Organizations: School of Electronic and Computer Engineering, Peking University, Shenzhen, China · Harbin Institute of Technology, Shenzhen, China · School of Computer Science, Peking University, Beijing, China · Pengcheng Laboratory, Shenzhen, China
Regarded as the third generation of neural networks, Spiking Neural Networks (SNNs) have garnered significant traction due to their biological plausibility and energy efficiency. Recent advancements in large models necessitate spiking neurons capable of high performance, adaptability, and training efficiency. In this work, we first propose a novel functional perspective that provides general guidance for designing the new generation of spiking neurons. Following the insightful guidelines, we propose the Adaptive Spiking Neuron (ASN), which incorporates trainable parameters to learn membrane potential dynamics and enable adaptive firing. ASN adopts an integer training and spike inference paradigm, facilitating efficient SNN training. To further enhance robustness, we propose a specialized variant of ASN, the Normalized Adaptive Spiking Neuron (NASN), which integrates normalization to stabilize training. We evaluate our neuron model on 19 datasets spanning five distinct tasks in both vision and language modalities, demonstrating the effectiveness and versatility of the ASN family. Our ASN family is expected to become the new generation of general-purpose spiking neurons.
Figures & tables
Spiking Neurons
Efficient Training
Adaptive Firing
Architecture Compatibility
Spike-Driven Inference
LIF ( Gerstner et al., 2014 )
✗
✗
✓
✓
PLIF ( Fang et al., 2021 )
✗
✓
✓
✓
PSN ( Fang et al., 2023b )
❍
✓
✗
✓
ILIF ( Luo et al., 2024 )
✓
✗
✓
✓
NILIF ( Lei et al., 2025 )
✓
✗
✓
✓
ASN, NASN
✓
✓
✓
✓
Table 1. Comparison with other foundation spiking neurons in the SNN domain, including LIF ( Fang et al., 2022 ) , PLIF ( Fang et al., 2021 ) , PSN ( Fang et al., 2023b ) , ILIF ( Yao et al., 2023 ) , and NILIF ( Lei et al., 2025 ) . ❍ indicates that PSN is substantially faster than LIF and PLIF, while remaining notably slower than ILIF and NILIF in training. Our ASN family (including ASN and NASN) satisfies the four characteristics: Efficient Training, Adaptive Firing, Architecture Compatibility, Spike-Driven Inference.
Figure 1. Overview of Adaptive Spiking Neuron(ASN). (a) schematics of a biological spiking neuron. (b) depicts a neuron lacking membrane potential adaptation, thereby suppressing the transmission of membrane potential information. (c) illustrates the proposed neuron, where a learnable shift parameter enables adaptive firing, thereby ensuring better membrane potential information transmission. From the perspective of spike dynamics at the neural level. (d) and (e) show the dynamics of LIF (spike training and spike inference) and ILIF (integer training and spike inference) neurons, respectively. Both employ fixed-interval clamping, resulting in significant membrane potential quantization bias. (f) shows the spike dynamics of our proposed ASN with adaptive firing, achieving more precise spike quantization of membrane potential.
Methods
Architecture
Param (M)
Train Size
Test Size
Time Step
E (mJ)
Top-1 Acc ( % )
SEW ResNet ( Fang et al., 2022 )
SEW-ResNet-34
21.79
224 2
224 2
4
-
67.04
MS-ResNet ( Hu et al., 2021 )
MS-ResNet-34
21.80
224 2
224 2
6
-
69.42
Spikformer ( Zhou et al., 2023b )
Spikformer-8-384
16.81
224 2
224 2
4
12.43
70.24
SD-Transformer ( Yao et al., 2023 )
SD-Transformer-8-384
16.81
224 2
224 2
4
3.90
72.28
LIF (Spikingformer) ( Zhou et al., 2026b )
Spikingformer-8-384
16.81
224 2
224 2
4
4.69
72.45
LIF (Spikingformer † ) ( Zhou et al., 2026b )
Spikingformer-8-384
16.81
224 2
224 2
4
5.61
74.35
Table 2. The results on ImageNet-1k classification. "E (mJ)" means "Energy Consumption (mJ)". Spikingformer † is the variant of Spikingformer with the improved downsampling ( Zhou et al., 2023a ) .
Method
CIFAR-10
CIFAR-100
Param
T
Acc
Param
T
Acc
Spikformer ( Zhou et al., 2023b )
5.76
4
94.80
5.76
4
76.95
Spikformer ( Zhou et al., 2023b )
9.32
4
95.51
9.32
4
78.21
SD-Transformer ( Yao et al., 2023 )
10.28
4
95.60
10.28
4
78.4
Spikingformer ( Zhou et al., 2026b )
9.32
4
95.81
9.32
4
79.21
Spikingformer † ( Zhou et al., 2026b )
9.32
4
95.95
9.32
4
80.37
Table 3. Comparision on CIFAR-10 and CIFAR-100. "Param" denotes "Parameter (M)", "Acc" denotes "Top-1 Accuracy (%)", and " T " denotes "Time Step". Our model adopts the architecture of Spikingformer † .
Model
Time Step
MNLI
QQP
QNLI
SST-2
CoLA
STS-B
MRPC
RTE
Avg.
BERT base ( Devlin et al., 2019 )
–
83.4
71.2
90.5
93.5
52.1
85.8
88.9
66.4
79.6
Q2BERT ( Zhang et al., 2020 )
–
47.3
67.0
61.3
80.6
0.0
4.7
81.2
52.7
49.1
BERT 3L ( Devlin et al., 2019 )
–
77.1
85.2
85.8
88.1
31.7
85.7
86.4
66.4
75.9
SpikeLM ( Xing et al., 2024b )
4
77.2
83.9
85.3
87.0
38.8
84.9
85.7
69.0
76.5
LIF-BERT ( Xing et al., 2024b )
4
35.2
0
50.5
50.9
0
0
81.2
52.7
34.6
PSN-BERT ( Xing et al., 2024b )
4
35.2
0
50.5
50.9
0
6.8
81.2
52.7
34.7
Table 4. The results on the Natural Language Understanding task (GLUE datasets). "Avg." denotes "Average Accuracy (%)". The results of LIF-BERT and PSN-BERT are reported in Xing et al. (2024b) . Our model adopts the architecture of WE-SpikingFormer.
Model
E (mJ)
Time Step
ARC-e
ARC-c
BoolQ
HeadQA
OBQA
Avg.
Qwen-1.5B ( Yang et al., 2024 )
3398.3
-
26.2
26.9
60.3
25.7
27.1
33.2
NILIF (WD-SpikingFormer-0.4B) ( Zhou et al., 2026a )
238.4
4
30.0
22.3
37.8
26.0
26.1
28.4
Ours
245.1
4
35.3
20.6
37.9
25.9
27.2
29.4
Table 5. The results on Question Answering Tasks (QAT). "E (mJ)" means "Energy Consumption (mJ)". "Avg." denotes "Average Accuracy (%)". Our model adopts the architecture of WD-SpikingFormer-0.4B.
Model
E (mJ)
Time Step
HellaSwag
PIQA
WinoGrande
Avg.
SpikeLLM-7B ( Xing et al., 2024a )
-
4
33.9
53.4
51.5
46.3
Qwen-1.5B ( Yang et al., 2024 )
3398.3
-
26.5
53.7
52.8
44.3
NILIF (WD-SpikingFormer-0.4B) ( Zhou et al., 2026a )
238.4
4
25.9
53.4
50.2
43.2
Ours
245.1
4
26.0
54.6
49.4
43.3
Table 6. The results on Commonsense Reasoning Tasks (CRT). "E (mJ)" means "Energy Consumption (mJ)". "Avg." denotes "Average Accuracy (%)" . Our model adopts the architecture of WD-SpikingFormer-0.4B.
Figure 2. Our model (WD-SpikingFormer with NASN) on parameter scaling and pretraining data scaling. (a) and (b) show increasing model parameters for pretraining on Question-Answering Tasks (QAT) and Commonsense Reasoning Tasks (CRT). (c) and (d) show increasing tokens for model pretraining on the above two Tasks.
Neuron model
CIFAR-100 ( % )
LIF ( Zhou et al., 2026b )
80.22
PLIF ( Fang et al., 2021 )
80.25
PSN ( Fang et al., 2023b )
77.84
ILIF ( Luo et al., 2024 )
80.32
ASN (Ours)
80.36
NILIF ( Lei et al., 2025 )
80.98
Table 7. Ablation study for Spiking Neurons on CIFAR-100. Base model adopts the architecture of Spikingformer † ( Zhou et al., 2026b ) .
Figure 3. Neuron comparison on scaling. The base model is WD-SpikingFormer. (a) and (b) show the neuron comparison with the setting of 0.4B param with 0.1B Tokens on Question-Answering Tasks (QAT) and Commonsense Reasoning Tasks (CRT). (c) and (d) show the neuron comparison with the setting of 1.0B param with 0.5B Tokens on the above two Tasks.
Center for Semiconductor Technology, Korea Institute of Science and Technology (KIST), Seoul, Republic of Korea · Department of Computer Science and Engineering, Korea University, Seoul, Republic of Korea · LIG Defense&Aerospace, Seongnam, Republic of Korea +1