditioned language model behavior.
Engineering Breakdown
Plain English
This paper discovers that LLMs have a geometric structure in their internal representations that encodes the 'granularity' of social roles—whether they're reasoning as an individual person or as a large organization. The researchers found that in Qwen3-8B, a single direction in the model's hidden state space (the Granularity Axis) explains 52.6% of variance across 75 different social roles spanning from micro-level individuals to macro-level institutions, and this axis aligns nearly perfectly (0.972 cosine similarity) with the model's principal component of role representations.
Key Engineering Insight
LLM role representations aren't scattered randomly in latent space—they're organized around interpretable geometric axes that directly correspond to semantic properties like granularity. This means you can predict and potentially control how a model reasons about a role simply by identifying where that role sits on this axis, rather than treating role behavior as a black box.
Why It Matters for Engineers
If you're building systems that prompt LLMs to reason from specific perspectives (customer service agents, policy makers, individual users), this reveals that model behavior is structured and measurable in a way that wasn't previously documented. You can now diagnose whether a model is reasoning at the right 'scale' for your use case, and potentially manipulate this axis to steer model outputs without retraining—useful for applications requiring consistent perspective-taking.
Research Context
Prior work has shown LLMs can adopt social roles when prompted, but it was unclear whether the model's internal representations actually encode role properties or if role-switching was just surface-level behavior. This paper advances interpretability research by proving that semantic properties of roles map to stable geometric directions in activation space, similar to how prior work found interpretable axes for concepts like bias or sentiment, but here applied systematically to the previously unmapped domain of social role granularity.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::