acceptodds
Withdrawn from ICLR 2027

GEAR: Response-Aware Test-Time Compute Allocation for Frozen Quantized Spiking Transformers

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Frozen quantized spiking classifiers cannot change their weights at deployment, but they can spend extra computation on selected inputs. We study when an alternative view helps rather than merely changes a prediction. GEAR routes low-margin images from a frozen 4-bit QSD-Transformer through nested high-resolution and flipped views, then fuses their predictions. Its empirical response curve models a fixed action's net repair utility—repairs minus newly introduced errors—along the cheap-view margin rank. On ImageNet validation and five ImageNet-derived shifts, a fixed 26:8:4 policy gains 0.79–2.46 percentage points without target-set retuning. A labeled target-like probe calibrates the curve and operating width, evaluated on disjoint held-out rows; geometry interventions show why a source-optimal crop need not transfer. Tests on a second spiking family, synthetic corruptions, detection, and segmentation clarify its scope: margin routing helps when escalation corrects errors concentrated among uncertain inputs, but not when refinement mainly benefits confident ones. The response profile is conditioned on the chosen action and deployment distribution.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Spiking Neural Networks, Quantized Spiking Transformers, Response-Aware Inference, Test-Time Compute Allocation, Adaptive Test-Time Augmentation, Distribution Shift

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_Sdm7OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

In this paper, the authors propose GEAR, which addresses test-time compute allocation for frozen quantized spiking classifiers by distinguishing sample uncertainty from the actual utility of applying an additional inference action. It routes low-margin inputs through nested higher-resolution and flipped views, while using an action-conditioned response curve based on repair-minus-harm utility. Experiments on ImageNet and several ImageNet-derived distribution shifts show that the fixed routing policy improves accuracy, while probe-based calibration further adapts the operating width to target-like streams. The results also show that the benefit of escalation is action- and distribution-dependent, with failure cases occurring when improvement caused by GEAR is not concentrated among uncertain samples.

Critical strengths

Strengths

  1. This paper tackles a meaningful problem that can occur during inference of transformer models.
  2. The authors provide broad experimental results.
  3. The authors adopt their mechanism in a practical processing pipeline, including JPEG decoding, beyond simple image classification tasks.
Critical weaknesses

Weaknesses

  1. The introduction of this paper is really hard to read. The problem statement is not structured well, and I cannot follow how the authors decide to propose GEAR based on the mentioned problems.
  2. I think the core problem pointed out by the authors is that uncertainty alone does not tell whether additional test-time computation will actually improve a prediction. However, I do not think this core problem is specific to the SNNs. This problem is also related to ANN-based transformers.
  3. The discussion of dynamic networks at the beginning of Section 2 feels somewhat abrupt. While these methods share the high-level goal of input-dependent compute allocation with GEAR, their underlying mechanisms are different. It would improve readability to clarify GEAR's positioning within adaptive inference and then introduce dynamic networks as related but distinct approaches.
  4. Equation (2) approximates the additional computational workload using terms proportional to the input image size. However, it is not clear why the inference cost of the QSD-Transformer should scale proportionally with spatial resolution squared. Actually, different components of a Transformer may scale differently with the number of input tokens.
  5. The choice of k1=26k_1=26 appears to be an empirically selected operating point rather than a value derived from a theoretical analysis. I think this value can depend on the dataset and batch size, as Table 6 in the Appendix. Furthermore, this kind of design makes GEAR seem hardly dependent on heuristic approaches.
  6. Although the authors deal with SNNs, they do not consider event-based or sequential datasets at all. Experimental results with these datasets can enhance SNN-specific aspects of GEAR.
  7. The comparative results with existing approaches are in the Appendix. I recommend that the authors show these results in the main text of their paper.
Questions for authors

Questions

  1. What are SNN-specific aspects of GEAR?
  2. Could you provide a theoretical foundation for k1k_1 selection? If you cannot, please provide an intuitive, brief explanation to justify the selection.
  3. The authors mention dynamic neural networks in Section 2. Could you provide the advantages and disadvantages of GEAR compared to dynamic neural networks?
Reviewer_Er3EOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes GEAR, which allocates additional test-time computation to a frozen quantized spiking transformer. After a cheap prediction, low-margin inputs receive higher-resolution or flipped views through a nested cascade, and the predictions are fused. A response curve estimates the net benefit of these actions from repaired and newly introduced errors. The paper evaluates a fixed policy across distribution shifts and a variant calibrated with a separate labeled probe.

Critical strengths

The paper demonstrates that selective additional views can improve the accuracy of a frozen model without updating its weights. The fixed policy reports gains of 0.79-2.46 percentage points on five external datasets without target-set fine-tuning.

Critical weaknesses
  1. The main idea is not sufficiently justified in terms of the accuracy–efficiency trade-off. For resource-constrained deployment of a quantized spiking transformer, the key question is whether the accuracy improvement warrants the additional latency and energy consumption relative to the original single-pass inference. For example, Route 26/64 improves accuracy by 1.266 percentage points while increasing latency by 67.0% and energy per image by 93.2%; Cascade 26:8 gains 1.608 percentage points at the cost of 86.5% higher latency and 112.5% higher energy consumption. These modest accuracy gains do not convincingly justify the substantial additional overhead.

  2. The presentation lacks clarity, making it difficult to understand the proposed method and assess its claimed practical benefits.

Questions for authors
  1. Can you clarify why the modest accuracy gains in Table 3 justify the substantial increases in latency and energy consumption?

  2. Do these gains remain worthwhile on resource-constrained edge devices or for small-batch, latency-sensitive workloads? Could you evaluate end-to-end latency under these conditions?

  3. Can the proposed method be applied to other models, such as Spikformer or SDT-V3?

Reviewer_9xEbOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The authors propose GEAR, a response-aware test-time compute allocation strategy for frozen quantized spiking transformer. GEAR is practical and empirically grounded adaptive test-time augmentation (TTA) policy, From a cheap initial baseline view, the method identifies low-margin images and routes them through an empirically chosen 26:8:4 cascade to improve the classification accuracy without changing any model parameters. GEAR realizes a top-1 accuracy gain of 0.79 to 2.46% on ImageNet and various distribution shifts.

Critical strengths
  1. The safety-gated calibration probe is a robust solution to distribution shift, which overcomes a critical limitation of standard adaptive TTA.
  2. The repair-hard accounting identity explicitly measures the overall advantages in an efficient manner.
Critical weaknesses
  1. Irrelevance to SNNs. This method is neither tailored to SNNs nor capable of making use of dynamics of SNNs, which is evidenced by the fact that it can transfer to standard ResNet and ViT.
  2. Limited generalization. For segmentation tasks, high-margin images benefit more from spatial scaling than low-margin inputs, which significantly undermines the assumption that low-confidence routing is optimal.
  3. Routing thresholds and heuristics. The 26:8:4 routing thresholds and fixed 0.3/0.7 fusing weights are too empirical to be generalized to various tasks.
Questions for authors
  1. The 26:8:4 routing threshold is strictly tailored to a batch size of 64. How does the response curve and compute allocation generalize to dynamic batch sizes or online inference (batch size=1)?
  2. Why are the fixed 0.3/0.7 fusing weights used in place of dynamic confidence-weighted fusion?
  3. Why does the title emphasize “Frozen quantized spiking transformers”, this method is indeed architecture-agnostic?

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

Related papers

Discussion (0)

Sign in to comment.