- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper presents a framework for exploring large Cayley and Schreier graphs through scalable exact computation and learning-guided search. The authors report substantial computational speedups for exact graph-growth calculations and show that learned heuristics can improve search performance, with Bellman refinement providing additional gains. The framework is used to generate and test a large catalogue of conjectures on graph diameters, growth statistics, antipodes, and extremal generator families. Notable mathematical outputs include a fixed-conjugacy-type quasi-polynomiality conjecture, a proposed Babai-type diameter bound, and conjectures related to the Glushkov and unitriangular-group settings. Overall, this is an interesting AI4Math contribution that combines scalable exact computation with learning-guided exploration and mathematically verifiable outputs.
Critical strengths
1. The combination of scalable algebraic computation with learned search and verifiable certificates is a useful AI4Math design principle.
2. The idea seems novel, and the software contribution could be valuable research infrastructure if the reported scalability and correctness are validated. Moreover, several structural conjectures appear mathematically interesting and could motivate follow-up work beyond the immediate benchmarks.
3. The paper is appropriately cautious in several places about distinguishing search outputs from proofs and about the unresolved question of whether learning actually increases the rate of discovery.
Critical weaknesses
1. Correctness, internal consistency and explainability
I found enough concrete inconsistencies that my confidence in some parts is reduced. Examples include:
-
Conflicting CPU benchmark tables: Table 11 and a later unlabeled table report substantially different timings for the same families and sizes.
-
A table on page 27 is labeled as undirected but reproduces the directed diameter values.
-
The generalized symmetric-group section contains apparent algebraic/notation errors, including the codomain of parity maps, an “even n” versus “even m” inconsistency, and a mismatched group relation in the proof.
-
The discussion of Coxeter generators and braid groups is inaccurate as written.
-
The odd-n antipode discussion contradicts its own n=5 example.
-
The LARX+inverse odd-n formula has inconsistent denominators between the catalogue and detailed section.
-
The LSL section should clarify whether it is treating a directed graph, an inverse-closed graph, or a symmetrized operator, especially in the spectral discussion.
-
There are also several instances of visibly unfinished text, such as “Current work:” and “We were able to achieve ...”. More generally, the manuscript would benefit from careful proofreading to remove incomplete sentences, grammatical issues, and other presentation inconsistencies.
I should emphasize that these are only some of the issues related to inconsistencies in the manuscript. Therefore the authors should review the manuscript carefully and rigorously to ensure that no inconsistencies remain. Many of these issues may be individually repairable, but collectively they raise concerns about the rigor of the study. I therefore strongly encourage the authors to conduct a systematic check across the manuscript.
Beyond correctness and internal consistency, I think the paper would benefit from a stronger emphasis on explainability and mathematical intuition. A particularly valuable aspect of AI4Math systems is not only that they search efficiently or generate candidate conjectures, but that they help expose new structure that can guide human understanding. In the current manuscript, the outputs are often verifiable and some appendices do provide useful mathematical interpretation, but for several headline discoveries it remains unclear what genuinely new intuition the learning component reveals. I would encourage the authors to explore this more explicitly—for example, by identifying structural or algebraic features associated with difficult/extremal states, explaining why particular scaling or residue-class patterns emerge, or extracting interpretable principles from the learned search. Incorporating such insights more directly into the main presentation would strengthen both the mathematical contribution and the broader AI4Math significance of the work.
2. Novelty and positioning
The paper would benefit from distinguishing more clearly between its three sources of novelty: the scalable graph calculations, the learning-guided search procedure, and the resulting mathematical discoveries. In my view, the main novelty is mostly on the integration of learned search with scalable exact algebraic computation and independently verifiable mathematical evidence (not on particularly new RL method for example). Clarifying which mathematical discoveries specifically depended on the learning component would strengthen the paper's AI4Math positioning.
Some mathematical novelty statements should also be calibrated more carefully. For example, the Babai-type discussion should distinguish conjectural polynomial behavior from proved general diameter bounds, and several literature attributions should be tightened. More broadly, it should be made more explicit which of the conjectures actually depended on learning.
I also noticed several issues with the existing literature attribution that would benefit from correction or further clarification. Some illustrative examples include, but are not limited to:
-
“In line with the literature, our computations confirm the exact formula .” It would be helpful to provide the specific literature reference supporting this exact formula.
-
“For nilpotent groups, we formulate conjectural improvements to results of J. S. Ellenberg on upper unitriangular groups over ...” It would be helpful to include the relevant citation to the work of J. S. Ellenberg being referred to here.
-
“We also conjecture an answer to a question posed by V. M. Glushkov in 1968 on directed Cayley graphs generated by a cyclic shift and a transposition.” It would be helpful to provide the corresponding citation for the original Glushkov result/question.
-
The Babai-type framing would benefit from more clearly distinguishing proved and conjectural bounds. Babai–Seress and Helfgott–Seress provide respectively subexponential and quasipolynomial general upper bounds, while a universal quadratic bound for arbitrary generators remains conjectural. Thus, may be more accurately presented as a sharpening of a conjectural prediction rather than as an improvement in the constant of an established theorem.
Beyond correcting and clarifying the existing references, I believe the related-work discussion could benefit from broader coverage of learning-assisted mathematical discovery and automated conjecturing in graph/combinatorial settings. Relevant examples include:
-
Adam Zsolt Wagner (2021), “Constructions in Combinatorics via Neural Networks.” Uses reinforcement learning to search for constructions and counterexamples to open problems in extremal combinatorics and graph theory.
-
Randy Davila (2024), “Automated Conjecturing in Mathematics with TxGraffiti.” Develops an automated conjecturing system based largely on relationships between graph invariants.
-
Ján Pastorek (2026), “AutoGraphForge: Towards Automated Graph Theory Discovery.” Develops a graph-theoretic pipeline combining conjecture generation, counterexample search, and formal verification.
-
Hakan Akgün et al. (2026), “KnottedGraph: Scalable Knotted-Graph Topology for Scientific and Mathematical Discovery.” Combines scalable computation of graph-topological invariants with LLM-assisted discovery of closed-form mathematical relations.
-
Giorgi Butbaia et al. (2026), “Hierarchical Reinforcement Learning for Sparse-Reward Search in Commutative Algebra,” ICML 2026. Casts the search for counterexamples to mathematical conjectures as sparse-reward reinforcement learning on graph-structured search spaces.
These works, among other related studies that the authors may consider, would help place the present contribution more clearly within the broader literature on learning-assisted and automated mathematical discovery.
Questions for authors
1. Which mathematical discoveries genuinely required learned search rather than exact BFS and formula fitting alone?
2. What explainability or interpretability analyses have the authors explored to understand what the learned model is capturing mathematically (e.g., probing correlations with algebraic invariants such as cycle structure or parity, analyzing which generator choices drive the score, or extracting symbolic or structural rules from highly ranked states)?
3. How automated is the conjecture-generation stage itself? In particular, which parts of the process—identifying patterns, proposing functional forms, selecting candidate extremal states, and formulating conjectures—are performed automatically, and which rely on manual mathematical inspection?