acceptodds
Under review as a conference paper at ICLR 2027

Toward Accurate Few-Timestep ANN to SNN Conversion of LLMs through Learned Rotations

Abstract

ANN-to-SNN conversion offers a practical route to transferring pretrained large language models (LLMs) to spiking neural networks, but preserving model accuracy with only a few timesteps remains challenging. Multi-threshold neurons (MTNs) have enabled nearly lossless conversion of smaller-scale models with only a few timesteps. However, directly applying MTNs to large language models at similarly low timestep budgets leads to substantial accuracy degradation. We identify a mismatch between pretrained activation distributions and the nonuniform output levels of MTNs, with pronounced conversion errors concentrated in channels containing activation outliers. We propose a task-driven rotation calibration framework that adapts activation distributions to MTN conversion. The framework learns orthogonal rotations through the converted model's next-token loss, redistributing activations across channels to substantially mitigate the impact of activation outliers on MTN conversion. To enable efficient and accurate calibration, we introduce the equivalent parallel multi-threshold neuron (EPMTN). EPMTN requires only a single forward and backward pass, reducing calibration cost while mitigating the accumulation of surrogate-gradient approximation errors across timesteps. Experiments on LLaMA2-7B and LLaMA3-8B across four language-modeling corpora and four reasoning benchmarks demonstrate SOTA conversion performance with only four timesteps. These results highlight the potential of learned rotations and EPMTN calibration as a path toward accurate, low-timestep spiking LLMs.

Primary area:
applications to neuroscience & cognitive science
Keywords:
ANN-to-SNN Conversion, Spiking Neural Networks, Multi-threshold Neurons, Orthogonal Rotation, Large Language Models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_w99LOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This work integrates a multi-threshold spiking model with a rotation calibration strategy to propose an ANN-SNN conversion framework tailored for Large Language Models (LLMs). The proposed approach is empirically validated on benchmarks related to language modeling and reasoning.

Critical strengths
  1. Building upon the conventional multi-threshold spiking model, the authors have implemented relevant optimizations regarding parallelization and surrogate gradient consistency computation.
Critical weaknesses
  1. Lack of substantial academic contribution to the SNN domain: Although the paper superficially appears to focus on ANN-SNN conversion learning, the proposed EPMTN model is essentially equivalent to applying bit-width compression directly to the activation layer. In essence, the overall methodology reduces to a straightforward application of established post-training quantization (PTQ) algorithms—typically developed for LLMs—solely to the activation layers, lacking genuine novelty in SNN research.

  2. Failure to address core architectural challenges: Regarding the deployment of SNNs on Transformer architectures, this work fails to propose a genuine solution to effectively circumvent the floating-point multiplications inherent in operations such as LayerNorm and Softmax. For instance, it does not explore replacing these with BatchNorm and spiking neurons, or other SNN-friendly computational modules, as demonstrated in prior works like Spiking Vision Transformers. Consequently, the converted spiking LLMs exhibit no substantial architectural or computational differences from standard low-bit quantized LLMs.

Questions for authors

Baseline Comparisons: Regarding the experimental evaluation, the primary baselines for comparison should be quantized LLMs rather than full-precision ANNs. The authors must further substantiate whether the spiking LLMs derived from conversion learning offer any genuine advantages over standard quantized LLMs in terms of both performance and power consumption (energy efficiency).

4 comments

Weakness 1

We thank the reviewer for highlighting the connection between EPMTN and activation quantization. However, EPMTN is not constructed by directly imposing an arbitrary bit-width constraint on activations. It is derived from the adopted MTN dynamics as a parallel formulation that is exactly equivalent to the time-averaged output. This equivalence addresses end-to-end MTN calibration by preserving the original forward response while avoiding temporal unrolling and directly defining a surrogate gradient for the complete response. The controlled experiments in Table 1 distinguish forward equivalence from differences in surrogate gradients. Table 2 shows that this design reduces calibration time and memory consumption while improving final perplexity. We therefore view the contribution as extending beyond discrete activation representation. It establishes an equivalent calibration formulation for multi-timestep spiking dynamics and validates its training efficiency and optimization benefits. We acknowledge and appropriately cite the prior quantization research underlying rotation calibration. We do not regard differences in numerical representation or execution alone as evidence of superiority over quantized LLMs. Performance and energy advantages relative to standard quantized models must be assessed through direct experiments.

Question 1

We thank the reviewer for the suggestion. We add a direct comparison between quantized LLMs and quantized EPMTN using the same base model and evaluation data within each model. We compare the FP16 ANN baseline and W4 configurations in terms of perplexity and estimated energy per token.

Supplementary Table 5: QuaRot and EPMTN on Llama3-8B and Qwen3-8B

ModelMethodWT2 PPL ↓Pile PPL ↓Estimated energy (mJ/token) ↓Energy relative to ANN ↓
Llama3-8BANN FP166.13585.353535.1503100%
Llama3-8BQuaRot W4A4KV48.20536.83855.210414.82%
Llama3-8BOurs W47.67626.84445.297015.07%
Qwen3-8BANN FP169.71516.336435.5229100%
Qwen3-8BQuaRot W4A4KV411.31837.33055.771116.25%
Qwen3-8BOurs W410.18016.99235.747916.18%

On Llama3-8B, our method achieves lower WikiText-2 PPL than QuaRot, while Pile PPL is close but slightly higher. Its estimated energy subtotal is approximately 1.66% higher. On Qwen3-8B, our method achieves lower PPL on both datasets, with an estimated energy subtotal approximately 0.40% lower. Relative to FP16 ANN, our method reduces estimated energy by 84.93% on Llama3-8B and 83.82% on Qwen3-8B, trading an increase in PPL for a substantial reduction in cost. Relative to QuaRot, an ANN quantization method, our method reduces WikiText-2 PPL by 6.45% on Llama3-8B while maintaining nearly identical Pile PPL. On Qwen3-8B, it reduces WikiText-2 and Pile PPL by 10.06% and 4.61%, respectively. These results show that our method retains effective conversion accuracy after weight quantization and improves selected performance metrics at estimated costs comparable to those of an established ANN quantization method.

Weakness 2

We thank the reviewer for raising the issue of floating-point operations in Transformer nonlinear modules. Our method focuses on conversion accuracy with few timesteps through a framework based on learnable rotations and MTNs. This framework is independent of, and can be combined with, neuron-based approximations of nonlinear modules. First, these nonlinear modules account for a small fraction of the operation count in the evaluated configuration. We perform an operator-level analysis of a complete dense forward pass of Llama2-7B with batch size 1 and sequence length 2048. The analysis covers weight-activation products, attention matrix multiplications, nonlinear function evaluation, and operations associated with rotations, encoding, gating, and residual connections. RMSNorm, Softmax, and SiLU together account for 0.0730% of the operations and special-function calls in the original EPMTN implementation, compared with 0.0759% in ANN. Most computation remains concentrated in linear projections and attention matrix multiplications, which are therefore the primary focus of our framework.

Supplementary Table W2-a: operations and special-function calls (millions per token)

Operation categoryANNOriginal EPMTN
Linear projections and LM head13214.154813214.1548
Attention QK and AV1073.74181073.7418
RMSNorm1.06510.7988
Softmax8.38768.3876
SiLU1.40901.4090
Rotations, encoding, and other operations5.6095209.4881
Total14304.367714507.9802
Combined share of the three nonlinear modules0.0759%0.0730%

Second, retaining floating-point function evaluation does not imply that our accuracy gains require exact evaluation of these functions. To test this directly, we fix T=4, the original rotations learned through CE calibration, FP16 weights, and all Q/A/SiLU encoding locations. We then use populations of threshold neurons to approximate the square and reciprocal square root in RMSNorm, the exponential and reciprocal in Softmax, and the scalar SiLU function. We also evaluate replacing all of these functions simultaneously.

A neuron population approximates a scalar function using fixed thresholds and weighted spike outputs:

F^(x)=v0+∑i=1KwiH(x−bi).\widehat{F}(x)=v_0+\sum_{i=1}^{K}w_i H(x-b_i).

Here, K is the number of additional spatial thresholds per scalar function, and H emits a spike when its input strictly exceeds the threshold. Thresholds and output weights are constructed offline using predefined grids, without fitting or selection on the test set. An additional threshold is used in the negative tail of the exponential to preserve zero outputs for causally masked positions.

Supplementary Table W2-b: Neuron-based approximation of nonlinear functions for Llama2-7B at T=4

ConfigurationWT2 PPL ↓Pile PPL ↓
Original method with nonlinear function evaluation6.20865.6993
Approximate only the square and reciprocal square root in RMSNorm6.21355.7140
Approximate only the exponential and reciprocal in Softmax6.20545.6931
Approximate only SiLU6.21215.7002
Replace all of the above functions6.28145.8165

When all functions are replaced simultaneously, WikiText-2 and Pile PPL are 6.2814 and 5.8165, corresponding to relative increases of 1.17% and 2.06%. These results show that, for the evaluated model and settings, the performance of the original rotation-calibrated method is largely preserved when nonlinear functions are approximated by neurons.

For temporal execution, we draw on the nonlinear state accumulation and output compensation mechanism in DCGS [1]. Its differential graded units still evaluate the original function F. For example, Equations (15) and (16) in Section 4.1.1 of the published version are:

m[t]=m[t−1]+xin[t]t,xout[t]=t(F(m[t])−F(m[t−1])).m[t]=m[t-1]+\frac{x_{\mathrm{in}}[t]}{t},\qquad x_{\mathrm{out}}[t]=t\bigl(F(m[t])-F(m[t-1])\bigr).

This mechanism specifies how nonlinear modules are incorporated into temporal computation.

[1] Differential Coding for Training-Free ANN-to-SNN Conversion, Huang et al., ICML 2025.

Thank you again for your time and thoughtful feedback. Your comments prompted us to clarify the contribution of EPMTN beyond activation quantization and provide additional evidence addressing your concerns. Specifically, our rebuttal includes direct comparisons with QuaRot in terms of perplexity and estimated energy consumption, as well as experiments evaluating neuron-based approximations of Transformer nonlinear functions.

We would greatly appreciate hearing whether these clarifications and additional results address your concerns, or whether any questions remain that we could help clarify. If you find that our responses resolve some of the issues raised, we would be grateful if you could reconsider your overall assessment in light of the additional evidence.

I acknowledge the authors' response to the reviews. After thoroughly evaluating the rebuttal and the perspectives shared by other reviewers, I will maintain my original recommendation.

Reviewer_xg1fOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Multi-threshold neurons are a promising technique for LLM-to-SNN conversion; however, for the abundant outliers in LLM activations, the wide intervals between thresholds naturally lead to degraded activation precision. This paper proposes applying orthogonal rotations to SNN-converted LLMs to make the activation distribution more uniform and flatten out outliers, ensuring that most activations lie within a similar range of the thresholds to enhance precision. Rather than calibrating layer-wise, it improves prediction accuracy by learning rotation matrices optimized against the full network’s next-token prediction loss, maintaining orthogonality via Cayley updates. Furthermore, it introduces a novel Equivalent Parallel Multi-Threshold Neuron (EPMTN) to parallelize multi-threshold neuron training, making it both efficient and accurate by preventing surrogate-gradient error accumulation. Evaluations on language modeling and reasoning tasks across two converted LLMs demonstrate that the proposed method retains accuracy on par with original ANNs while outperforming prior LLM-to-SNN conversion techniques.

Critical strengths

By introducing EPMTN, the paper effectively overcomes the major drawbacks of MTNs, which is time-sequential training overhead and surrogate-gradient error accumulation, through parallelization.

Compared to existing SNN-based LLMs, the proposed method achieves a noticeable accuracy margin across various language modeling and reasoning datasets, reaching accuracy competitive with FP16 ANNs.

Critical weaknesses

The core technical novelty is relatively limited. Most key techniques, such as rotation-based activation smoothing [1,2], next-token cross-entropy loss [3], and Cayley updates [2], are directly adapted from existing LLM quantization literature to MTNs. Moreover, EPMTN serves as an orthogonal optimization technique that improves the training efficiency and accuracy of MTNs, rather than a fundamental enabling technique for the proposed framework.

The proposed techniques worsen the inherent neuromorphic infeasibility of MTNs. Specifically, applying rotations destroys activation sparsity, which in turn increases the spike transmission routing traffic and memory bandwidth overhead in neuromorphic chips.

[1] Ashkboos, S., Mohtashami, A., Croci, M. L., Li, B., Cameron, P., Jaggi, M., ... & Hensman, J. (2024). Quarot: Outlier-free 4-bit inference in rotated llms. Advances in Neural Information Processing Systems, 37, 100213-100240.

[2] Liu, Z., Zhao, C., Fedorov, I., Soran, B., Choudhary, D., Krishnamoorthi, R., ... & Blankevoort, T. (2025, May). Spinquant: Llm quantization with learned rotations. In International Conference on Learning Representations (Vol. 2025, pp. 92009-92032).

[3] Shao, W., Chen, M., Zhang, Z., Xu, P., Zhao, L., Li, Z., ... & Luo, P. (2024, May). Omniquant: Omnidirectionally calibrated quantization for large language models. In International Conference on Learning Representations (Vol. 2024, pp. 45472-45496).

Questions for authors

Is the proposed rotation approach applicable to standard LIF neurons rather than MTNs? Since the core benefit of the learned rotation is making activation distributions more uniform and suppressing outliers, couldn't a single-threshold, unsigned LIF neuron also achieve sufficient effectiveness under this framework?

When applying the proposed technique to actual MTN-based SNN hardware (if exists), wouldn't the increased activation density (due to rotation smoothing out outliers) lead to a noticeable drop in hardware efficiency and energy savings?

Many neuromorphic hardware chips support only low-bit integer weight precision. If the rotated weight matrices are quantized to low-bit integers, can they still preserve their intended rotation (and inverse rotation) functionality without significant precision degradation?

How does energy consumption specifically change before and after applying the proposed techniques (especially the learned rotations)? Table 6 currently reports only the relative energy usage after applying the proposed methods compared to the FP16 ANN baseline, lacking a direct comparison with the unrotated SNN counterpart.

5 comments

Weakness 1

We acknowledge that rotations, the cross-entropy calibration objective, and Cayley updates build on prior quantization research, which we cite as the foundations of our approach. Our contribution is to establish a conversion framework compatible with rotations and, within this framework, develop and validate task-driven calibration that addresses the mismatch between pretrained activations and the nonuniform responses of MTNs under a finite timestep budget. The framework jointly addresses inference implementation and calibration feasibility. Learnable rotations are folded into the weights, while online transformations use structured Hadamard operations. To address the temporal unrolling cost and surrogate-gradient propagation challenges of end-to-end MTN calibration, we further derive EPMTN as a forward-equivalent formulation. EPMTN is therefore directly connected to rotation calibration, rather than serving only as an independent efficiency improvement. Our ablations demonstrate both the accuracy gains from learned rotations and the improvements in calibration resource requirements and final performance enabled by EPMTN.

Weakness 2

We thank the reviewer for raising this important concern. Rotations can increase the firing rate, but they can also improve conversion accuracy and reduce the number of timesteps required to achieve comparable performance. Our method targets accurate conversion with few timesteps. Evaluating its hardware implications therefore requires considering not only firing rates at a fixed timestep budget, but also the required number of timesteps and the total number of spike events at comparable accuracy. Increased activity per timestep may be offset by a reduction in the number of timesteps.

To examine this tradeoff, we add two direct comparisons. First, we compare performance and total spike events between MTN without rotation and EPMTN at four timesteps. Second, we increase the timestep budget of MTN without rotation and measure the timesteps and total spike events required to achieve performance comparable to the four-step EPMTN model.

Table 1: Performance and spiking activity at T=4

ModelMethodWT2 PPL ↓Pile PPL ↓Events (M/token)Estimated energy (mJ/token)
Llama2-7BANN5.47214.5134-31.0119
Llama2-7BMTN without rotation8.57687.04826.997121.1351
Llama2-7BEPMTN6.20865.69937.650724.2491
Llama3-8BANN6.13585.3535-35.1408
Llama3-8BMTN without rotation8.23146.81266.940724.4722
Llama3-8BEPMTN6.99346.22517.631927.5918

At T=4, EPMTN achieves lower PPL on both datasets for both models. Meanwhile, spike events increase by 9.34% and 9.96%, and estimated energy increases by 14.73% and 12.75%, respectively. These results confirm that rotations improve conversion accuracy while increasing spiking activity at a fixed timestep budget.

Table 2: Timesteps required to satisfy the predefined accuracy criterion

ModelMethodTWT2 PPL ↓Pile PPL ↓Events (M/token)Estimated energy (mJ/token)
Llama2-7BANN-5.47214.5134-31.0119
Llama2-7BMTN without rotation76.02115.144012.266736.8657
Llama2-7BEPMTN46.20865.69937.650724.2491
Llama3-8BANN-6.13585.3535-35.1408
Llama3-8BMTN without rotation66.94045.967110.449335.7920
Llama3-8BEPMTN46.99346.22517.631927.5918

We progressively increase the timestep budget of MTN without rotation until its PPL on each dataset is not less than that of EPMTN at T=4. Llama2-7B and Llama3-8B without rotation first satisfy this criterion at T=7 and T=6, respectively. Compared with these configurations, EPMTN at T=4 reduces estimated energy by 34.22% and 22.91%, respectively. Thus, our method achieves lower estimated energy in comparisons subject to the predefined accuracy criterion.

Question 1

We thank the reviewer for the suggestion. Rotation calibration is not restricted to multi-threshold neurons. Figure 4(a) already includes experiments with single-threshold signed IF neurons. At 16 timesteps, fixed and learned rotations reduce PPL from 58,815.89 to 16.35 and 8.55, respectively, demonstrating that the approach also benefits single-threshold neurons. However, MTNs with learned rotations achieve a PPL of 6.21 using only four timesteps. Rotations improve activation distributions but do not eliminate the representational limitations of neurons under a finite timestep budget. Multiple thresholds therefore remain beneficial for accurate conversion with few timesteps. To further investigate different neuron types, we add comparisons of the complete configurations using Signed IF and LIF at T=4, 8, 16, 32, and 64, and our method at T=4, 6, and 8.

Supplementary Table 3a: Llama2-7B with FP16 weights

MethodTWT2 PPL ↓Pile PPL ↓Events (M/token)Estimated energy (mJ/token)
Signed IF4NFNF1.10194.2427
Signed IF81019.3918NF2.63359.6535
Signed IF168.547712.54285.956920.7897
Signed IF326.56825.380412.710443.2039
Signed IF645.81304.749626.284788.2011
LIF414655.74968090.63801.13683.9986
LIF82963.9650NF2.47319.1758
LIF16205.9636103.77145.441719.3664
LIF3232.306821.798411.495839.9289
LIF6417.123212.970023.632281.1216
Ours (MTN / EPMTN)46.20865.69937.650724.2583
Ours (MTN / EPMTN)65.60114.954111.460635.9394
Ours (MTN / EPMTN)85.49504.794615.265847.6139

Supplementary Table 3b: Llama3-8B with FP16 weights

MethodTWT2 PPL ↓Pile PPL ↓Events (M/token)Estimated energy (mJ/token)
Signed IF4336931.4063232509.12791.73737.1308
Signed IF82688.34722309.39532.335411.6979
Signed IF1627.534119.44435.208922.9658
Signed IF328.26436.881811.201946.1129
Signed IF646.73265.804523.228192.3593
LIF4442731.9113387125.70851.64786.6646
LIF816673.647011658.14172.647411.8463
LIF16392.6687319.86334.738921.6276
LIF3249.183535.745910.031942.7851
LIF6425.275818.549420.647485.1096
Ours (MTN / EPMTN)46.99346.22517.631927.6012
Ours (MTN / EPMTN)66.50645.754111.434740.1559
Ours (MTN / EPMTN)86.34395.585915.244352.7115

NF denotes nonfinite PPL on the full evaluation. The rotation framework is compatible with IF and LIF neurons, but single-threshold neurons still require more timesteps to achieve high accuracy. Among the evaluated configurations, Signed IF requires T=64 on both models to outperform our T=4 configuration on both datasets. Its estimated energy is 3.64 times and 3.35 times that of our method, respectively. Even at T=64, LIF remains less accurate than our T=4 configuration and consumes more estimated energy. Compatibility with IF and LIF therefore does not remove their timestep and energy costs. Multi-threshold neurons retain an advantage for accurate conversion with few timesteps.

Question 2

We thank the reviewer for the question. This is a tradeoff between spiking activity per timestep and the total timestep budget. Rotations increase activity at a fixed timestep budget, but they can also reduce the number of timesteps needed to reach the target accuracy. In Table 2, Llama2-7B and Llama3-8B without rotation require T=7 and T=6, respectively, to satisfy the predefined accuracy criterion, whereas our method requires only T=4. The corresponding estimated energy reductions are 34.22% and 22.91%. Thus, when accuracy is prioritized, reducing the timestep budget can offset the cost of increased activity per timestep. These values are energy estimates.

Question 3

We thank the reviewer for raising the issue of low-bit weight deployment. After calibration, our method folds the learnable rotations into the corresponding linear-layer weights, so these rotations and their inverses do not need to be explicitly executed during inference. Low-bit quantization is applied to the folded weights. The relevant question is how the resulting quantization error affects model performance, rather than whether the quantized folded weights remain orthogonal. To evaluate this compatibility, we add experiments that quantize the folded weights and compare MTN without rotation with EPMTN at different weight precisions. These experiments assess whether the benefits of rotation calibration persist after weight quantization.

At T=4, we compare MTN without rotation and EPMTN with FP16, W8, and W4 weights. We use the official GPTQ implementation from QuaRot with 128 WikiText-2 training blocks of 2048 tokens each and seed=0.

Table 4: Comparison across weight precisions at T=4

ModelWeightsWT2, no rotationWT2, EPMTNPile, no rotationPile, EPMTN
Llama2-7BFP168.57686.20867.04825.6993
Llama2-7BW88.55956.21217.02945.7005
Llama2-7BW4NF6.49189.46625.9885
Llama3-8BFP168.23146.99346.81266.2251
Llama3-8BW88.22026.99206.79876.2283
Llama3-8BW414.22447.391412.64846.6495

At W8, EPMTN maintains PPL close to its FP16 counterpart. At W4, EPMTN outperforms MTN without rotation in every corresponding comparison with finite PPL. The Llama2-7B configuration without rotation at W4 returns NaN on the full WikiText-2 test set and is marked as NF, while its Pile result remains valid. Although reducing weight precision introduces quantization error, these results show that the benefits of rotation calibration are retained.

Question 4

We thank the reviewer for identifying this missing comparison. The original Table 6 compares our method only with FP16 ANN and does not directly show the incremental cost relative to an SNN without rotation. We add comparisons between MTN without rotation and EPMTN at the same timestep budget and at comparable accuracy. We report PPL, timesteps, spike events per token, and estimated energy per token, together with the energy changes relative to MTN without rotation.

These comparisons are provided in Tables 1 and 2 of our response to Weakness 2. At T=4, the relative changes in energy are +14.73% for Llama2-7B and +12.75% for Llama3-8B. Relative to the configurations without rotation that satisfy the predefined accuracy criterion at T=7 and T=6, the changes are -34.22% and -22.91%, respectively. These results directly demonstrate the tradeoff between the accuracy benefits of rotation, the timestep budget, and energy.

Reviewer_TMb3OpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper addresses ANN-to-SNN conversion errors in LLMs at low time-step counts by learning orthogonal rotations through a task loss. It also introduces EPMTN, which expresses the time-averaged response of a particular MTN formulation in a parallel form without temporal unrolling. I appreciate the improvements in four-step conversion accuracy and calibration efficiency, but questions remain about the execution of the complete spiking network and the contribution beyond existing rotation-based methods.

Critical strengths
  1. Substantial improvements in four-step results. Table 3 reports average perplexities of 6.94 and 8.20 for LLaMA2-7B and LLaMA3-8B, respectively, while Table 4 reports average accuracies of 63.43% and 69.67%. These results are markedly better than the listed four-step Phase Coding baselines. The internal comparison in Figure 4(a) is also informative: for LLaMA2-7B on WikiText-2, perplexity decreases from 8.58 without rotation and 8.26 with fixed rotation to 6.21 with learned rotation, supporting the benefit of learning the rotations themselves.
  2. Concrete calibration benefits from EPMTN. Proposition 1 explicitly states the conditions under which equivalence holds, and the appendix provides a proof. Table 2 reports a reduction in four-step calibration training time from 59.67 to 17.85 minutes and in peak memory usage from 58.26 to 16.59 GiB. The paper also clearly distinguishes equivalence of forward outputs from differences in surrogate gradients.
Critical weaknesses

None

Questions for authors

None

1 comment

We thank the reviewer for recognizing the improvements in four-step conversion accuracy and calibration efficiency. We agree that orthogonal rotations, task-loss calibration, and Cayley updates build on existing quantization methods, which we explicitly acknowledge and cite. Our contribution is to establish a rotation-compatible ANN-to-SNN conversion framework that addresses the mismatch between pretrained activations and the finite-timestep responses of MTNs. To enable efficient calibration, we introduce EPMTN, which expresses the time-averaged response of the adopted MTN formulation in a parallel form. Under the stated conditions, EPMTN preserves the forward response while avoiding temporal unrolling and enabling surrogate-gradient optimization of the complete response. EPMTN is therefore directly connected to the feasibility of task-driven MTN calibration. Our ablations separately demonstrate the accuracy benefits of learned rotations and the calibration efficiency improvements enabled by EPMTN.

To evaluate the implications of increased spiking activity, the table below includes the FP16 ANN baseline and compares configurations at both a fixed timestep budget and a predefined accuracy criterion. At T=4, learned rotations improve PPL on both datasets but increase estimated energy by 14.73% and 12.75% for Llama2-7B and Llama3-8B, respectively. When the models without rotation are required to achieve PPL no less than that of EPMTN on both datasets, they require T=7 and T=6, respectively. Compared with these configurations, EPMTN at T=4 reduces estimated energy by 34.22% and 22.91%. These results demonstrate an accuracy–timestep–energy tradeoff, rather than an unconditional reduction in spiking activity.

Table: Accuracy and estimated energy of FP16 ANN, MTN without rotation, and EPMTN with learned rotations.

ModelMethodTWikiText-2 PPL ↓Pile PPL ↓Estimated energy (mJ/token) ↓
Llama2-7BANN FP16—5.47214.513431.0119
Llama2-7BMTN without rotation48.57687.048221.1351
Llama2-7BMTN without rotation76.02115.144036.8657
Llama2-7BEPMTN with learned rotations46.20865.699324.2491
Llama3-8BANN FP16—6.13585.353535.1408
Llama3-8BMTN without rotation48.23146.812624.4722
Llama3-8BMTN without rotation66.94045.967135.7920
Llama3-8BEPMTN with learned rotations46.99346.225127.5918

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response

We thank all reviewers for their time and constructive feedback. Their comments have helped us clarify the scope and contributions of our work and identify important directions for additional evaluation. In our rebuttal, we have provided further explanations and experiments addressing the relationship to activation quantization, the role of EPMTN, compatibility with different neuron types and low-bit weights, and the tradeoff between conversion accuracy and estimated energy consumption.

Our contribution builds on established rotation techniques from LLM quantization while addressing the calibration challenges of multi-threshold spiking dynamics under a limited timestep budget. Specifically, EPMTN is derived as a parallel formulation that exactly preserves the time-averaged forward response of the adopted MTN dynamics. It enables end-to-end calibration without temporal unrolling and provides a surrogate gradient for the complete response. Our controlled experiments and ablations examine this equivalence and demonstrate improvements in calibration time, memory consumption, and final perplexity. Additional comparisons with single-threshold IF and LIF neurons further support the benefits of MTNs for accurate conversion with few timesteps.

We have also expanded the evaluation of deployment-related tradeoffs. Direct comparisons with unrotated MTNs show that rotations increase spiking activity and estimated energy at a fixed timestep budget, but reduce the timesteps required to meet the specified accuracy criterion, yielding estimated energy reductions of 34.22% and 22.91% on the two evaluated models. Additional W8 and W4 experiments show that the benefits of rotation calibration persist after weight quantization. Comparisons with QuaRot demonstrate improvements on selected perplexity metrics at comparable estimated energy costs. We also provide an operator-level analysis and neuron-based approximations of Transformer nonlinear functions, with the latter largely preserving performance in the evaluated configuration.

We hope these clarifications and additional results help address the reviewers’ concerns and provide a clearer basis for assessing our contributions and their limitations.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.