acceptodds
Under review as a conference paper at ICLR 2027

Hierarchical geometry and reference-derived steering in Gemma-2-9B

Abstract

Language models can represent hierarchical relationships geometrically, but the presence of this structure does not establish when it influences an answer. We examine this distinction in Gemma-2-9B across 46 noun hierarchies containing 690 concepts, tracing representations through all 42 transformer blocks. Hierarchical distance structure strengthens in early blocks, while reference-fitted principal components capture broad category distinctions more strongly than finer distinctions. We then add reference-derived category directions to individual concept-token states, excluding query leaves from their own reference fits and using the known target category to orient each edit. At B10, these interventions improve coarse and fine answer accuracy by 2.85 and 3.40 percentage points, respectively. Benefits diminish at later sites despite substantial retained category structure. In a paired cathedral example, a B34 edit strengthens the correct geometric preference while leaving the answer incorrect. State transfers identify the final prompt position as carrying a consequential difference from the successful early edit. Component interventions reveal contributions through this state from later attention and MLP computations, including B23 attention. These findings distinguish hierarchical representation from behavioral influence and show why evaluating geometric interventions requires tracing their effects through subsequent computation.

Primary area:
interpretability and explainable AI
Keywords:
large language models, hierarchical representations, semantic geometry, mechanistic interpretability, activation interventions, category prediction

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_JUUCOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates hierarchical concept representations in Gemma-2-9B and their relationship to activation steering. The authors analyze representation geometry across layers, perform reference-derived interventions, and use activation patching to trace how these interventions influence final predictions.

Critical strengths

The paper presents an interesting observation that improving geometric representations does not necessarily lead to correct decisions. However, the authors do not develop this observation into a sufficiently novel mechanistic explanation. The work mainly combines established techniques, relies on ground-truth information for its central interventions, and provides evidence from only one model.

Critical weaknesses

The authors show that early interventions are more effective than late interventions and that transferring the final prediction-position residual state can restore some correct predictions. However, this is not particularly surprising in an autoregressive Transformer, where the prediction position aggregates information from preceding tokens. The paper does not adequately distinguish a hierarchy-specific computational mechanism from general information propagation across layers. Furthermore, the evaluation is restricted to Gemma-2-9B, limiting the generality of the conclusions.

A major concern is that the steering direction is explicitly determined using the known correct category, as stated in Section 2. Therefore, the intervention does not demonstrate autonomous error correction or discovery of the correct semantic category. Instead, it injects ground-truth information into the model. The reported improvements of 2.85 and 3.40 percentage points at B10 are insufficient to establish a novel hierarchical decision mechanism.

Questions for authors

Where is your contribution in Section 1?

Reviewer_5UnNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper investigates how hierarchical concept structure is represented across depths in Gemma-2-9B and tries to establish when that structure, if present, influences model behavior. They use wordnet heirarchies to create the relevant hierarchical targets and show that tree-structure representations peak around early to middle layers of the model. Upon adding reference category information to the concept tokens (alinged to the correct answers) they find some gains in accuracy when intervening on early layers. Overall, the paper tackles an important question of when representational structures influence behavior. The general design and experiments are well constructed, but it takes several reads to grok the relevant information pointing to revisions needed to the prose to make this paper ready for publication. I present more detailed comments in the sections below.

Critical strengths
  • The question of whether geometric structure in concept categories influences outputs is increasingly an important topic in model interpretability and this paper’s contribution in that space seems largely distinct.
  • The range of interpretability methods brought to bear on the question is impressive and are largely well-motivated.
  • The controls applied throughout are careful and lend credence to the core claims of the paper
Critical weaknesses
  • While the flavor of the contribution is unique, the failures in late-edits shown by the authors seem largely in line with earlier work in the field on factual recall, where editing concept-level tokens in late layers has limited bearing on the final answers. Could the authors distinguish their work from earlier work in this area (e.g., Geva et al., 2023)?
  • Since the baseline accuracy is already high (near 90%), the room for improvement is relatively limited. I wonder how much of these effects would hold with larger models. It would maybe be helpful to report gains as a proportion of correctable errors?
  • I found the paper generally hard to read. Perhaps it was the use of notations when not needed and the somewhat sterile presentations of the methods, which makes it hard to infer the motivation behind them. The figures are quite small as well and could benefit from larger labels. I’d recommend a thorough editing pass focusing on readability.
  • While case studies on single models are helpful, it is difficulty to justify a main conference acceptance without some generalization to other models, even if they are of comparable sizes. The limitation of the prompt template and task, though they help in the controlled nature of the methods, limits its generalizability.
Questions for authors
  • Is it a concern that the size of the edits are not matched across layers?
  • How do the authors reconcile the late-edit failure with known results on attention-mediated information in earlier work? Do the authors think late edits at the final position would succeed?
Reviewer_ZzAhOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper is a controlled study of interventions propagating through Gemma-2-9B using hierarchical trees from WordNet that are applied on some blocks of the LLM. It shows there are category-specific coordinates but that this does not necessarily imply a correct answer.

Critical strengths

This work might indicate important interpretability relationships within a LLM transformer blocks that might be mapped to how linguistics has studied concept hierarchies. It shows how intervening on different points in the model could indicate latent representations. The experiments with held-out constructions and tree transfer are substantial and are relevant to the specific case being presented.

Critical weaknesses

I could not identify what is novel from this work in comparison to prior literature and the current empirical experiments are limited to a single model and 46 trees, leading me to think the results could be specific to this LLM or explained by confounding variables. In my view, it is impossible to determine what is definitely hierarchical about these interventions versus just categorical steering or some simple information matching.

Also, none of the claimed contributions are particular novel and it would be important to asses these conclusions across multiple models of the same family, different families and so on, like Gemma 3, Gemma 4, Gemma 4 Encoder Free, Llama 3, etc to build statistically significant and scientific relevant conclusions. The paper needs better task definition and usage of nonhierarchical control tasks so we can at least try to compare what the interventions on each block would actually change versus some alternative approach.

In that line of thought, it leaves behind many works in hierarchical representations, linear concept geometry, activation steering, and causal information flow. For example, [1-4] have developed alternative methods for studying and steering with concepts and hierarchy structures that are important to be compared to.

  • [1] Hernandez and Andreas, The Low-Dimensional Linear Geometry of Contextualized Word Representations, CoNLL 2021
  • [2] Elazar et al., Amnesic Probing: Behavioral Explanation with Amnesic Counterfactuals, TACL 2021
  • [3] Park, Choe, and Veitch, The Linear Representation Hypothesis and the Geometry of Large Language Models, ICML 2024
  • [4] Park et al., The Information Geometry of Softmax: Probing and Steering, ICML 2026
Questions for authors
  1. How can you confirm there is a hierarchical or fine-grained nature in the reported mechanism?
  2. Can the authors distinguish what depends on the hierarchy versus a generic category membership?
  3. Could the core results be generalized across models and families and tasks in a more sophisticated evaluation procedure?
  4. Please compare with hierarchical-subspace and causal usage studies. The paper has no related work section, insufficient discussion and comparison, doesn't properly introduce the problem, its importance nor the field. How do your contributions meaningfully differ from the prior work?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.