acceptodds
Under review as a conference paper at ICLR 2027

The Geometry of Contextual Relations: Language Models Address Facts by Order of Mention

Abstract

Human reasoning depends on how objects are related within propositions. How do relations organize the language representations of contextual contents? We give an LLM a list of facts in its context (e.g., Alice eats an apple. Bob eats a pear.) and measure how its hidden state changes when the question switches from what Alice eats to what Bob eats. Averaged over many lists, this change is a steering vector, which we call the ordinal vector. It points to a fact by its order of mention, the order in which the facts were stated in the context. We find that LLMs represent the fact a question asks about by its order of mention, not by the name the question contains. We state this as the ordinal addressing hypothesis: each order of mention has a fact address in the model's state, shared by all contexts, and a question moves the state to the fact address of the fact it asks about, while the context supplies what that fact says. Across Qwen, Gemma, and Llama, fact addresses are (1) ordered by mention: query states are organized by the order of facts, not of names, even when one fact has multiple subjects; (2) steerable: added to a question about the first fact of a new list, the ordinal vector makes the model answer with the second fact of that list; (3) low-rank: they span a low-rank subspace in which the first-mentioned fact is the easiest to reach, surprisingly similar to human recall; and (4) emergent: they are shared in late-middle layers, hold from 1.5B to 32B parameters, and form early in pretraining. Language models reach a stated fact by where it was mentioned, deepening our understanding of LLM reasoning.

Primary area:
interpretability and explainable AI
Keywords:
Interpretability, Representation Learning, LLMs, Contextual Relations, Geometry

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_bCosOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper asks whether a model finds a fact in its context by the name in the question or by the order in which the fact was stated. It uses short lists of four facts, such as "Petra eats the buckwheat. Freya eats the amaranth." The difference in hidden state between questions about two positions, averaged over other lists, is added once to a new list. The question text stays the same. The model often answers with the fact at the target position. The authors report that this position code follows order of mention rather than names, transfers across several relations, and appears in Qwen, Gemma, and Llama.

Critical strengths

The transfer test separates position from names. The vector is estimated on other lists and other names, then added to a new list while the question still names the first person. In Qwen the answer switches to the target fact about 84% of the time, and a random direction almost never does. Reversing the facts reverses the direction, and a vector from one relation works on another. This is direct evidence that, in these lists, the question moves the state to a position and the current list supplies the content.

Critical weaknesses

The finding is not shown to be general, and the title claims more than the experiments support. "The geometry of contextual relations" suggests a broad account of how language models represent relations. The experiments stay inside short templated lists: four to eight facts, one content per fact, and a fixed question that asks for that content. Contents are not only food. Appendix D.3 includes animals, cities, jobs, instruments, and sports, and a vector from one relation transfers to another. The task does not change. Short paragraphs still contain four facts plus a few unrelated sentences. Long documents and multi-step reasoning are not tested, as the authors note. A result on this one task does not support the broader claim.

The effect is also uneven: about 84% in Qwen, 54% in Gemma, and 36% in Llama on the main test, against a ceiling near 92–97%. It should not be described as a uniform mechanism.

Questions for authors

Q1: Which result is supposed to show that order-of-mention addressing is a general property of contextual relations, rather than a property of these short one-content fact lists? If the same facts are placed in a longer passage, with unrelated material between them, does the vector still select the target fact? This answer would affect my score.

Q2: The edited layer was chosen by how well the vector works. How much of the result remains at a layer fixed before seeing the test lists?

Reviewer_1twFOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes that LLMs represent queried facts through shared addresses tied to their order of mention. It extracts ordinal vectors from query-state differences and uses them to redirect answers across contexts. Experiments across model families, layers, and training checkpoints examine the geometry, transferability, and emergence of these addresses.

Critical strengths
  1. The ordinal vector is an interesting finding. Its transfer across names, contents, and relations provides evidence for a shared structure in contextual fact retrieval.

  2. The controls for fact order, coordinated subjects, and token distance strengthen the interpretation beyond a simple correlation with input position.

Critical weaknesses
  1. The mechanism remains underdetermined. The results support an order-related representation, but also appear compatible with entity matching followed by position encoding. This limits the claim that models retrieve facts by order rather than by name.

  2. The evidence is limited to simple fact retrieval, including templated paragraphs where transfer can be weak. Its relevance to broader reasoning remains unclear. The theoretical propositions largely follow from the assumed address structure and do not independently resolve this gap.

Questions for authors
  1. What experimental prediction distinguishes ordinal addressing from entity matching followed by position encoding? Could the authors provide a test where these explanations predict different outcomes?

  2. Is there evidence that the proposed addresses predict or causally affect errors in a multi-step reasoning task? Such evidence would help establish the broader significance of the finding.

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

The paper argues that language models access facts from lists of parallel statements using vector representations of fact order, which they call ordinal vectors (something like "fact #3 vector"). They support the claims with steering experiments on language models from three families. Three corruptions/perturbations of the lists are applied. The claims are based on significant amounts of experimentation.

Critical strengths
  • The paper features very nice visuals throughout.
  • Experiments disentangling the number of names and statements (conjunctions of multiple names per statement) are, as far as I can see, novel and interesting.
  • analyzing the emergence of fact lookup mechanism through training of OLMo is novel and insightful.
Critical weaknesses
  • It's not clear to what extent the claims are novel. Claims 1 and 2 as stated could be equivalent to what is established in the previous literature on the topic, which the paper cites. Comparison to previous work reads as a side note in the related work section, but it should be the core of the argument if novelty is claimed in such a well-researched field. Claim 3 is fairly trivial, although experimental support for the robustness of ordinal vectors is a plus.
  • The presentation of claims is overly simplified: fact IDs, encoded in ordinal vectors, is all there is to the mechanism. But And Gur-Arieh et al. show that multiple mechanisms are at play. Further, obviously, facts cannot be looked up by order of mention alone without names playing any role, because the fact to retrieve depends on the name in the question. Previous work on the topic, especially Prakash et al., does discuss the relevant mechanics carefully in more detail. Unfortunately, the main text of the paper seems to pretend the names play no role at all (again, this is not possible). The role of names is only acknowledged in the appendix.
Questions for authors
  • Among the four claims as stated in the abstract and in the introduction, only claim 4 is non-trivially novel and insightful. Unfortunately, only a short section of the paper (7.2) is dedicated to it; I wish it was the main focus of attention, discussed in much more detail.
  • The math doesn't seem load-bearing, mainly rewriting claims using a formalism, and could cede place to other content.
  • How do you think fact recall extends to more natural types of text? Highly parallel lists of a few facts are a highly artificial setting. You start exploring this direction in Appendix G.3. What would it take to make the setting even more realistic?
  • What do you think explains differences between models? For example, some of your reported results (Table 1) suggest that Llama is different from other model families.
  • Why did you choose R^2 as the main metric to report?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.