acceptodds
Under review as a conference paper at ICLR 2027

Context Is King: How In-Context Specification Shapes the Geometry of Concepts

Abstract

Large language models place structured concepts on geometrically faithful manifolds: weekdays lie on a circle, months on another, often read as a fixed world-model the network stores and looks up. Prior work shows that context can affect entity-token geometry while a strong pretrained prior remains prominent. We show where context takes over: at the final prompt token, after the model has read the rule and query, geometry follows the in-context specification and dominates the prior. A declarative rule changes not just the ordering of entities but the geometry's relational topology: the same tokens form a cycle or a hierarchy with tree depth and branch relationships on command. This change cannot be explained by simple relabeling and holds even for arbitrary nouns randomly assigned to the structure's nodes. Under conflicting specifications, capable models in both tested families (Gemma, Qwen) show representational similarity of – to the imposed structure versus near-zero to the pretrained order, across the concept sets tested. To investigate the computation underlying these geometric findings, we use activation patching and find that replacing one entity's activation with another redirects the answer to the donor's successor under the context-specified order. Sweeping the intervention across layers localizes where entity identity enters this computation. Smaller models can already form rough context-specified geometry, while larger models more consistently suppress the pretrained order and follow the active relation under patching. This family-dependent trend cautions against transferring mechanistic conclusions across model sizes. Whether the model builds this geometry anew or reconfigures a stored one remains open; operationally, the geometry it uses is the one the context specifies. Code and artifacts: https://github.com/anonauthoriclr/context-is-king.

Primary area:
interpretability and explainable AI
Keywords:
Mechanistic Interpretability, Representation Learning, Representation Geometry

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_SvaTOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper analyzes the geometry of concept representations in language models, such as circular configurations of weekday representations. More specifically it studies if and how contextualization and in-context learning can change the "default" canonical structures.

Even though the topic falls squarely into my area of expertise, the paper is difficult to understand due to the incoherent writing.

Critical strengths

none

Critical weaknesses

From what I can tell, I see neither much novelty nor technical contributions or insights.

Lack of novelty: It is already known that in-context learning can override semantic priors and that this reflects in internal representations (e.g., the cited "In-context learning of representations" by Park et al. 2025), and it is also known that contextualization affects representational geometry. For example Tiblias et al. find that "Different continuations produce drastically different topologies."

Tiblias, F., Bigoulaeva, I., Niu, J., Balloccu, S., & Gurevych, I. (2026). Hypothesis-driven feature manifold analysis in LLMs via supervised multi-dimensional scaling. Transactions on Machine Learning Research. https://openreview.net/forum?id=vCKZ40YYPr

Lack of contributions or insights: The conclusions section summarizes the paper's findings as follows:

Across two model families, pretrained organization persists at the entity token but gives way down- stream to context-specified relational geometry. The same entities form a cycle or hierarchy on command, and entity identity is causally bound to the specified relation, with clean binding emerg- ing only at scale. Whether this downstream geometry is built or reconfigured remains open; either way, recovered “world-model” geometry must be interpreted relative to context and readout posi- tion.

None of this meets the standards of a top tier conference.

Questions for authors

none

1 comment

Dear Reviewer, thank you again for your review. As the discussion period ends on November 18, we wanted to check whether our response addressed your concerns, and whether there is anything further we could clarify or analyze in the remaining time.

Reviewer_fH9ZOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper looks into the structure of representations of LLMs and shows that the learned geometry in pre-training can be overwritten in context, more consistently in larger LLMs, but only if evaluated later in the input when the model has read the whole new conflicting context rules.

Critical strengths
  1. Interesting insight about how context can shape the representations of known entities.
Critical weaknesses
  1. The main question of the paper is mentioned as "can a context-specified relational geometry override the pretrained one, and if so, where in the computation does it take over?". However, Park et al. previously found that "When reference concepts have correlations in their semantics (e.g., Monday, Tuesday, etc.), the context-specified graph structure is still present in the representations, but is unable to dominate the pretrained structure.". So the main contribution of this paper seems to be changing the readout position.
Questions for authors
  1. To me, the most important distinction of this work is the readout position. Please share more about what substantial novelty this work provides compared with previous work.
1 comment

Dear Reviewer, thank you again for your review. As the discussion period ends on November 18, we wanted to check whether our response addressed your concerns, and whether there is anything further we could clarify or analyze in the remaining time.

Reviewer_m3oxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

LLMs, in their inner working, arrange concept representations in certain shapes that resemble the concept ordering relationships among them (i.e. weekdays, months, musical notes, etc). When prompted to modify these assumptions, their inner representations rearrange to match those new arrangements, which is observed in the state prior to responding. They test with cyclic-arrangements and tree-arrangements, to study the effect on geometry

Critical strengths
  • Well proposed question, clearly stated
  • I liked the fact that they introduced cycle-to-tree switch. It effectively shows that the geometry is adapted, and rules out the possible interpretation of mere label rearrangement without substantially modifying the geometry
Critical weaknesses
  • I'm not completely sure whether the reported geometry encodes what the authors claim (concept representation), or if it is just a way the model has to compute the answer. The centroids average the states of the final tokens over k-step queries, so they partly encode the answer that the model wants to output next
  • In the activation-patching result, the authors replace the activations of Monday (in vector space) by Friday. The results show that the model computes the next element from context, which explains a behaviour rather than the actual geometry used by the model to represent the concepts. This limitation is admited by the authors in the paper, and weakens the claims stated in the abstract ("operationally, the geometry it uses is the one the context specifies").
  • When the model answers directly, the geometry looks supportive to successful editing, while the actual answer of the model tells a different story (imposed order is successful only 44% for Qwen-27B and 57% for Gemma-31B). This weakens the relationship between the observed geometry and the actual meaning that the model is trying to convey in its answer.
  • The paper tries to make sense to the effect of model sizes, but the number of sizes to support those claims is low. Moreover, they are not consistent across architectures (qwen-9B keeps the ring, while llama-8B fails).
  • Unclear novelty when compared against Park et al. (2025a)
Questions for authors
  • When the LLM answers directly, imposed order's accuracy is low, while the final-token RSA to the imposed order is high. Could you split the final-token centroids by outcome (correct vs incorrect) and report the RSA of the natural-order and imposed-order for each group?
  • In your words, what is the novelty of your approach when compared against Park et al. (2025a)?
1 comment

Dear Reviewer, thank you again for your review. As the discussion period ends on November 18, we wanted to check whether our response addressed your concerns, and whether there is anything further we could clarify or analyze in the remaining time.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.