acceptodds
Withdrawn from ICLR 2027

The Granularity Axis: A Micro-to-Macro Latent Direction for Social Roles in Language Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Large language models (LLMs) are prompted to adopt social roles from individuals to institutions, but whether their representations encode this difference in social granularity remains unclear. We discover a contrast-based Granularity Axis, the difference between mean macro- and micro-role hidden states, that aligns with the dominant geometry of role space: in Qwen3-8B, its cosine with PC1 is , and PC1 explains of the variance. Across 75 roles at five levels, with 90,000 role-conditioned and 1,200 default responses per model, projections increase monotonically and remain stable across layers, prompts, and response-quality filters. The structure transfers to Llama-3.1-8B-Instruct and scales to Qwen3-32B and Llama-3.1-70B-Instruct, whose PC1 shares are and . Independent five-rater validation () preserves the ordering; identical continuations and affect/formality controls confirm scale-related variation beyond generated wording. The axis is also behaviorally causal: positive steering moves Llama from to on a five-point macro scale for prompts admitting local responses. Steering strength and stability depend on the model's default operating regime. Together, these findings establish social granularity as a representational primitive: a continuous, ordered, causally manipulable direction organizing role-conditioned generation across model families. Code and data are available at https://anonymous.4open.science/r/Granularity-Axis-anon/.

Primary area:
interpretability and explainable AI
Keywords:
large language models, representation engineering, activation steering, social simulation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_7qxJOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper find evidence for causal representations of 'granularity' in the residual stream of LLMs. The paper finds both geometric and causal evidence that representations derived from the diffmean of role playing datatsets can mediate granularity of a given role.

Critical strengths
  • This paper uses clear evidence and an understandable argument. Namely, I appreciate the use of random controls, causal evidence, human validation/rating of steered outputs. This is more thorough validation than some papers.
  • The use of PCA and diffmean are well motivated.
  • I also appreciate the use of qualitative examples. These are meaningful and helpful for understanding the argument.
  • Replication across multiple models is helpful as well
  • The teacher-forced continuation experiment and the matched-domain controls help the construct validity and reasoning-level concerns
Critical weaknesses
  • My main concern is that the notion of "granularity" doesnt seem to be well defined. That is that granularity as you use in this paper seems to be only related to levels of policy intervention, as opposed to a more general notion of granularity which would transfer to other tasks and situations. This is fine. But that notion of granularity is very limited in its applications and interpretation. This unclear definition also brings up questions about construct validity as well.
  • I also notice that there is limited external validation of the effect on reasoning. In your motivation you cite “individual reasoning where systemic analysis is needed, or institutional abstraction where local judgment is appropriate”, while we see LLM/human judgements and qualitative examples, to me, this seems to enough to support a notion of stylistic or surface level change (indeed the qualitative examples seem to only vary in small ways), but not necessarily that the reasoning itself is changing. I would like to see evaluations which substantively test the reasoning itself. For example, if we are motivated by representational validity, survey bias, social simulation, why not demonstrate that this intervention meaningfully addresses these concerns?
  • I also worry about the claim that this is a “representational primitive, beyond a stylistic surface variable”. I dont know that we see enough evidence to support this claim, namely there has been multiple works which supports the idea that behaviors that coherent on the surface level and can be steered together are not necessarily “primitives” (1, 2).
  • The concerns above come together make me question the overall usefulness of this study. If the papers notion of granularity is very narrow and it doesnt convincingly show that the narrow notion of granularity it studies is at a deeper reasoning level then it doesnt demonstrate that this granularity axis would actually solve representational validity, survey bias, social simulation. It remains open whether this result is practically useful. Then my other concern would be novelty. At this point it is well understood that essentially any coherent phenomenon/behavior is represented accessibly in the residual stream of these models. And especially since the “representational primitive” claim is insufficiently tested, I worry that this paper does not meaningfully advance our understanding in this regard.

References

  1. Vennemeyer et al. (2026), Sycophancy Is Not One Thing: Causal Separation of Sycophantic Behaviors in LLMs. ⁠arXiv:2509.21305
  2. Zhao et al. (2026), LLMs Encode Harmfulness and Refusal Separately. ⁠arXiv:2507.11878
Questions for authors

Please define granularity more specifically

Reviewer_PnkUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a contrast-based "Granularity Axis" framework to investigate how large language models represent and modulate social role granularity in internal activation space. By curating 75 social roles organized across five hierarchical levels of granularity (15 roles per level), the proposed pipeline executes five stages, including 1) generating model responses to shared evaluation questions under diverse prompts, 2) extracting role-level hidden-state representations from intermediate layers, 3) constructing a linear contrast-based granularity axis, 4) validating the geometric properties of the axis (e.g., checking alignment with principal components), and 5) conducting causal probing via activation steering to test behavioral shifts. The framework is evaluated across four open-weight checkpoints (Qwen3-8B, Qwen3-32B, Llama-3.1-8B-Instruct, and Llama-3.1-70B-Instruct). The authors report that the extracted Granularity Axis strongly aligns with the first principal component (PC1) of the role vectors and correlates with the five-level hierarchical ordering, claiming that social granularity serves as a linear representational primitive in LLMs.

Critical strengths

• Intriguing and Well-Motivated Direction: The paper addresses a conceptually interesting and relevant topic in mechanistic interpretability—understanding how internal hidden representations encode social hierarchies and varying granularities of social roles.

• Systematic Conceptual Pipeline: The end-to-end framework systematically integrates multiple stages (from elicitation and representation extraction to geometric validation and causal activation steering), offering a structured blueprint for testing internal latent directions.

Critical weaknesses

• Severe Presentation Deficiencies and Poor Editing: The manuscript appears to be largely unedited AI-generated text (as also acknowledged) and is difficult to read. The text is heavily congested with disjointed clauses separated by semicolons and walls of model-specific numbers devoid of intuitive context (e.g., Section 3.6: "Held-out prompts/questions remain strong; Qwen role hold out is slightly fragile."). Crucial terms and acronyms are introduced abruptly without definition (e.g., "PC1" in the abstract, "L3" in the conclusion).

• Missing Problem Formalisms and Technical Ambiguities: Fundamental aspects of the methodology are underspecified. Key concepts—including the hidden activation dimension d across architectures, the definition and design of the level assignment function y(u), the metric formulation of "role-level Spearman," and reporting conventions like interval brackets and "joint scale partial R^2"—are left ungrounded, significantly hurting technical clarity and reproducibility.

• Absence of Demonstrated Downstream Utility: While the study demonstrates geometric correlations and activation shifts, it provides no concrete application, downstream task, or practical use case demonstrating why this Granularity Axis is useful. Broad assertions such as "these findings establish social granularity as a representational primitive" remain vague and fail to substantiate practical impact.

• Limited Scope of Evaluation: The evaluation is restricted to a small set of open-source models, leaving it unclear how generalizable these representations are across broader architecture families, training paradigms, or closed frontier models

Questions for authors

**********Clarifications Materially Impacting the Assessment:

--Formal Problem Formulation (Section 2.1):

--How is the hidden dimension d defined across different LLM architectures? Is it assumed to be known and identical, or does it vary per model?

--How is the level function y(u) formally designed and mapped? Why is its operational design omitted from the text?

2.Role Selection and Curation (Section 2.2):

--How were the 75 roles and the five granularity levels chosen? Are these levels and roles considered exhaustive or merely representative? How sensitive are the findings to prompt phrasing and role selection?

3. Statistical and Metric Clarifications (Sections 2.3, 3.3, 3.5):

--What is the exact mathematical definition of "role-level Spearman," and how does it formally measure agreement across individual roles?

--In Section 3.3, what does Krippendorff’s α measure in this specific experimental context?

--In the statement: "Matched-domain continuation scale coefficients are 0.979 [0.898, 1.064] and 0.253 [0.238, 0.267], with joint scale partial R2 = 0.036/0.153", what do the bracketed intervals represent (e.g., 95% CIs?), and what is the exact definition of "joint scale partial R^2"?

--In Section 3.5, what does "prompt-level SEM" formally denote?

4. Intervention and Model Setups (Sections 2.4 & 4.3):

--In Section 2.4, how exactly is the intervention layer determined by the layer sweep? What objective function or threshold governs this selection?

--In Section 4.3, the text states "Our four checkpoints", whereas earlier sections introduce three model families/scales. Could you explicitly clarify the exact set of four checkpoints evaluated?

5. Practical Downstream Utility:

--What concrete downstream task or user-facing application (e.g., role-playing fidelity, alignment steering, instruction following) demonstrates the practical utility of manipulating the Granularity Axis?

*************Secondary Clarifications and Suggestions for Revision:

1. Exposition and Logical Transitions:

--In Section 3.7, what specific baselines or data do "These historical aggregates" refer to?

--In Section 4.1 ("representation should be validated before control"), the logical progression between sentences is unclear. Could the authors restructure this paragraph to clarify the core argument?

--What is the concise, actionable takeaway intended from Section 4.2?

2. Abstract and Conclusion Clarifications:

--Define technical acronyms upon first use (e.g., "PC1" in the abstract; "L3" in the conclusion).

--In the conclusion, the paper states: "Representation is robust; control depends on the model's default regime. This enables..." Which specific representation is robust, and what empirical evidence justifies that claim?
Reviewer_2SVAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes the “Granularity Axis,” a single latent direction in LLMs that orders social roles from individual to institutional scale and can be used to steer the granularity of generated responses.

Critical strengths

The motivation is convincing and offers a new perspective on how we should approach training/evaluating LLMs for social roles: can agents internally understand and distinguish between different social roles? This perspective will enable more interpretable insights into model behavior, which may help us better steer models and train them for specific roles.

The authors provide enough experimental evidence that the granularity axis can be a robust representational structure: it aligns with the dominant PC of role space, preserves the L1–L5 ordering, generalizes across model families and scales, and survives in surface wording or formality.

Critical weaknesses
  1. Some design choices are not fully justified. For example, the axis is constructed by contrasting {L1, L2} with {L4, L5}, but the paper does not explain why L3 is held out or why only a single level is used for this test. Although Appendix C.2 reports some ablations regarding this, it does not specify enough details (e.g., which endpoint definitions were tested). Would the results, particularly the placement of L3, change if we included L3 in the contrast or held out a different level instead?

  2. The validity of granularity levels is also unclear. The authors mention that the levels were first generated by an LLM and then reviewed by them, but it does not explain how each level was defined or why the roles were divided into five levels. Could the levels instead be grounded in an established framework from the social sciences? This could affect how you interpret the results. For example, L4 and L5 are barely distinguishable in Table 2. Is this because the model does not distinguish these two levels, or because the level definitions themselves overlap?

  3. The steering results are relatively weak, which makes the positioning of the paper unclear. The abstract describes the axis as “behaviorally causal,” but the steering effects are small or unstable in several settings. How do the authors position this work: primarily as an interpretability finding about how LLMs represent social roles, or as a practical tool for controlling social granularity in generation? I believe either positioning is reasonable, but the claims and evaluation should be aligned with the chosen one.

Questions for authors

Please see the Critical Weaknesses part!

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.