acceptodds
Under review as a conference paper at ICLR 2027

Lotus: A Logic-based Object-centric Topological Unified Solver

Abstract

In bipartite graph constraint satisfaction problems such as channel error-correction decoding, existing graph neural networks (GNNs) and belief propagation (BP) algorithms suffer from severe error floors caused by trapping set deadlocks induced by overlapping short cycles. To address this bottleneck, we propose Lotus (Logic-based Object-centric Topological Unified Solver), a neuro-algebraic framework that integrates deep unfolded networks with local discrete algebraic elimination across a Base GNN, an Expert GNN, and an algebraic elimination module termed xLoop. Lotus is structured around three core mechanisms: (i) an isotropic Variable Module and parameter-free Check Module devoid of positional encodings, avoiding graph-specific eigenspectrum overfitting to preserve permutation equivariance and enable zero-shot length and topology generalization; (ii) closed-form forward analytical gradients derived from a soft parity potential function, paired with Targeted Constraint Relaxation (TCR) and dual dynamic gating to directly steer forward trajectories without test-time optimization; and (iii) the xLoop module, which pairs weighted Laplacian spectral bisection with streaming Gaussian elimination (SGE) and Local Maximum Likelihood (Local ML) selection to algebraically resolve long-tail trapping sets within size-capped candidate subgraphs. Evaluated on 3GPP 5G New Radio (NR) and regular quasi-cyclic LDPC (QC-LDPC) matrices, Lotus achieves a 75.0%–83.9% hard-instance recovery rate, suppresses suboptimal convergence (SC) by approximately 50%, approaches the near-ML empirical performance bound of ordered statistics decoding (OSD-3) (with BP+xLoop also achieving near-ML performance), and preserves an asymptotic linear latency scaling of O(N) that operates 15x to 41x faster than OSD-3.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Representation Learning, Neural-Symbolic Reasoning, Constraint Satisfaction Problems, Channel Decoding, Deep Unfolding, Permutation Equivariance

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_xVWnOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper proposes a neural solver for decoding parity-check codes. The solver combines two main components: a message-passing network that acts on relaxed Boolean variables, and a more expensive Gaussian elimination component that handles deadlocks. The message-passing network is used to resolve most of the constraints, while elimination is applied only to the variables it fails to resolve. The approach is validated against belief propagation and algebraic elimination methods.

Critical strengths
  • The approach appears well-motivated, combining the scalability of continuous neural relaxations with the robustness of a more classical decoding method.

  • Allowing the xLoop component to be used independently of the proposed neural solver makes it more widely applicable to existing systems.

Critical weaknesses
  • The introduction could be tightened. The related work paragraph seems redundant, as it largely repeats work already cited earlier. I also found the 'core contributions' section too dense. It might be better to give a more concise summary of the model there, leaving the more technical terminology to subsequent sections.

  • Section 3 was difficult to follow. The latent state update described in Section 3.2 makes forward references to variables defined in subsequent sections without sufficiently explaining them when they first appear. More generally, the section describes how the model works, but does not adequately explain why it is designed this way, or the trade-offs between alternative designs.

  • The method is compared only against non-ML baselines. Experiments should ideally compare the solver's first two stages against some of the graph neural decoding methods described in the introduction to isolate the impact of this particular neural architecture.

  • Minor: it would be beneficial to include the latency scaling curves in the main paper, since Lotus's speedup is one of the headline results reported in the abstract.

Questions for authors
  • Have you considered alternative methods of aggregating node messages, such as attention or degree-normalized aggregation? How might this affect the model's performance?

  • How expensive is the xLoop stage of the solver compared to its first two stages?

Reviewer_b41AOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The proposed method, Lotus, is a two-stage GNN decoder for short LDPC codes, followed by a solver termed xLoop. xLoop solves a GF(2) system on a small set of unreliable bits and keeps the most likely codeword. The paper claims near-SOTA results on LDPC codes, with zero-shot transfer to longer codes with different topologies.

Critical strengths

Message-passing decoders like BP are fast and handle most noisy frames well. However, they get stuck on a few frames where a small cluster of wrong bits keeps reinforcing each other in a loop.

On the other hand, exact algebra (Gaussian elimination) can break such a loop, but doing it over the whole code is slow.

Lotus splits the work accordingly between the learning part and the algebraic part. The fast iterative part does the bulk of the cleanup using GNNs and, when it gets stuck, points to where the trouble is: the few checks still failing. The algebra part then works only on that small neighborhood, so its cost stays fixed no matter how long the code is.

I find this to be well motivated and elegant, and also (since I come from a specific background) novel. However, this neural+algebraic framework is not a new idea in the literature. For example, the papers below combine neural decoders with OSD to decode LDPC codes: Rosseel, Joachim, et al. "Decoding short LDPC codes via BP-RNN diversity and reliability-based post-processing." IEEE Transactions on Communications 70.12 (2022): 7830-7842. Rosseel, Joachim, et al. "Sets of complementary LLRs to improve OSD post-processing of BP decoding." 2023 12th International Symposium on Topics in Coding (ISTC). IEEE, 2023.

Critical weaknesses

I noticed some issues with the reported literature results and the paper's tables, so I ran Opus 5.5 medium with the following prompt: "I think I see some gaps in the experimental results between the paper and the literature and maybe even the various tables of the paper. Please survey all paper results and make sure they are consistent, match the literature, and that references to results match published results."

The results were an extremely extensive list of issues. These are provided in the AI disclosure section. I'll just give the section titles here:

A. Tables contradict each other: the same matrix and SNR get different FERs in Tables 1, 2, 4, and 5. Among other issues, there is a column copy error. B. Headline numbers match no table. 10.09× FER reduction over BP-50 at 4.0 dB on H96A is maybe half as much. C. Their own tables undercut the claims. BP+OSD-3 (Table 6) often beats Lotus. D. Literature comparisons are unsupported ( ECCT and W-BP do not contain the results that are said to be "compiled from published evaluations" for these citations) E. Complexity and latency accounting is understated. F. Several references are wrong. Veličković, Petar, and Charles Blundell. "Neural algorithmic reasoning." Patterns 2.7 (2021), for example, is cited as Petar Veličković, Rex Ying, Matilde Padovano, Raia Hadsell, and Charles Blundell. Neural algorithmic reasoning. Patterns, 2(7):100273, 2021.

Behind these 6 sections, there are tens of issues that undermine the validity of the results. A,B is critical since the results cannot be trusted. C considerably reduces the claims.

Questions for authors

Please run a thorough check for inconsistencies between the tables, for the validity of cited results (whether they appear in the original manuscripts)

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

The paper introduces LOTUS, a hybrid neural-algebraic framework for LDPC decoding, with the objective to overcome trapping sets while maintaining low computation complexity. It combines two GNNs, uses analytical gradients to help enforcing the constraints, as well as targeted constraint relaxation and dynamic gating, resorting to a local algebraic solver for remaining difficult cases. It is benchmarked against classical and neural decoders, claiming near-ML performance while showing no latency and zero-shot generalization across code lengths and topologies.

Critical strengths
  1. The framework seems carefully designed, with several adaptive mechanisms to handle conflicting constraints and difficult cases. Specifically, the combination of analytical gradients, targeted constraint relaxation and dynamic gating provides a flexible and seemingly robust manner to guide the decoding process without excessively disrupting the GNN's estimates and behavior.

  2. The experimental results are encouraging, with strong performance compared to classical and neural decoding baselines.

Critical weaknesses
  1. The ML novelty seems relatively limited. Analytical gradients and constraint relaxation (as well as dynamic gating) are closely related to existing methods, particularly gradient-based guidance for iterative processes, which is commonly used for conditional diffusion models for instance.

  2. The contribution of the neural components is unclear, as BP + xLoop already achieves very strong results, sometimes outperforming LOTUS. A large part of the improvement may come from the algebraic solver rather than the proposed GNN architecture.

  3. The architecture seems relatively complex, with two separately trained GNNs and several additional mechanisms. While some of them appear essential, the Expert GNN, with its specialized training, raises questions about generalization to unseen trapping sets and topologies. It is also unclear whether this complexity is justified compared to simpler approaches, such as a single GNN with more iterations.

Questions for authors
  1. Could the authors better clarify the methodological novelty of LOTUS beyond combining existing neural and algebraic techniques for LDPC decoding? Identifying one or two particular methodological ideas of interest that define the point the paper is making would make it more impactful.

  2. Could the authors better isolate the benefits of analytical gradient guidance, TCR and dynamic gating compared to simpler alternatives?

  3. Given the strong performance of BP + xLoop, what are the main advantages of the proposed neural architecture over this simpler approach? Could the authors compare the Base GNN + Expert GNN approach with a single GNN using more iterations?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.