ICML2025-MainTrack-Submission#280-AllReviewers

[magicparrots]

Table 1. Performance Comparison on Benchmark Datasets

Method	CIFAR-10 (Acc %)	CIFAR-100 (Acc %)	TinyImageNet (Acc %)	Params (M)	FLOPs (G)
ResNet-18	94.5	76.3	64.1	11.2	1.8
ViT-Small	95.2	77.9	65.7	21.7	4.6
Ours (GraphFormer)	96.1	79.5	67.3	19.8	3.9

---

Table 2. Ablation Study on Temporal Encoding

Method Variant	CIFAR-10 (Acc %)	TinyImageNet (Acc %)
Ours (No Time Encoding)	95.4	66.1
Ours (Sinusoidal Only)	95.8	66.8
Ours (Learnable Time)	96.1	67.3

---

Table 3. Robustness to Input Noise on CIFAR-10 (% Accuracy)

Method	No Noise	Gaussian (σ=0.1)	Gaussian (σ=0.3)	FGSM (ε=0.1)
ResNet-18	94.5	91.2	83.7	78.9
ViT-Small	95.2	92.4	85.1	80.3
Ours	96.1	93.5	87.0	83.7

---

Table 4. Generalization to Out-of-Distribution (OOD) Data

Method	In-Domain (CIFAR-10)	OOD (SVHN)	OOD (CIFAR-10-C)
ResNet-18	94.5	76.3	71.2
ViT-Small	95.2	78.1	73.0
Ours	96.1	80.5	75.3