acceptodds
Under review as a conference paper at ICLR 2027

Numerical Reasoning on Knowledge Graphs with Convolutional Histograms

Abstract

Numerical complex query answering methods often encode scalar values as latent vectors or learned density parameters. Although expressive, these representations do not preserve numerical order, distance, and uncertainty together in an explicit numerical space. Their intermediate numerical states can therefore be difficult to inspect in the original units. We present CHOREOGRAPH, a framework that represents each observed scalar as a localized histogram on an ordered grid defined separately for each numerical attribute. Attribute-wise quantile binning allocates finer resolution where training literals are dense. Relation projection kernels (RPKs) use one-dimensional convolution to model how numerical mass changes across relational edges and can be fitted efficiently. Distributional message passing then estimates unobserved numerical slots. CHOREOGRAPH composes numerical operations using histogram-valued intermediate states. This discrete representation supports comparison, set, and arithmetic operations without operator-specific learning. Each intermediate numerical state can therefore be traced to intervals in the original units. We evaluate CHOREOGRAPH on FB15K-237, FB15K, DB15K, and YAGO15K using benchmark queries from prior work. Across all four datasets, CHOREOGRAPH outperforms the evaluated baselines on most query structures. On FB15K, all RPKs are fitted in 9.4 seconds, and total training is – faster than recent baselines.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Knowledge Graphs, Complex Query Answering, Numerical Reasoning, Histogram Representations

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_6H73OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CHOREOGRAPH, a framework for numerical complex query answering over knowledge graphs. It represents numerical states as histograms and uses relation projection kernels and distributional message passing to support numerical reasoning. Experiments on four benchmarks show improvements over existing methods on most query structures, with reduced training time.

Critical strengths

Important and Underexplored Problem: The paper addresses numerical reasoning in knowledge graphs, an important yet relatively underexplored aspect of complex query answering. Explicitly handling numerical constraints within compositional queries is a meaningful research direction.

Critical weaknesses

Limited Methodological Novelty: The core idea of discretizing numerical values into bins has been explored in prior work, such as KGA. The paper does not sufficiently clarify how its approach fundamentally differs from existing numerical graph reasoning methods.

Limitations of Quantile-Grid Transport: The relation projection kernels model numerical shifts in quantile-bin space, where bins can have substantially different widths in the original numerical domain. This raises concerns about whether the learned kernels can consistently capture meaningful numerical changes across different regions of the value range.

Questions for authors
  1. Could the authors clarify what the main differences are between CHOREOGRAPH and KGA, beyond representing numerical values using bins?

  2. Since different bins can cover very different numerical ranges, why is shifting a histogram by the same number of bins a reasonable way to model numerical changes?

  3. Could the authors clarify how well CHOREOGRAPH performs on arithmetic queries compared with existing methods? Also, how does decoding later operands into scalar values affect the benefits of maintaining histogram representations?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 6H73

Q1. Differences from KGA

The distinction is how bins are represented and used, not binning itself. KGA supports fixed-width/quantile bins, adds them as graph nodes, chains neighboring and hierarchical bins, and trains an off-the-shelf KGE model. Its numerical prediction selects the highest-scoring bin and reads its median.

CHOREOGRAPH uses histograms as intermediate numerical query states, not embeddings of additional bin nodes. Each coordinate denotes an attribute interval and its mass. Multiple possible answers can remain in one state through comparison, set operations and arithmetic. Relation-attribute RPKs transform these states by convolution and are fitted by independent ridge problems followed by simplex projection.

Our contribution is the connection between an explicit numerical state, relation-conditioned transport and compositional query execution. KGA enriches the graph used to learn link predictors; CHOREOGRAPH exposes where intermediate numerical answers lie and how operations move and combine them. A complex-query executor is not supplied by bin augmentation alone. We added this distinction to Sec. 2.

Q2. Quantile-grid transport

The model in Sec. 4.2 estimates source-target displacement distributions in the attribute's quantile coordinates. Its rationale is statistical sharing: one K-coefficient kernel pools relation-attribute training paths across positions, instead of estimating a separate transition at each position. This reduces estimation requirements when numerical observations are sparse. A kernel can place mass on several shifts, representing spread or multiple possible changes rather than one fixed offset.

Quantile bins allocate resolution where training literals concentrate. The shared transformation therefore models changes in relative numerical position; it does not assert a constant physical-unit offset across the range. Arithmetic uses scalar representatives, not index shifts. All kernel coefficients are estimated from training-path targets, rather than prescribed from bin widths.

Table 4 supports allocating resolution according to literal density. Macro MRR (%), three-split mean ± sample SD:

BinningFB15K-237FB15KDB15KYAGO15K
Quantile17.30 ± 1.4321.64 ± 0.2010.43 ± 0.815.79 ± 0.81
Uniform15.45 ± 1.8319.87 ± 0.4310.23 ± 0.965.09 ± 0.89

Q3. Arithmetic comparison and scalar decoding

On the submitted Table 3 queries, we replaced only arithmetic with CNR-NST's released function, fixing input predictions and downstream execution. Three-split macro MRR (%):

ArithmeticFB15KDB15KYAGO15K
Submitted CHOREOGRAPH5.274.1810.14
CNR-NST function3.772.695.50

CHOREOGRAPH is higher on all three datasets.

Singleton ground truth does not eliminate predictive uncertainty; later scalar decoding removes it. With fixed queries and input predictions, retaining both histograms changes three-split macro MRR (%) from 5.27 to 5.86 on FB15K, 4.18 to 4.56 on DB15K and 10.14 to 10.46 on YAGO15K. On FB15K, 2b/3b improve from 7.20/8.36 to 10.31/11.20, while bp, pb and 2pb favor set-scalar execution. Retaining both distributions improves the overall mean on all three datasets and better matches CHOREOGRAPH's representation. Table 3, Secs. 4.5/5.2 and Appendix B.4 have been updated accordingly.

Thank you for the detailed response and additional experiments. The clarification of the differences from KGA is helpful, although the main methodological contribution beyond existing approaches could be further clarified. The new arithmetic comparisons and histogram-preserving execution also address my concerns regarding Q3.

However, my concern about quantile-grid transport remains. While the explanation of position-independent kernels is reasonable, the improvements of learned RPKs over simple identity transport are very limited, as shown in Table 14. This raises questions about the practical contribution of the proposed transport mechanism. Overall, while the response addresses some of my concerns, I maintain my original rating.

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

The paper introduces CHOREOGRAPH to improve numerical complex query answering in KGQA. It represents numerical states as histograms on attribute-specific quantile grids, estimates relation-conditioned convolution kernels through ridge regression followed by simplex projection, and propagates these distributions to infer unobserved numerical attributes. A ComplEx-based executor connects entity and numerical states, while explicit histogram operators implement comparisons, set operations, and arithmetic. Experiments evaluate LitCQD and NRN query structures across four datasets, together with restricted set–scalar arithmetic. The strongest results concern numerical-filtering performance and low numerical-component training overhead.

Critical strengths
  1. Substantial improvements on explicitly numerical queries. The method outperforms baselines across multiple dataset–query-structure comparisons. The paper also separately evaluates queries containing numerical operators. These results support the method’s practical value for the targeted task.

  2. Inspectable intermediate representations with low training overhead. Attribute grids and relation kernels directly expose the support of intermediate numerical distributions and their transport, without increasing training costs. The combination of traceable representations and inexpensive estimation is an important strength, even though query-time inference is not faster.

Critical weaknesses
  1. The gains are not sufficiently attributable to the core mechanism. Table 4 compares learned RPKs only with uniform transport, without an identity or copy-transport control holding attribute-existence scores fixed. The contribution of the existence model is also not separately isolated. Moreover, the “ComplEx-only” baseline in Tables 2 and 10 changes the graph representation and execution procedure (Appendix C.2). The current results therefore do not sufficiently establish whether the gains come from relation-conditioned histogram transport or other system components.

  2. Recent directly relevant baselines are insufficiently covered experimentally. The paper primarily compares with LitCQD and NRN, without including the directly relevant CNEQ (2024) and CNR-NST (2025). In particular, it lacks a common-task comparison with CNR-NST, which supports binary arithmetic. Although the appendix explains reproduction difficulties, the current results remain insufficient to establish the incremental value over these methods. A comparison on a shared query subset, a more specific and verifiable account of the reproduction obstacles, or a theoretical comparison would strengthen the assessment.

  3. The semantics and compositional consistency of the numerical operators require clarification. Eq. (6) uses products of probability masses, implicitly assuming operand independence; operands sharing entities, variables, or evidence may violate this assumption. Furthermore, comparison scores are normalized and subsequently processed as histograms, while numerical intersection uses normalized elementwise products (§4.5). The authors should distinguish probability distributions from query scores and specify the conditions under which these operators can be composed. Nonnegative, normalized outputs alone do not guarantee correct probabilistic or set semantics.

Questions for authors
  1. Mechanism clarification and focused validation: How does the “w/o Message Passing” ablation change the histograms and attribute-existence scores, respectively? With existence scores and other executor settings fixed, how does learned RPK transport compare with identity or direct copy transport? Stable additional gains would strengthen support for the core mechanism.

  2. Comparison setup clarification: Do the structural-only query results in Table 1 use the same ComplEx checkpoint and entity executor as LitCQD? Can a recent directly relevant method be compared on a shared query subset, or can the reproduction obstacles and remaining verifiable capability differences be described more concretely? Existing implementation evidence or a limited common-task comparison could affect my assessment.

  3. Arithmetic scope and assumptions: Does Table 3 exclusively use Eq. (9)? Are there existing results or tests for Eq. (6) with uncertainty in both operands? Does the query language allow operands to share entities, variables, or reasoning evidence? If so, how does the product joint distribution account for their dependence? How are the candidate-value set and derived-output domain constructed? If the contribution is restricted to independent inputs or set–scalar arithmetic, please state this scope explicitly.

  4. Validation of the existing operators: Are there executor checks using ground-truth input values or known input distributions to separate input-prediction errors from arithmetic approximation errors? How does unary scaling transport probability mass on nonuniform grids, and does arithmetic error vary with binning resolution? Existing implementations, small worked examples, or existing tests could clarify these issues; a new large benchmark is not requested.

2 comments
AuthorsOpenReview ↗

Response to Reviewer WvvW

Q1. Mechanism clarification and focused validation

The submitted w/o Message Passing (L=0) retains observed histograms but removes both inferred histograms and existence scores for unobserved slots.

With existence scores and other settings fixed, Table 4's numerical-filter macro MRR (%) over three splits is:

TransportFB15K-237FB15KDB15KYAGO15K
Learned RPK17.30 ± 1.4321.64 ± 0.2010.43 ± 0.815.79 ± 0.81
Identity17.09 ± 1.8421.21 ± 0.5210.43 ± 0.715.64 ± 0.88
Neighbor copy/mean16.75 ± 1.6020.98 ± 0.0310.52 ± 0.875.67 ± 1.06

Learned RPK exceeds both controls on FB15K-237, FB15K and YAGO15K (Appendix D.6, Table 14), adding relation-specific transport to informative neighboring literals.

We separately controlled attribute-existence scores with histograms and other components fixed:

InterventionFB15K-237FB15KDB15KYAGO15K
CHOREOGRAPH17.30 ± 1.4321.64 ± 0.2010.43 ± 0.815.79 ± 0.81
Attribute-existence score control18.02 ± 1.7421.55 ± 0.1510.47 ± 0.876.42 ± 0.26

MRR changes by at most 0.73 points, indicating robustness to this change in existence scores.

To address the ComplEx-only concern, we used its checkpoint in CHOREOGRAPH, keeping everything else fixed. On the same FB15K Table 10 queries, three-split MRR (%) is reported as mean ± sample SD:

MethodAvg_M1p2p2i3ipiip2uup
ComplEx13.25 ± 0.371.88 ± 0.238.02 ± 0.2428.06 ± 0.9840.00 ± 0.569.42 ± 0.6315.27 ± 0.751.05 ± 0.062.31 ± 0.13
CHOREOGRAPH, ComplEx baseline checkpoint18.85 ± 0.2224.14 ± 1.5014.94 ± 0.5921.29 ± 0.5030.64 ± 0.9716.82 ± 0.6517.58 ± 0.3411.15 ± 0.2914.23 ± 1.66
CHOREOGRAPH, original checkpoint19.19 ± 0.2024.14 ± 1.5015.43 ± 0.4521.69 ± 0.3730.18 ± 0.8317.68 ± 0.6617.79 ± 0.2811.96 ± 0.2614.67 ± 1.63

With the same checkpoint, CHOREOGRAPH improves macro MRR by 42.24%, remaining within 0.34 points of its original result. Different ComplEx training alone therefore does not explain the gain.

Q2. Comparison setup clarification

No. LitCQD jointly learns relations and numerical regression; our backbone uses relational facts only. Executors also differ (max-plus/z-sum versus fuzzy); Q1 addresses these factors.

CNEQ lacks processed graphs/queries and TransEA artifacts for its Entity Filter. CNR-NST releases Multi-ComplEx code and KBC arrays, but not numerical-predicate edge reconstruction. Reproduction needs mapping and preprocessing choices, not just checkpoint retraining.

Replacing only arithmetic with CNR-NST's released function, submitted Table 3 queries yield three-split macro MRR (%):

ArithmeticFB15KDB15KYAGO15K
Submitted CHOREOGRAPH5.274.1810.14
CNR-NST function3.772.695.50

CHOREOGRAPH is higher on all three datasets.

The adapter maps normalized candidate memberships and fuzzy outputs to the fixed grid, retaining released thresholds, collision handling and caps.

Q3. Arithmetic scope and assumptions

Submitted Table 3 used Eq. (9), decoding later operands to scalars. Keeping both histograms under Eq. (6) raises three-split macro MRR (%) from 5.27 to 5.86 on FB15K, 4.18 to 4.56 on DB15K and 10.14 to 10.46 on YAGO15K. FB15K 2b/3b rise from 7.20/8.36 to 10.31/11.20; bp, pb and 2pb favor set-scalar. Table 3, Secs. 4.5/5.2 and Appendix B.4 now use both distributions.

Shared entities/subqueries are allowed, but Eq. (6) uses independent marginals, without shared-evidence correlations. Comparison and product-intersection outputs are query scores, not calibrated posteriors.

Candidates include benchmark values and finite generated answers, not extra training facts. Output grids include out-of-range predictions. Same-attribute constrains provenance, not physical dimensions.

Q4. Validation of existing operators

In 32 known-input addition/subtraction checks with collisions and 64-512 bins, masses matched direct enumeration (maximum total variation 1.4e-17); representative error was within half the widest output bin.

Unary scaling inverse-maps output representatives to input coordinates, interpolates masses and renormalizes, approximating continuous-density change of variables.

Figure 4(A) tests binning resolution B: finer grids help Freebase, barely affect DB15K and do not outperform coarser YAGO15K grids. It measures numerical-filter performance; known-input checks isolate arithmetic approximation.

Thank you for providing the supplementary results, which alleviate some of my concerns.

However, the latest results still do not resolve the questions about the effectiveness of the learned RPK and the attribute-existence score. e.g., in Tables 1 and 2 of the response, the learned RPK improves over other settings by at most 0.43% MRR and performs worse than the simple neighbor copy/mean on DB15K. After replacing the attribute-existence scores, half of the datasets also show either a marginal improvement (+0.04%) or a negative improvement (−0.09%).

Considering the insignificance of these experiments, I tend to maintain my score.

Reviewer_gFz4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper investigates numerical complex query answering on knowledge graphs. The proposed method, CHOREOGRAPH, models numerical values as histograms via quantile bins and Gaussian smoothing, learns relation-specific convolution kernels, and applies message passing for missing value estimation. Additionally, it features a hybrid executor that fuses entity reasoning with histogram-based comparison, set, and arithmetic operations. Evaluations on four datasets indicate performance gains on most tested query structures and lower training costs compared to the evaluated baselines.

Critical strengths
  1. Explicit and traceable numerical execution. Histograms keep intermediate numerical states inspectable on attribute grids tied to the original units. Relation kernels make the direction and spread of transport visible. This gives a concrete operational benefit for examining composed numerical queries, beyond final answer accuracy.

  2. Empirical gains with modest training overhead. The method improves most tested ranking structures across several datasets, with supporting ablations and variability estimates. On FB15K, total training takes approximately 1,146 seconds, versus 3,826 seconds for LitCQD. The controlled graph splits and retraining of evaluated baselines strengthen the comparison. These results support a useful practical contribution.

Critical weaknesses
  1. The contribution of RPK is not fully isolated. Table 4 shows that message passing brings the largest improvement, while learned RPKs are compared mainly with uniform kernels. Comparisons with identity kernels, shared kernels, or direct neighbor aggregation would better establish the value of relation-conditioned convolution. Also, the same shift on a quantile grid can represent different numerical changes across regions, which deserves further analysis.

  2. Arithmetic evaluation is limited. Table 3 has no arithmetic baselines and uses only 100 queries per structure. Although Equation 6 supports operations between two distributions, the experiments decode later operands into scalars. The benefit of full distributional arithmetic therefore remains unclear, particularly against simpler scalar-based methods.

  3. Comparisons with recent methods are insufficient. The main experiments focus on LitCQD and NRN, both from 2023. More recent methods such as CNEQ (2024) and CNR-NST (2025) are discussed but not directly compared. Appendix C.4 explains the reproduction difficulties, but without these comparisons it is still hard to judge competitiveness against recent work. The paper also gives little discussion of LLM-based KG reasoning approaches.

Questions for authors
  1. Contribution of RPK. How would learned RPKs compare with identity kernels, shared kernels, and direct neighbor aggregation under the same setup?

  2. Arithmetic capability. Could the authors evaluate full distributional arithmetic under Equation 6 and compare its accuracy and cost with scalar decoding on the same queries?

  3. Comparison with recent methods. Could the authors better explain how CHOREOGRAPH compares with CNEQ and CNR-NST, beyond the reproduction issues noted in Appendix C.4?

  4. Comparison with LLM-based approaches. Why are LLM-based KG reasoning approaches not discussed or evaluated? How does CHOREOGRAPH differ in numerical accuracy, uncertainty handling, and computational cost? If direct comparison is not feasible, could the authors clarify the differences in task settings?

1 comment
AuthorsOpenReview ↗

Response to Reviewer gFz4

Q1. Contribution of RPK

With backbone and attribute-existence scores fixed, we evaluated Table 4's seven numerical-filter structures. Macro MRR (%), mean ± sample SD over three KG splits:

TransportFB15K-237FB15KDB15KYAGO15K
Learned RPK17.30 ± 1.4321.64 ± 0.2010.43 ± 0.815.79 ± 0.81
Identity17.09 ± 1.8421.21 ± 0.5210.43 ± 0.715.64 ± 0.88
Shared kernel16.83 ± 1.6321.07 ± 0.2510.41 ± 0.855.65 ± 0.99
Neighbor copy/mean16.75 ± 1.6020.98 ± 0.0310.52 ± 0.875.67 ± 1.06

Learned RPK exceeded all three controls on FB15K-237, FB15K and YAGO15K. The settings and results are included in Appendix D.6, Table 14.

Strong copy-based results are consistent with informative neighboring literals. RPK retains that evidence while learning relation-attribute-specific shifts and spread, rather than merely copying or applying the same transformation everywhere.

RPK shares an estimated distribution of source-target displacements within each relation-attribute pair in quantile coordinates. Sharing K coefficients across positions reduces estimation requirements under sparse path supervision; multiple shifts model spread rather than a single offset. Quantile bins allocate resolution where training literals concentrate. This is a learned statistical transport model, not a constant physical-unit shift. Table 4 compares this coordinate choice with uniform binning.

Q2. Arithmetic capability

On the submitted Table 3 queries with fixed input predictions, we compared set-scalar execution with Eq. (6), retaining both histograms. Three-split macro MRR (%):

ExecutionFB15KDB15KYAGO15K
Submitted set-scalar5.27 ± 0.134.18 ± 1.0310.14 ± 0.75
Both histograms retained5.86 ± 0.394.56 ± 0.9410.46 ± 0.81

On FB15K, 2b increases from 7.20 to 10.31 and 3b from 8.36 to 11.20; bp, pb and 2pb favor set-scalar execution. Retaining both distributions improves the overall mean on all three datasets and better matches CHOREOGRAPH's representation. We updated Table 3, Secs. 4.5/5.2 and Appendix B.4 accordingly.

Whole-query execution and terminal scoring decreased from 62.98 to 26.85 ms/query. The speedup arises because distribution-based execution applies arithmetic directly to bin representatives and probability masses, avoiding candidate-value scoring within arithmetic and scalar selection for later operands.

Q3. Comparison with recent methods

The distinction concerns numerical representation and execution, not arithmetic support alone. CNEQ combines magnitude-aware value encoding with learned query representations and predicted Gaussian-mixture parameters. Its learned query coordinates are not explicit numerical intervals. CNR-NST already maintains fuzzy candidate memberships and performs explicit arithmetic. It learns numerical comparison relations as value-node links, whereas CHOREOGRAPH computes comparison profiles directly on the ordered grid.

CHOREOGRAPH instead represents numerical answers as probability mass on attribute-specific ordered intervals. RPKs learn relation-conditioned transport on that same grid through independent ridge fitting and simplex projection. Intermediate support, modes and numerical transport can be inspected without decoding every state to a point or fitting an operator-specific neural network. This connects the representation, transport estimator and query algebra in one numerical coordinate system. Sec. 2 now makes these distinctions explicit.

Q4. LLM-based approaches

Our task requires estimating missing numerical facts and composing them. Retrieval-based KG RAG instead supplies retrieved facts to a generator; missing numerical evidence creates a task mismatch that can favor our setting. Nevertheless, we evaluated a SubgraphRAG adaptation with Llama-3.1-8B-Instruct on numerical-filter queries, retrieving only from the training KG.

On FB15K-237, three-split mean answer-set F1/Recall and matched-query costs were:

MethodF1 (%)Recall (%)ms/queryGPU peak (GiB)
CHOREOGRAPH5.6142.9821.810.24
SubgraphRAG0.720.4894,371.2517.91

SubgraphRAG does not output an exhaustive ranking, so both use answer-set metrics. Ratios before rounding are 7.82x/89.67x for F1/Recall, 4,326x for time and 74.6x for GPU peaks on RTX A6000. SubgraphRAG time covers generation alone. Table 5's 1.94 ms uses a different workload dominated by simple queries at batch 128; these numerical-filter timings use batch 1.

Comments on the paper

1 comment
Area_Chair_uVYrOpenReview ↗

Please Respond to the Authors’ Rebuttal

Dear Reviewers,

The Author–Reviewer Discussion period is now nearing its end and will close on 18 November. The authors have submitted their response and, in some cases, a revised manuscript.

Could you please take a moment to:

  • Read the authors’ rebuttal and any updated version of the paper;
  • Leave a reply to the authors on the points you raised, indicating whether your concerns have been addressed.

Even if your evaluation remains unchanged, a short reply confirming your final position is very helpful, as it allows me to track each reviewer’s stance and write an accurate meta-review.

If you are unable to participate further, please let me know as soon as possible.

Thank you for your time and careful reviewing.

Best regards,

Area Chair, ICLR 2027

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.