acceptodds
Under review as a conference paper at ICLR 2027

ReasAtlas: Shared Expositions as a Representation for Mathematical Knowledge

Abstract

Mathematical knowledge is developed through expositions that organize definitions, theorems, and algorithms around related mathematical objects. Such organization situates individual statements within broader mathematical contexts and can serve as a representation of mathematical knowledge. We introduce ReasAtlas, which organizes mathematical knowledge from heterogeneous sources within a shared hierarchical exposition. Each statement can consolidate formulations across sources while preserving their source-specific content and provenance throughout structural revisions. The hierarchy provides each statement with a primary organizational context, while relations and cross-topic links capture mathematical connections beyond the primary hierarchy. To build this representation, ReasAtlas aligns source-local expositions, retrieves relevant regions of the shared organization, and uses source context to place new statements and revise the structure. The resulting atlas provides a shared organizational context for locating and retrieving mathematical knowledge across sources. Its hierarchy also identifies mathematical neighborhoods in which relevant statements must be distinguished from nearby alternatives. We introduce information-calibrated margin cross-entropy (IC-MCE), which uses shared ancestry to strengthen these comparisons during training while the resulting encoder operates on statement text alone at inference time. Our evaluation examines whether shared exposition can support both the localization of mathematical knowledge and the learning of retrieval representations, while also assessing the mathematical coherence of the resulting topic organization.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Mathematical Knowledge Representation, Hierarchical Knowledge Organization, Mathematical Retrieval, Representation Learning

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_y5EjOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper presents ReasAtlas, which organizes mathematical statements from multiple textbooks into a shared hierarchy. The hierarchy is then used to provide retrieval context and supervise embedding training through IC-MCE.

The paper reports some positive results, but I remain unconvinced by its motivation and the necessity of the proposed representation. The experiments support the usefulness of hierarchical information, rather than the broader claim that a shared exposition provides a meaningfully better representation of mathematical knowledge. I do not think the current contribution is sufficiently established for ICLR.

Critical strengths
  1. The construction of a large-scale mathematical knowledge hierarchy across 63 textbooks involves considerable effort. Preserving source formulations and provenance is potentially useful.

  2. The paper includes reasonably controlled retrieval experiments, particularly the comparisons between flat cross-entropy and IC-MCE, with encouraging improvements across two embedding models.

Critical weaknesses
  1. The motivation is not convincing. It is unclear why mathematical knowledge needs to be reorganized into a shared textbook-based hierarchy. The paper does not identify concrete limitations of existing retrieval methods that would justify this additional construction process.

  2. The constructed hierarchy lacks convincing validation. The paper claims to align equivalent mathematical knowledge across sources, but provides limited direct evidence that these alignments are accurate. The evaluation largely focuses on topic-level quality rather than the correctness of cross-source statement consolidation.

  3. The experiments do not fully support the central claims. Most retrieval queries are generated from statements in the constructed corpus, raising concerns about generalization. While the results show that hierarchical information can improve retrieval, they do not establish why the proposed shared exposition is preferable to simpler alternatives.

Questions for authors
  1. The paper evaluates topic-level quality, but how accurate is the cross-source statement consolidation itself? Is there any statement-level human evaluation, especially for formulations with different assumptions?

  2. Do the improvements from IC-MCE generalize to independently authored mathematical queries? The Stacks experiments evaluate contextual representations rather than IC-MCE.

  3. How does IC-MCE compare with existing hierarchy-aware contrastive learning objectives under the same training and sampling settings?

1 comment

Thank you for the additional experiments and clarifications. The new evaluation on independently authored mathematical queries and the comparison with HiMulCon address some of my previous concerns. However, I still have some reservations about the overall contribution. While the additional experiments provide useful evidence, I am not yet fully convinced that they sufficiently establish the necessity and broader value of the proposed representation.

Overall, the rebuttal improves the empirical evidence, but does not fully resolve my concerns regarding the motivation and broader significance of the work. I therefore raise my score from 1 to 2.

Reviewer_yU53OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces ReasAtlas, a framework for organizing mathematical knowledge from multiple textbooks into a unified hierarchical structure. Such structure could improve mathematical statement retrieval and topic identification. paired with ICMEC, it also help model to differ between closely related but mathematically different statements.

Critical strengths
  1. the paper introduces an interesting approach to organizing mathematical knowledge across heterogeneous sources through a shared hierarchical structure.
  2. empirical evidence that shared expository structure benefits mathematical knowledge retrieval.
Critical weaknesses
  1. The evaluation is largely circular and never uses real queries to test the learned encoder. Test queries are generated by GPT-5.6-sol from LLM-written representative formulations, and relevance is judged by another LLM call. As a result, the retrieval gains may not reflect real search behavior. Evaluating on mathematician-written queries, such as those of NaturalProofs etc, is necessary
  2. The atlas’s quality is not established. The human study shows 59.5% agreement with LLM, generally is about what chance predicts given both raters’ heavily skewed marginals (5 pts). Using the paper’s own Table 9, chance agreement is roughly 58.6%, So the human study does not validate the LLM scores
Questions for authors
  1. it seems IC-MCE is quite similar to hierarchical triplet loss (Deep Metric Learning with Hierarchical Triplet Loss, eccv18), wondering why there are no citations
  2. why no frontier-LLM baseline? such as reranking an off-the-shelf embedder’s top candidates with a strong model or an agent navigating native tables of contents.
  3. why not compare the shared-atlas path against simple baselines such as each statement’s native book chapter path?
Reviewer_nNYyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Shared Expositions as a structured representation for mathematical knowledge, aiming to capture the reusable reasoning patterns that underlie diverse mathematical problems. The authors propose a pipeline that extracts expositions from problem solutions using GPT-5.6-sol, clusters them by semantic similarity, and selects representative expositions via Iterative Contrastive Maximization with Entropy (IC-MCE). The resulting repository is used to augment LLM reasoning through retrieval-augmented generation. Experiments on mathematical reasoning benchmarks show modest improvements (1-3% absolute gains) over baselines, with the authors claiming that shared expositions provide a more interpretable and reusable alternative to full problem solutions.

Critical strengths

S1. The concept of Shared Expositions as reusable reasoning patterns is well-motivated and addresses a real limitation of current retrieval-augmented methods, which typically retrieve full problem solutions rather than abstract reasoning strategies. This abstraction could improve generalization to novel problems that share underlying structure but differ in surface features.

S2. The IC-MCE selection method is a reasonable approach to balancing diversity and representativeness in the exposition repository. The iterative contrastive mechanism avoids the common pitfall of selecting redundant expositions, and the entropy regularization encourages coverage of the reasoning space.

S3. The paper provides qualitative examples showing how shared expositions capture cross-cutting reasoning strategies (e.g., "consider the boundary case," "exploit symmetry"), which are more interpretable than full solution retrieval and could serve as useful pedagogical tools.

Critical weaknesses

W1. The use of GPT-5.6-sol for exposition extraction creates a self-reinforcing loop: the quality and style of shared expositions are entirely determined by a single proprietary model, and there is no validation that the extracted patterns are actually reusable or that they generalize beyond the training distribution. The paper does not provide human evaluation of exposition quality, ablation on extraction prompt variations, or analysis of failure cases where the extracted expositions are misleading or inapplicable.

W2. There is no downstream evaluation beyond mathematical reasoning benchmarks. If shared expositions truly capture reusable reasoning patterns, they should improve performance on tasks requiring similar abstract reasoning (e.g., scientific problem-solving, logical deduction, or even non-mathematical domains). The paper does not test this transferability, making it unclear whether the method captures general reasoning skills or merely memorizes mathematical heuristics.

W3. The reported gains (1-3% absolute) are thin and lack statistical significance tests. Given that the baseline already uses retrieval-augmented generation with full solutions, the incremental benefit of abstracting to shared expositions is modest. The paper does not report confidence intervals, ablation on repository size or selection strategy, or comparison with simpler baselines like retrieving the top-k most similar full solutions. Without this analysis, it is impossible to assess whether the added complexity of exposition extraction and clustering is justified.

W4. The construction process is opaque: how many expositions are extracted, how is the clustering threshold selected, and how sensitive are the results to these choices? The paper does not provide ablation on these hyperparameters, making it difficult to reproduce or extend the work. Additionally, the reliance on GPT-5.6-sol makes the method expensive and inaccessible to researchers without API access.

Questions for authors
  1. Can the authors provide statistical significance tests (e.g., bootstrap confidence intervals) for the 1-3% gains and demonstrate that they exceed run-to-run variance?

  2. How do the authors validate that the extracted expositions are actually reusable and not just GPT-5.6-sol-specific heuristics? Is there human evaluation of exposition quality or analysis of failure cases?

1 comment

I appreciate the work done by the SPC and the response provided by the authors. Having reviewed the relevant details, I am maintaining my rating.

Comments on the paper

1 comment

Dear Reviewers,

The authors have responded to your comments. Could you please review their responses and update your reviews?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.