Neural Algorithmic Reasoning for Hypergraphs with Looped Transformers
Zekai Huang, Yingyu Liang, Zhenmei Shi, Zhao Song, Zhen Zhuang

TL;DR
This paper extends Loop Transformers to hypergraphs, enabling neural algorithmic reasoning for complex higher-order structures through novel encoding schemes and theoretical guarantees.
Contribution
It introduces a hyperedge-aware encoding and a degradation mechanism to adapt Loop Transformers for hypergraph algorithms, bridging neural networks and combinatorial optimization.
Findings
Successfully simulated hypergraph algorithms like Helly's algorithm.
Provided theoretical guarantees for hypergraph algorithm simulation.
Demonstrated potential of Transformers as general-purpose structured data solvers.
Abstract
Looped Transformers have shown exceptional neural algorithmic reasoning capability in simulating traditional graph algorithms, but their application to more complex structures like hypergraphs remains underexplored. Hypergraphs generalize graphs by modeling higher-order relationships among multiple entities, enabling richer representations but introducing significant computational challenges. In this work, we extend the Loop Transformer architecture's neural algorithmic reasoning capability to simulate hypergraph algorithms, addressing the gap between neural networks and combinatorial optimization over hypergraphs. Specifically, we propose a novel degradation mechanism for reducing hypergraphs to graph representations, enabling the simulation of graph-based algorithms, such as Dijkstra's shortest path. Furthermore, we introduce a hyperedge-aware encoding scheme to simulate…
Peer Reviews
Decision·Submitted to ICLR 2026
- Neural algorithmic reasoning and hypergraphs are both interesting and important topics - The paper proposes valid next steps in this domain: It extends the results from Back de Luca et al. to port the graph algorithms to hypergraphs (which is simple). Beyond that, it presents one hyper graph-related algorithm as an example and describes the power of simulating deterministic hypergraph motif algorithms more generally. This is a decent contribution overall. - It's a good idea to cover the simula
- The paper is hard to read for readers not familiar with the domain. For instance, the paper jumps from the transformer definition in the preliminaries to Algorithm 2, without specifying the relationship (beyond the definition of a simulation). The introduction, related work, or preliminaries would be good places for giving a basic introduction. - More generally, technical details/definitions are often given without any intuition. Especially in a graph/algorithmic setting, pictures could provid
The theoretical development appears careful and rigorous, with clear statements of theorems and proofs. Establishing expressivity for hypergraph algorithms is a natural and timely question, and the results could be of interest to the algorithmic reasoning community.
The paper is entirely theoretical. While I appreciate the focus, even a small empirical demonstration would strengthen the case that the abstractions matter in practice (e.g., whether a looped Transformer can learn to execute hypergraph Dijkstra on synthetic tasks; sample efficiency; robustness to noise; comparison to non-looped baselines). In the current form, the bar for acceptance rests solely on theoretical novelty and significance. Many variations of BFS/DFS/Dijkstra are already known to b
1) The paper is well written and easy to follow despite the heavy theoretical content. 2) The mathematical formalization is precise and consistent. 3) The logical flow is clear: the reader can easily understand the motivation and the main theoretical results.
**Novelty** The paper represents a structured extension of an existing theoretical framework [1]. While the move from graphs to hypergraphs is non-trivial and technically well executed, the conceptual novelty is limited. Can you pleas clarify more explicitly how this work positions itself with respect to that prior paper? **Experimental validation** I understand that this paper is heavily theoretical, and experiments may not be strictly required. However, I have a question rather than a critic
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsGraph Theory and Algorithms · Data Visualization and Analytics · Distributed and Parallel Computing Systems
MethodsAttention Is All You Need · Adam · Softmax · Absolute Position Encodings · Residual Connection · Dropout · Byte Pair Encoding · Linear Layer · Multi-Head Attention · Position-Wise Feed-Forward Layer
