acceptodds
Withdrawn from ICLR 2027

The Geometry of Knowledge Accessibility in Large Language Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Large language models (LLMs) contain broad knowledge, but they cannot access all of it reliably. We study this problem through knowledge accessibility, which describes whether the knowledge needed for a query can be recalled from the model. We find that knowledge accessibility has a simple geometric structure in the model's representation of the query alone, before any generation. More accessible queries are closer to a center in the representation space, while less accessible queries are farther away. This geometry reveals a knowledge boundary that separates more accessible queries from less accessible ones. Accessibility consistently decreases with distance from the center, and this distance-based ordering transfers across datasets even when the centers differ. Controlled experiments further show that the centered geometry is more closely related to knowledge accessibility than to reasoning difficulty. The geometry also reveals when different interventions are useful. Query rewriting helps more for accessible queries, chain-of-thought reasoning helps more near the boundary, and retrieval gives larger gains beyond the boundary. These findings not only provide a new geometric view of how knowledge is organized in language models, but also suggest a useful pre-generation signal for adaptive inference.

Primary area:
interpretability and explainable AI
Keywords:
Knowledge Accessibility, Knowledge Boundary, Representation Geometry, Language Models

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_V5t7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper focuses on the geometric organization of knowledge accessibility in LLM representation spaces. The authors hypothesize that queries whose required knowledge is more accessible lie closer to a learned accessibility center. The authors compare isotropic and anisotropic center-based gemoetries against linear and nonlinear probing baselines. Across two LLMs and several QA datasets, distance from the learned center correlates with factual recall, transfers partially across datasets, and behaves differently from reasoning capability. Then, the authors further examine how query rewriting, chain-of-thought reasoning, and retrieval-augmented generation help in different regions of the learned knowledge boundary.

Critical strengths

S1. The question of focus in this paper is interesting and well-motivated. The attempt to separate accessibility from reasoning capability is a valuable distinction.

S2. The formulation is simple and interpretable. The comparison against linear probing, anisotropic distance, and an MLP provides initial evidence about this inductive bias.

S3. The downstream analyses are useful and thorough. Correctness stability near the boundary, cross-dataset transfer, and the differing effects of rewriting, reasoning, and retrieval make practical suggestions for adapting inference strategies to knowledge accessibility accordingly.

Critical weaknesses

W1. The experiment scope is quite limited. The main experiments use only two relatively small instruction-tuned LLMs, and the intervention experiments are restricted to TriviaQA. The support is thin for a claim about a general property of LLM knowledge accessibility. Larger models, additional families, and reasoning-capable models would help. It also remains unclear whether the geometry exists in the corresponding pre-trained models or emerges during instruction tuning, and whether tuning shifts the center or the shape of the boundary. In addition, testing on temporal updates, conflicting knowledge, and safety-related queries would be interesting.

W2. The single-center assumption is insufficiently justified. LLM knowledge spans diverse domains that may form multiple disconnected or non-convex regions, and the paper's own transfer experiments show that learned centers and dimension weights differ considerably across datasets. Comparisons against multi-center or mixture-based geometrics would establish whether this assumption is appropriate or overly resttrictive. It would also help to test whether paraphrases of the same question occupy similar regions.

W3. The empirical advantage over nonlinear baselines is marginal. In Figure 1, the anisotropic centered model beats the MLP only slightly on three of four factual QA settings and loses on the fourth. It is unclear whether these gaps are statistically significant. The setup also needs clearer description. The main paper, instead of Appendix A, should explain how the center and dimension weights are optimized, what objectives are used, how scores become correctness predictions, how per-layer AUROC is computed, and how hyperparameters are chosen.

W4. The writing is generally clear, but several paragraphs repeat the motivation and implications of the accessibility-center hypothesis over and over again. The related-work discussion could also be more comprehensive, especially regarding the organization of knowledge in representation space, knowledge-status characterization, and the effects of instruction tuning on knowledge accessibility.

Questions for authors

See weaknesses.

Reviewer_j3AsOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies knowledge accessibility: whether an LLM can recall the facts needed to answer a query from its parameters. Using the hidden state of the final query token before answer generation, the authors learn an accessibility center and a diagonal distance metric, defining interior, boundary, and exterior regions. They compare this geometry with linear and MLP probes and examine its relationship to factual recall, reasoning failures, and inference-time interventions.

Critical strengths

The paper asks a meaningful question beyond whether accessibility information is encoded in hidden states: how that information is geometrically organized. The effort to distinguish factual recall from reasoning failures provides a clear motivation and a useful direction for analysis.

Critical weaknesses
  1. The accuracy and stability patterns may also arise from generic correctness predictors. They do not clearly establish what centered geometry adds beyond correctness prediction.

  2. Centered and Linear+Norm use different training objectives and regularization, making their performance gap difficult to attribute to the centered inductive bias.

Questions for authors

Do the stability and intervention patterns also hold when queries are grouped by linear or MLP probe scores? This comparison would help establish whether these patterns specifically support centered geometry.

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

The paper looks at whether an LLM’s ability to recall a fact is already possible in the hidden state of the query before generation, and argues that it is organized around a “center”: queries whose knowledge is accessible sit close to it, and accessibility drops the farther out you go. The authors fit distance-to-center scorers on Llama-8B and Qwen-14B and compare them with linear and MLP probes. They then show the boundary lines up with where answers become unstable and that the distance ordering transfers across datasets. Further analysis show that rewriting, CoT and RAG each help most in a different region (interior, boundary and exterior).

Critical strengths

The paper asks whether knowledge accessibility is organized in the representation, not only whether it can be probed, which is a more interesting than most knowledge-boundary work. The Linear+Norm comparison is an important control and addresses the obvious concern that the centered model is just a linear probe with a norm feature. The stability analysis is the strongest result: repeated sampling isn’t used to fit the boundary, yet correctness is least stable right around it. The paper is well written and honest about its limitations such as that the evidence is correlational.

Critical weaknesses
  • Coming from a retrieval background, the method reminded me of SetFit-style embedding classification: representations followed by a lightweight classifier, here implemented as distance to a learned center c. The important difference is that SetFit explicitly trains the encoder to create useful embedding structure, while here the interesting claim is that such a structure already exists in the pretrained representation. I am not convinced that the experiments show this. The center c is learned from correctness labels, so at least part of the observed geometry is introduced by the supervised classifier itself. In that sense, the experiments clearly show that accessibility can be predicted from the hidden states with a very simple head, but I am less sure they show that accessibility is inherently organized around a center before fitting this head.

  • The paper's Linear+Norm experiment makes this concern stronger. Please correct me if I am wrong, but expanding −∥h−c∥2-\|h-c\|^2 gives a linear term together with a norm, and Appendix B.1 shows that a freely learned Linear+Norm model also converges toward the centered form. To me this mainly shows that this is a good structure for separating correct and incorrect queries, but I don't think it necessarily shows that the representation space itself has an accessibility center. A simple additional experiment could make this much more convincing, by e.g. using the mean representation of correctly answered queries as the center without optimizing c, or directly showing that correct queries are more tightly clustered than incorrect ones. The difference to the MLP is also quite small on TriviaQA, so I currently read the result more as a simple structured classifier performing about as well as a more flexible one rather than as strong evidence for the geometric claim.

Questions for authors
  • How should we interpret the centered scorer beyond a simple structured classifier on frozen representations? What evidence shows that the centered structure is a property of the pretrained representation space itself, rather than just a "useful" supervised inductive bias?

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

Related papers

Discussion (0)

Sign in to comment.