FloydNet: DP-Style Global Graph Learning

• FloydNet is a graph learning architecture that uses global DP-style iterative refinement to capture complex combinatorial and relational reasoning.
• It employs a dense all-pairs relationship tensor with a learnable operator inspired by the Floyd-Warshall algorithm to achieve high-order expressive power.
• Empirical results show state-of-the-art performance on benchmarks including CLRS-30, TSP, BREC, and molecule property prediction.

FloydNet is a graph learning architecture that realizes global, dynamic programming (DP)-style iterative refinement for combinatorial, algorithmic, and relational reasoning tasks. It departs from message-passing graph neural networks (MPNNs) by operating on a dense, all-pairs relationship representation, using a learnable operator inspired by the Floyd-Warshall algorithm. FloydNet attains higher-order expressive power, precisely implements generalized k-Folklore Weisfeiler-Lehman (k-FWL) color refinement, and achieves state-of-the-art empirical performance across a suite of challenging benchmarks, including CLRS-30, BREC, Traveling Salesman Problem (TSP), and molecule property prediction (Yu et al., 27 Jan 2026).

1. Global All-Pairs Representation

At the core of FloydNet is a dense, global "relationship tensor" maintained at every layer $l$:

$$\mathbf{R}^{(l)} \in \mathbb{R}^{N \times N \times d_r}$$

Here, $N$ is the number of graph nodes and $d_r$ is the hidden dimension. Each entry $\mathbf{R}^{(l)}_{i, k} \in \mathbb{R}^{d_r}$ encodes the current embedding of the relationship between nodes $i$ and $k$. The initialization $\mathbf{R}^{(0)}$ aggregates node features $\mathbf{X}_i \in \mathbb{R}^{d_n}$, edge features $\mathbf{E}_{i, k} \in \mathbb{R}^{d_e}$, and global features $$\mathbf{R}^{(l)} \in \mathbb{R}^{N \times N \times d_r}$$0 using a multi-layer perceptron:

$$\mathbf{R}^{(l)} \in \mathbb{R}^{N \times N \times d_r}$$1

Such a design enables direct modeling of long-range and high-order dependencies, in contrast to local aggregation schemes in standard GNNs.

2. FloydBlock: Learned DP-Style Refinement

Each FloydNet layer, termed a "FloydBlock," performs a global update of the relationship tensor using a learnable analogue of the Floyd-Warshall update:

$$\mathbf{R}^{(l)} \in \mathbb{R}^{N \times N \times d_r}$$2

In FloydNet, the scalar min and addition are replaced with high-dimensional attention-based operators ("Pivotal Attention"). Given $$\mathbf{R}^{(l)} \in \mathbb{R}^{N \times N \times d_r}$$3, normalized as $$\mathbf{R}^{(l)} \in \mathbb{R}^{N \times N \times d_r}$$4:

• Pairwise queries $$\mathbf{R}^{(l)} \in \mathbb{R}^{N \times N \times d_r}$$5, keys $$\mathbf{R}^{(l)} \in \mathbb{R}^{N \times N \times d_r}$$6 and $$\mathbf{R}^{(l)} \in \mathbb{R}^{N \times N \times d_r}$$7, and values are linearly projected for each path $$\mathbf{R}^{(l)} \in \mathbb{R}^{N \times N \times d_r}$$8.
• Keys and values along two-hop paths are combined (element-wise addition, default).
• Scaled dot-product attention over all pivot nodes $$\mathbf{R}^{(l)} \in \mathbb{R}^{N \times N \times d_r}$$9 computes:

$N$0

The full FloydBlock updates via residual connections and a feed-forward network:

$N$1

This global pattern can be interpreted as learning a task-specific relational calculus and enables long-range reasoning in $N$2 layers for exponential receptive field growth.

3. Expressive Power and Theoretical Properties

FloydNet with $N$3 directly implements the 2-Folklore WL (2-FWL) color refinement, which is equivalent in expressive power to the 3-WL test. Each pairwise embedding is refined by attending over the multiset of all length-2 paths $N$4, matching the combinatorial update rule of 3-WL (Yu et al., 27 Jan 2026).

In general, #1-FloydNet$N$5 can update each $N$6-tuple via attention over all pivots, effectively realizing the k-Folklore WL test. This positions FloydNet firmly within the k-FWL hierarchy and endows the architecture with expressive power strictly beyond 1-WL MPNNs and previously established message-passing GNNs. The achievable expressive power can thus be systematically increased via the tuple size, providing a theoretically principled mechanism for higher-order reasoning.

4. Model Architecture and Implementation

FloydNet is structured as a stack of $N$7 FloydBlocks, each utilizing the Pre-LN Transformer pattern:

• PivotalAttention: multi-head, head dimension $N$8.
• Feed-forward network (FFN): $N$9 with GeLU activation.
• Normalization: LayerNorm by default; BatchNorm, RMSNorm, and QK-Norm are also supported.
• Combine operator $d_r$0: additive (default), multiplicative (for geometric tasks).
• CUDA kernel: optimized to reduce memory from $d_r$1 to $d_r$2.

A prototypical iterative refinement pseudocode:

$i$7

5. Training Regimes and Hyperparameters

FloydNet is trained under domain-specific configurations:

• BREC (graph isomorphism): $d_r$3 layers, $d_r$4, single head, BatchNorm, FFN removed, float64, AdamW with $d_r$5, batch size 64, no positional encodings.
• CLRS-30 (algorithmic reasoning): up to $d_r$6, $d_r$7, 6 heads, AdamW with $d_r$8, linear warmup and cosine decay, up to 80k steps, tested OOD up to $d_r$9.
• TSP (combinatorial optimization): $\mathbf{R}^{(l)}_{i, k} \in \mathbb{R}^{d_r}$0, 6 heads, $\mathbf{R}^{(l)}_{i, k} \in \mathbb{R}^{d_r}$1, DDPM formulation (binary cross-entropy on edges), 400 epochs × 100 steps, 64 GPUs, batch size 1 (with accumulation), trained on $\mathbf{R}^{(l)}_{i, k} \in \mathbb{R}^{d_r}$2, tested on $\mathbf{R}^{(l)}_{i, k} \in \mathbb{R}^{d_r}$3, optimality filtered via Concorde.

6. Empirical Evaluation

FloydNet establishes strong or state-of-the-art empirical results across domains:

• Homomorphism Counting: Near-zero mean absolute error on all 8 tasks, surpassing GIN, Subgraph-GNN, 2-GNN, 2-FGNN (all $\mathbf{R}^{(l)}_{i, k} \in \mathbb{R}^{d_r}$4 2-WL).
• BREC: FloydNet (2-FWL/3-WL) accuracy 67.5% vs. 1-WL GNNs <$\mathbf{R}^{(l)}_{i, k} \in \mathbb{R}^{d_r}$5, Graphormer 19.8%, PPGT 58.5%, KP-GNN 68.8%; #1-FloydNet$\mathbf{R}^{(l)}_{i, k} \in \mathbb{R}^{d_r}$6 achieves 95.0%, #1-FloydNet$\mathbf{R}^{(l)}_{i, k} \in \mathbb{R}^{d_r}$7 99.8%, matching 4-WL.
• CLRS-30: Aggregated test accuracy:

Class	Triplet-GMPNN	RANR	G-ForgetNet	RT	ET	FloydNet
Sort (4)	75.6	94.2	78.1	50.0	82.3	100.0
Search (3)	58.8	82.9	63.8	65.3	63.0	91.6
Greedy (2)	76.4	83.5	91.8	85.3	81.7	93.2
DP (3)	82.0	42.7	86.7	83.2	83.5	90.0
Graph (12)	86.4	74.2	88.8	65.3	86.1	98.6
String (2)	49.1	49.1	54.7	32.5	54.8	99.7
Geometry (3)	94.1	88.4	95.1	84.6	88.2	99.7
Total (30)	80.0	75.8	82.9	66.2	80.1	96.6

• FloydNet maintains >90% accuracy up to $\mathbf{R}^{(l)}_{i, k} \in \mathbb{R}^{d_r}$8 in OOD settings (no-hint), while hint-augmented models often degrade or run out of memory.
• TSP ($\mathbf{R}^{(l)}_{i, k} \in \mathbb{R}^{d_r}$9): Linkern heuristic optimality 38.8% (general), 16.1% ($i$0); FloydNet (10 samples) 99.8% overall, 99.4% large $i$1.
• LRGB/ZINC: Competitive or superior on selected molecule and vision tasks (e.g., ZINC-full MAE 0.016).

7. Comparison with Message Passing Paradigms and Future Directions

MPNNs propagate information solely along local graph edges, causing over-squashing and constraining expressivity to 1-WL (sometimes 2-WL with extensions). FloydNet, by contrast, leverages a global $i$2 pairwise representation and DP-style refinement, resulting in exponentially fast receptive field growth ($i$3 paths in $i$4 layers) and provable 3-WL (2-FWL) expressiveness.

The major tradeoff is an increased computational and memory cost: FloydNet’s standard implementation incurs cubic complexity $i$5. This is mitigated for practical graph sizes ($i$6) with optimized kernels. Its strengths include exact emulation of dynamic programming routines (e.g., Floyd–Warshall), and superior suitability for tasks requiring long-range or combinatorial reasoning.

Current limitations include the cubic scaling and tailoring to moderate-size dense graphs. Open research areas include sparse pivot selection, approximate DP refinement, and multimodal extensions. FloydNet establishes learned DP refinement as a high-expressivity, empirically strong, and theoretically principled alternative to local message passing for global graph reasoning tasks (Yu et al., 27 Jan 2026).