acceptodds
Under review as a conference paper at ICLR 2027

Overlapping Subspace Communication for Spiking Decision Transformer

Abstract

Offline reinforcement learning can be formulated as sequence modeling, enabling policies to learn directly from fixed trajectories. While spiking Decision Transformers provide an efficient alternative to dense Transformers through sparse event-driven computation, the internal communication structure of sparse spiking attention remains underexplored. We observe that spike-driven communication supports vary across return-ranked trajectories, while a subset of token interactions repeatedly recurs, suggesting the coexistence of reusable and context-dependent communication patterns. Motivated by this observation, we propose the Spiking Subspace Decision Transformer (SS-DT) with Overlapping Spiking Subspace Attention (OSSA). OSSA factorizes intra-head spike communication into multiple overlapping subspaces, where shared components capture reusable structures and private components preserve context-specific variations. Spike-guided allocation dynamically composes these subspaces according to spiking activity, while predictive feedback and return-conditioned modulation further adapt communication to trajectory dynamics and decision context. Together, SS-DT introduces structured adaptive communication into sparse spiking sequence modeling. Across 21 tasks on MuJoCo and Adroit, SS-DT outperforms the strongest spiking baseline by 3.7 normalized-return points on average, with gains of 1.3 and 5.4 points on MuJoCo and Adroit, respectively, establishing state-of-the-art average performance among spiking offline RL methods.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Spiking Neural Networks, Spiking Decision Transformer, Offline Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_XwGeOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SS-DT, using intra-head shared–private overlapping subspaces, spike-guided soft allocation and latent prediction feedback for offline decision making. Across 21 D4RL MuJoCo and Adroit tasks, it reports higher mean normalized return than PSSA on 19 tasks and a roughly 3.7-point overall average improvement. It includes progressive component ablations, an added-head comparison, seed-wise plots for selected tasks, and qualitative/entropy-based communication analyses. The method is plausible, but the evidence does not yet isolate the contribution and cost of the central inductive bias

Critical strengths
  1. Broad task coverage and several meaningful gains. In Table 1, the proposed method improves the reported means on 19 of 21 tasks relative to PSSA, notably pen-cloned (80.05 vs 44.8). The MuJoCo/Adroit split demonstrates that the design is not evaluated solely on locomotion.
  2. The work attempts to connect a communication-structure hypothesis to learned behavior rather than reporting only performance. Figure 1 motivates recurrent and variable token-pair supports; Figures 4–5 and supplementary Appendix C provide subspace allocations, Jaccard similarities and communication-entropy statistics. These are useful descriptive results, albeit not proof of causal necessity.
Critical weaknesses
  1. The central shared–private design is not independently justified by the current ablations. Equation (4) is an algebraic decomposition of concatenated projections, not evidence that common and private information are disentangled. Table 2 does contain an ablation, but its 'PSSA + Shared–Private' row improves only three of five listed tasks and decreases two relative to PSSA. The full method adds spike allocation, predictive feedback, a two-stage process and additional capacity; the PSSA+Heads comparison does not match those components and operation budgets. Moreover, no corresponding controlled ablation is reported for Adroit, where the largest benefits occur. Correlation regularization is mentioned in Eq. (19) but lacks a sufficiently explicit evaluated contribution. This weakens attribution to the proposed overlapping subspace inductive bias.
  2. Robustness and spiking-efficiency claims remain underdocumented. Figure 3 does not identify the comparison baseline sufficiently clearly, and its positive hopper-medium-replay seed-wise gains are difficult to reconcile with the Table 1 mean being 0.89 points below PSSA without further protocol/baseline details. Figure 3 also shows considerable seed variation on hopper-medium. Table 2 provides no uncertainty for incremental gains. In addition, the learned shared–private projections follow binary LIF Q/K/V, and the two-stage predictor plus mean-normalized sigmoid gating introduce dense real-valued operations; no end-to-end parameter/operation, latency or energy accounting is supplied. This does not make a spiking implementation impossible, but it leaves the return-versus-cost tradeoff unsupported.
  3. Figures 4–5 and supplementary analyses illustrate context-dependent subspace allocation, but these observations do not establish that such adaptation improves decision-making. Although Table 2 evaluates spike allocation, it does not distinguish the benefit of context-dependent gating from that of learned but fixed communication weights. This leaves the functional necessity of adaptive subspace routing insufficiently demonstrated
Questions for authors
  1. Please provide a resource-matched comparison of shared–private OSSA against fully independent and/or fully shared subspaces with the same gating, predictive objective and two-stage protocol. Can you distinguish the prediction loss from feedback modulation and add uncertainty on representative MuJoCo and Adroit tasks?
  2. Please identify Figure 3's exact baseline, checkpoint selection, training/evaluation seeds and episode counts, and explain the hopper-medium-replay discrepancy relative to Table 1. Which improvements are reliably larger than training/evaluation variance?
  3. After Eq. (3), are projected Q/K/V subspace features re-spiked or real-valued? How does supplementary Attn_spk map to main Eq. (5), and what are the complete inference costs (both stages, all subspaces, prediction and gating) relative to PSSA?
  4. Can the authors compare the proposed adaptive gating with fixed or context-shuffled gating while keeping the subspace projections and model capacity unchanged? This would help determine whether context-dependent communication itself contributes to improved decision-making performance.
5 comments

Q2: We apologize for not distinguishing the two sources of PSSA results in the manuscript. Table 1 cites the result published in the original PSSA paper, whereas Fig. 3 compares SS-DT with our rerun of the released PSSA code using matched training seeds 123–126. For each run, we select the highest-scoring evaluated checkpoint–RTG pair. Each target is evaluated over five environment-seed groups of ten episodes. The table below lists the selected iteration, target RTG, and score for each Hopper-medium-replay seed. The published PSSA Hopper-medium-replay score in Table 1 is 96.3±1.396.3 \pm 1.3, whereas our paired PSSA rerun achieves 92.38±2.4592.38 \pm 2.45, compared with 95.41±2.5895.41 \pm 2.58 for SS-DT. The four paired improvements are +2.17+2.17, +3.54+3.54, +2.77+2.77, and +3.61+3.61 points, respectively, yielding a mean paired improvement of 3.023.02 points with a sample standard deviation of 0.680.68. Thus, the Table 1 comparison and Fig. 3 paired comparison use different PSSA reporting sources. We identify both sources, the checkpoint-selection rule, and the evaluation episode count in the revised Table 1 note and Fig. 3 caption. All four training seeds yield positive paired improvements, ranging from 2.17 to 3.61 points, indicating consistent gains across the evaluated seeds. The sample standard deviation of the paired gains is 0.68 points, suggesting relatively low variation across training seeds.

Table: Provenance of the Hopper-medium-replay comparison in Fig. 3. Both models use d=256d=256, four Transformer blocks, one attention head, and training seeds 123-126. Each evaluated target uses five environment-seed groups of ten episodes.

Training seedPSSA IterationPSSA Target RTGPSSA ScoreSS-DT IterationSS-DT Target RTGSS-DT ScorePaired gain
12348360088.8440360091.01+2.17
12439360094.4039360097.94+3.54
12533360093.4041360096.17+2.77
12637360092.9037360096.51+3.61
Mean----92.38----95.41+3.02

Q3: As in PSSA, Linear-Norm-LIF projections produce binary Q/K/V spikes. OSSA applies learned shared and private projections to these spikes; the resulting subspace features and their attention products are real-valued. After the subspace messages are combined, the attention output passes through an LIF neuron before continuing through the spiking Transformer. We replace the supplementary Attnspk\mathrm{Attn}_{\mathrm{spk}} shorthand with the explicit masked and positionally modulated QK⊤VQK^\top V computation corresponding to the main-text attention equation.

We profile the complete inference path, including both Transformer passes, all subspaces, dense projections, attention matrix products, latent prediction, communication gating, and the action head. The following table compares PSSA, its widened variant (PSSA-MoreHeads), and SS-DT under the same batch size and context length.

Table: Computational cost and performance comparison (B=1B=1, context length 20).

ModelParams (M)FLOPs (G)Compute (μ\muJ)Latency (ms)Hopper-m-rWalker2d-m-rPen-cloned
PSSA3.180.50768.211.5592.38±2.4592.38 \pm 2.4571.29±4.3271.29 \pm 4.3244.92±13.4944.92 \pm 13.49
PSSA-MoreHeads6.271.000139.018.293.10±2.6693.10 \pm 2.6671.65±4.1071.65 \pm 4.1049.95±17.7649.95 \pm 17.76
SS-DT6.4371.274664.137.3895.41±2.5895.41 \pm 2.5874.63±3.5574.63 \pm 3.5580.05±7.9080.05 \pm 7.90

Although SS-DT introduces additional parameters and inference overhead compared with the original PSSA, it achieves better performance and stability. And, as demonstrated in our response to Q1, resource-matched comparisons against fully shared and fully independent subspace constructions show that SS-DT consistently achieves better performance under closely matched parameter counts, FLOPs, and latency. This provides further evidence that the shared-private subspace design contributes to the performance gains beyond simply increasing model capacity.

Q4: We compare learned token-independent gates, context-shuffled gates, and adaptive gates while retaining the same OSSA projections, predictor, and two-pass protocol. Context shuffling preserves the marginal gate-value distribution while changing which trajectory receives each gate vector; the causal/local attention mask remains unchanged. Table below reports the resulting returns.

Table: Effect of context-dependent communication allocation.

TaskLearned fixed gateContext-shuffled gateAdaptive gate (SS-DT)
Hopper-medium-replay92.9±2.992.9 \pm 2.991.7±3.291.7 \pm 3.295.41±2.5895.41 \pm 2.58
Walker2d-medium-replay71.6±3.871.6 \pm 3.870.8±4.170.8 \pm 4.174.63±3.5574.63 \pm 3.55
Pen-cloned75.2±8.575.2 \pm 8.573.9±8.873.9 \pm 8.880.05±7.9080.05 \pm 7.90
Pen-human83.8±5.883.8 \pm 5.881.9±6.281.9 \pm 6.287.37±5.3887.37 \pm 5.38

Adaptive gates exceed fixed gates by 2.51, 3.03, 4.85, and 3.57 points across the four tasks, respectively. Their gains over context-shuffled gates are 3.71, 3.83, 6.15, and 5.47 points. The shuffled control tests whether learned gate values suffice without being assigned to the corresponding decision context; the consistent reduction in returns supports the functional contribution of context-aligned allocation.

W1: The answer to Q1 provides the missing resource-matched structural comparison on both MuJoCo and Adroit: OSSA outperforms both fully shared and fully independent channels while retaining the other model components. The accompanying ablation separately evaluates predictive supervision, feedback modulation, and the explicitly defined private-projection correlation penalty. These controlled comparisons provide evidence that the shared--private design contributes to the observed performance gains.

W2: The answer to Q2 identifies the matched PSSA rerun behind the Hopper-medium-replay panel of Fig. 3 and reconciles it with the separately published PSSA aggregate in Table 1. For the distinct Hopper-medium task, we extended the paired evaluation from four to seven training seeds (123–129). SS-DT achieves higher returns on five of the seven seeds, with a mean paired improvement of 2.00 points. Although the gains vary across seeds, the overall mean remains positive. The answer to Q1 also reports seed-level uncertainty for the controlled component comparisons. Q3 clarifies that binary LIF outputs are followed by real-valued operations and accounts for both Transformer passes in the parameter and dense-equivalent FLOP estimates.

Table: Paired-seed comparison between PSSA and SS-DT on Hopper-medium.

Training seedPSSA ScoreSS-DT ScorePaired gain
12379.3276.98−2.34
12478.8477.87−0.97
12572.2473.30+1.06
12671.8980.99+9.10
12776.3077.05+0.75
12875.6978.45+2.76
12973.2976.91+3.62
Mean ± sample SD75.37 ± 3.0277.36 ± 2.29+2.00 ± 3.74

W3: Figures 4-5 visualize context-dependent subspace allocation, while the additional experiment in Q4 evaluates its contribution to performance. With subspace projections and model capacity unchanged, adaptive gating outperforms both learned fixed gating and context-shuffled gating across all reported tasks. In particular, the shuffled control retains the learned gate-value distribution but removes its alignment with the input trajectory, providing direct evidence that context-dependent communication contributes to improved decision-making.

Thanks for the constructive suggestions. Given the limited time, we are currently incorporating all relevant clarifications into our revised manuscript.

Q1: We thank the reviewer for identifying two attribution questions: whether shared-private communication is beneficial under matched resources, and whether predictive feedback contributes beyond auxiliary prediction supervision. We address them with controlled comparisons using the same outer attention-head count, training and evaluation protocol, return targets, checkpoint-selection rule, and paired training seeds. All reported returns are means ± standard deviations over matched training seeds, using the same seed set across compared variants.

Shared-private subspace construction. We compare three ways of constructing four communication channels within each attention head. Fully shared channels use the same query, key, and value projections; fully independent channels use separate projections; OSSA combines projections shared across channels with channel-specific private projections. The adaptive gate, latent prediction objective, feedback modulation, and two Transformer passes are held fixed. Channel widths are adjusted to match parameter and operation budgets as closely as possible. Table below reports the resulting costs and normalized returns. Under closely matched parameter counts, FLOPs, and inference latency, OSSA consistently outperforms both controls across all three representative MuJoCo and Adroit tasks.

Table: Resource-matched subspace constructions in the complete model.

VariantParams (M)FLOPs (G)Latency (ms)Hopper-m-rWalker2d-m-rPen-cloned
Fully shared6.4331.27636.9090.06±2.7390.06 \pm 2.7372.64±2.1572.64 \pm 2.1579.11±9.8379.11 \pm 9.83
Fully independent6.4381.27337.7093.57±2.8993.57 \pm 2.8972.87±3.8472.87 \pm 3.8478.80±12.4178.80 \pm 12.41
Shared–private OSSA (SS-DT)6.4371.27437.3895.41±2.5895.41 \pm 2.5874.63±3.5574.63 \pm 3.5580.05±7.9080.05 \pm 7.90

Auxiliary prediction versus feedback modulation. We hold OSSA fixed and compare three configurations: no latent prediction, auxiliary prediction without error-based gate modulation, and the full model with both. Each retains the two-pass protocol. We also remove only the private-projection correlation penalty from the full model to evaluate its separate contribution. Auxiliary prediction consistently improves performance over OSSA across all three tasks. Using the prediction error to further modulate the communication gate brings additional gains, reaching 3.85 points on Pen-cloned.

Removing only Lcorr\mathcal{L}_{\mathrm{corr}} also leads to performance drops across all three tasks, confirming its additional contribution. These results demonstrate the complementary benefits of predictive supervision, error-based feedback modulation, and private-projection correlation regularization. We define it explicitly in the Method: it penalizes similarity among private projections, while the shared projection remains available to every channel.

Table: Attribution of predictive supervision, feedback modulation, and private-projection correlation regularization.

VariantLpred\mathcal{L}_{\mathrm{pred}}Error modulates gateLcorr\mathcal{L}_{\mathrm{corr}}Hopper-m-rWalker2d-m-rPen-cloned
OSSA×\times×\times✓\checkmark92.5±2.392.5 \pm 2.372.0±2.772.0 \pm 2.775.8±2.575.8 \pm 2.5
OSSA + Aux-only✓\checkmark×\times✓\checkmark93.7±2.193.7 \pm 2.172.6±1.972.6 \pm 1.976.2±2.376.2 \pm 2.3
Full model w/o Lcorr\mathcal{L}_{\mathrm{corr}}✓\checkmark✓\checkmark×\times94.30±2.5294.30 \pm 2.5273.80±3.7873.80 \pm 3.7879.19±8.3679.19 \pm 8.36
Full model✓\checkmark✓\checkmark✓\checkmark95.41±2.5895.41 \pm 2.5874.63±3.5574.63 \pm 3.5580.05±7.9080.05 \pm 7.90
Reviewer_cAchOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This work proposes Spiking Subspace Decision Transformer (SS-DT), a spiking Transformer architecture for offline reinforcement learning. The Overlapping Spiking Subspace Attention (OSSA) decomposes intra-head communication into multiple overlapping subspaces with shared and private projections. The shared components capture recurring communication patterns, while the private components preserve context-dependent variations. The method is evaluated on 21 D4RL tasks covering MuJoCo locomotion and Adroit dexterous manipulation. The authors report average normalized returns of 80.1 on MuJoCo and 54.0 on Adroit, outperforming the strongest spiking baseline by 1.3 and 5.4 points.

Critical strengths
  1. The paper investigates the internal communication structure of spiking Transformers for offline decision making, which is an interesting and potentially useful direction.

  2. The experiments cover 21 tasks across locomotion and dexterous manipulation, with comparisons against several offline RL and spiking sequence-modeling baselines. The reported improvements are promising.

Critical weaknesses
  1. The presentation is difficult to follow. Several figures are difficult to interpret (their axes, legends, or construction procedures are insufficiently explained). Some concepts are not defined precisely enough to make the proposed method and its motivation understandable. I suggest that the authors proofread and clarify the manuscript.

  2. The computational and energy-efficiency claims are insufficiently supported. The proposed architecture introduces additional subspace projections, communication gates, a latent predictor, and a two-stage Transformer computation, the additional computation could be substantial. The use of LayerNorm and dense floating-point projections and matrix multiplications also complicates the computation. The paper should clarify the exact implementation used in the experiments and account for these operations when assessing efficiency.

Questions for authors
  1. Could the authors compare SS-DT with some representative ANN-to-SNN approaches such as conversion (https://openreview.net/pdf?id=goVeG3ui1Y) and distillation (https://ieeexplore.ieee.org/document/10204749)? If a suitable ANN teacher has already been trained, conversion and distillation would not require additional environment interaction or could be performed using fixed offline data without online interaction. SNN-based VLA (https://openreview.net/pdf?id=W86R5sIsxE) may also be trained without interactions. How would direct spiking offline RL approaches and SS-DT compare with these alternatives?

  2. Can you report BC and CQL baselines with both ANN and SNN variants in Table 1?

  3. What is the underlying cause of the recurring attention supports observed across trajectories? Do these observation also occours in ANN Decision Transformers or other SNN transformer architectures besides DT?

  4. How much additional computation is introduced by OSSA relative to other SNN Decision Transformers? How are LayerNorm, dense projections, attention matrix multiplications, and other non-spike-driven operations accounted for in the energy estimates?

5 comments

Q1: Following your suggestion, we train an IQL ANN teacher on Hopper-medium-replay using fixed offline data, without additional environment interaction. Using the same teacher checkpoint, we evaluate an SNN actor obtained through ANN-to-SNN conversion with CRPI and another SNN actor trained through offline action distillation. We compare both approaches with directly trained PSSA and SS-DT under the same training-seed and evaluation-episode protocol. Regarding the SNN-based VLA, we could not identify an officially released implementation. Its vision-language-conditioned robotic action generation task also differs from our setting, making direct numerical comparison inappropriate.

Table: D4RL normalized returns on Hopper-medium-replay.

PolicyNormalized return
IQL ANN teacher98.4±2.198.4 \pm 2.1
ANN-to-SNN conversion + CRPI91.1±2.591.1 \pm 2.5
Offline ANN-to-SNN distillation92.7±2.492.7 \pm 2.4
Directly trained PSSA92.38±2.4592.38 \pm 2.45
SS-DT95.41±2.5895.41 \pm 2.58

Both teacher-based approaches produce competitive SNN policies using only fixed offline data. Notably, offline action distillation improves the converted SNN policy by 1.6 points, suggesting that transferring knowledge from an offline-trained ANN teacher can improve SNN policy performance. Compared with directly trained PSSA, SS-DT achieves a 3.03 point improvement. Overall, SS-DT obtains the highest return among the evaluated SNN policies, outperforming ANN-to-SNN conversion, offline distillation, and directly trained PSSA by 4.31, 2.71, and 3.03 points, respectively, without requiring an ANN teacher.

We thank the reviewer for highlighting these valuable complementary approaches and have cited and discussed them in the revised manuscript.

Q2: Thank you for pointing out the ambiguity in Table 1. The BC and CQL columns report ANN-based offline RL baselines from the original publications. Given the limited rebuttal time and the non-trivial adaptation of CQL's critic and conservative objective to SNNs, we were unable to include SNN-CQL results. Instead, following your suggestion, we additionally trained ANN-BC and SNN-BC on Hopper-medium-replay using the same offline transitions, state normalization, training seeds, minibatch sequence, optimizer, learning rate, and 10,000 gradient updates. Both policies use two hidden layers of width 256 and have 69,635 trainable parameters. ANN-BC uses ReLU activations, while SNN-BC uses LIF neurons with four simulation steps. Both produce continuous actions and are evaluated for 20 episodes per training seed. SNN-BC achieves a 3.80 point higher mean return than ANN-BC in this matched experiment. This improvement may be related to the temporal integration dynamics of LIF neurons. We have incorporated these additional experiments and findings into the revised manuscript.

Table: Matched ANN-BC and SNN-BC results on Hopper-medium-replay.

PolicySeed 123Seed 124Seed 125Mean ± SD
ANN-BC19.1914.918.2914.13±5.4914.13 \pm 5.49
SNN-BC14.1224.0315.6617.93±5.3417.93 \pm 5.34

Q3: The recurring attention supports may reflect shared temporal structure across trajectories, together with input-dependent communication patterns. Causal constraints and position-dependent preferences may contribute to recurring communication patterns, while input-dependent dynamics allow the specific active supports to vary across trajectories. To investigate whether support recurrence is specific to spiking Decision Transformers, we further evaluated an ANN Decision Transformer and a causal adaptation of Spikformer's spiking self-attention (SSA) on Hopper-medium-replay. The ANN Decision Transformer uses two layers, hidden width 128, and 3,000 training updates. The causal SSA adaptation uses spike-based Q/K/V, four simulation steps, the same trajectory data, and the same training seed and update count. For both models, we sampled 48 twenty-position windows per return-ranked group and selected the top 30% of communication supports within a common causal, six-step local window.

The low-high return-context Jaccard similarity was 0.164 for ANN Decision Transformer and 0.683 for the causal SSA adaptation. These results suggest that recurring communication supports are not unique to our proposed architecture and can also emerge in other spiking attention mechanisms. The relatively low overlap in ANN Decision Transformer further suggests that support recurrence may depend on the underlying attention mechanism and representation dynamics, at least in our experimental setting. This observation motivates our investigation of adaptive communication mechanisms that explicitly accommodate both recurring and trajectory-dependent patterns.

Q4: We report complete-model parameters, dense-equivalent FLOPs, estimated compute energy, and end-to-end latency under the same batch size and context length. The accounting covers both Transformer passes, all subspaces, dense shared/private projections, attention matrix products, the predictor, and communication gates. Dense matrix operations are counted as MACs, while spike-triggered accumulations are counted as SOPs. Estimated compute energy applies 4.6 pJ per raw MAC and 0.9 pJ per SOP. LayerNorm, LIF state updates, nonlinearities, memory access, and runtime overhead are excluded from this analytical energy estimate but are included in the measured end-to-end latency. A MAC is counted as one operation for energy estimation but as two FLOPs in computational reporting.

Table: Computational cost and performance comparison across tasks (B=1B=1, context length 20).

ModelParams (M)FLOPs (G)Compute (μ\muJ)Latency (ms)Hopper-m-rWalker2d-m-rPen-cloned
PSSA3.180.50768.211.5592.38±2.4592.38 \pm 2.4571.29±4.3271.29 \pm 4.3244.92±13.4944.92 \pm 13.49
PSSA-MoreHeads6.271.000139.018.293.10±2.6693.10 \pm 2.6671.65±4.1071.65 \pm 4.1049.95±17.7649.95 \pm 17.76
Fully shared channels6.4331.276657.036.990.06±2.7390.06 \pm 2.7372.64±2.1572.64 \pm 2.1579.11±9.8379.11 \pm 9.83
Fully independent channels6.4381.273670.037.793.57±2.8993.57 \pm 2.8972.87±3.8472.87 \pm 3.8478.80±12.4178.80 \pm 12.41
SS-DT6.4371.274664.137.3895.41±2.5895.41 \pm 2.5874.63±3.5574.63 \pm 3.5580.05±7.9080.05 \pm 7.90

Although SS-DT incurs higher computational costs than the original PSSA, increasing parameters from 3.18M to 6.437M, estimated compute energy from 68.2 to 664.1 μ\muJ, and latency from 11.55 to 37.38 ms, the resource-matched comparisons demonstrate consistent performance gains from the proposed design. With nearly identical parameter counts and FLOPs, SS-DT outperforms both fully shared and fully independent channel controls across all three tasks. It also outperforms the expanded PSSA-MoreHeads baseline. These results support the effectiveness of the proposed shared-private, context-adaptive communication design, while highlighting its additional computational cost.

W1: We have addressed the presentation concerns by defining the model's spike outputs, real-valued subspace features, communication weights, and two-stage inference procedure explicitly. The revised figure captions specify the axes, legends, trajectory grouping, and support extraction procedure. Q3 further clarifies the possible sources of recurring supports and distinguishes shared temporal structure from trajectory-dependent communication. Together, these revisions make the motivation, method, and visual evidence directly interpretable.

W2: addressed the cost-accounting concern in Q4 by profiling the complete two-stage inference path, including every subspace, dense projection, attention product, predictor, gate, and action head. Dense matrix operations and spike-triggered accumulations are accounted for separately, while LayerNorm, LIF state updates, and other non-MAC/SOP operations are excluded from analytical compute energy and reflected in end-to-end latency. The comparison explicitly reports SS-DT's additional computational cost relative to PSSA. The capacity-controlled results further demonstrate that its performance improvement is associated with the shared-private adaptive communication design, rather than merely a larger parameter budget.

We sincerely thank the reviewer for the constructive suggestions. We have incorporated the relevant experiments, analyses, and clarifications into the revised manuscript.

Reviewer_5gtfOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This manuscript proposes a specific version of a spiking transformer.  Based on the observation that some sequence of tokens repeatedly occur, the authors propose a mechanism that they call “Overlapping Spiking Subspace Attention” (OSSA). This OSSA module process information in a dual way with some shared information across attention heads and some private information. The authors benchmark their version of spiking transformer and showed that for a list of MuJoCo and Adroit tasks, their method outperforms other spiking transformers.

Critical strengths
  • The benchmarking is well done. Indeed, the manuscript systematically compares the performance of their SSDT algorithm with many other algorithms across many tasks
Critical weaknesses
  • As I understand, the biggest claim of the paper is that the proposed “Overlapping Spiking Subspace Attention” (OSSA) mechanism is key to improve spiking decision transformers. (The title, abstract and text emphasise this point). However, the simulation results do not support this claim. Indeed, in the ablation study (Table 2), if we compare the baseline model (PSSA) or the PSSA + this “Shared-private” method (which is at the heart of OSSA), then performance decreases in 2 out 5 experiments and increases in the remaining 3. So I find difficult to make the point that this “Shared-private” method really improves performance.

  • The manuscript is not well written. Notations are not defined

  • Eq 1: what is E? what is LN? What is the domain of definition of x_i^0 ? etc…

  • Eq 2: LIF produce spikes, so are Q, K, V binary matrices? If yes, why are they expressed as real matrices? What are the dimensions of X^l? etc..

  • …. and so on up to the last equation

  • Eq 19: How is L_gate defined? It is just mentioned that “L_gate regularizes subspace activation, usage balance, and correlations among private projections”. This is not a definition.

  • Given the critical lack of clarity in the exposition, it is not possible to reproduce the simulations

Questions for authors
  • We request the authors to resolve the aforementioned readability issues and better demonstrate how their results back their central claim.
2 comments

W1: We acknowledge that the original Table 2 shows performance decreases on two of the five tasks for the PSSA + Shared-Private variant. These results alone do not provide sufficient evidence for the benefit of the shared-private design. To address this concern, we conducted additional multi-seed experiments using more controlled and resource-matched comparisons in the complete model. Specifically, we compare fully shared, fully independent, and shared-private OSSA constructions while keeping the adaptive gate, predictive objective, feedback modulation, and two-stage inference protocol unchanged. We also closely match their parameter counts, FLOPs, and inference latency. As shown in the table below, shared-private OSSA consistently outperforms both alternatives across all three representative MuJoCo and Adroit tasks. Compared with fully independent channels, it improves the normalized returns by 1.84, 1.76, and 1.25 points, respectively. These results provide stronger evidence that the shared-private construction contributes to performance beyond simply increasing model capacity.

Table: Computational cost and performance comparison across tasks (B=1B=1, context length 20).

ModelParams (M)FLOPs (G)Compute (μ\muJ)Latency (ms)Hopper-m-rWalker2d-m-rPen-cloned
PSSA3.180.50768.211.5592.38±2.4592.38 \pm 2.4571.29±4.3271.29 \pm 4.3244.92±13.4944.92 \pm 13.49
PSSA-MoreHeads6.271.000139.018.293.10±2.6693.10 \pm 2.6671.65±4.1071.65 \pm 4.1049.95±17.7649.95 \pm 17.76
Fully shared channels6.4331.276657.036.990.06±2.7390.06 \pm 2.7372.64±2.1572.64 \pm 2.1579.11±9.8379.11 \pm 9.83
Fully independent channels6.4381.273670.037.793.57±2.8993.57 \pm 2.8972.87±3.8472.87 \pm 3.8478.80±12.4178.80 \pm 12.41
SS-DT6.4371.274664.137.3895.41±2.5895.41 \pm 2.5874.63±3.5574.63 \pm 3.5580.05±7.9080.05 \pm 7.90

We further conducted component-level ablations to examine the contributions of predictive supervision, feedback modulation, and private-projection correlation regularization. Auxiliary prediction improves performance over OSSA on all three tasks. Adding prediction-error-based gate modulation brings further improvements of 1.71, 2.03, and 3.85 points, respectively. Removing the private-projection correlation regularizer from the full model reduces performance by 1.11, 0.83, and 0.86 points, respectively. These results confirm that predictive feedback and private-projection correlation regularization provide additional contributions to the overall performance. Table: Attribution of predictive supervision, feedback modulation, and private-projection correlation regularization.

VariantLpred\mathcal{L}_{\mathrm{pred}}Error modulates gateLcorr\mathcal{L}_{\mathrm{corr}}Hopper-m-rWalker2d-m-rPen-cloned
OSSA×\times×\times✓\checkmark92.5±2.392.5 \pm 2.372.0±2.772.0 \pm 2.775.8±2.575.8 \pm 2.5
OSSA + Aux-only✓\checkmark×\times✓\checkmark93.7±2.193.7 \pm 2.172.6±1.972.6 \pm 1.976.2±2.376.2 \pm 2.3
Full model w/o Lcorr\mathcal{L}_{\mathrm{corr}}✓\checkmark✓\checkmark×\times94.30±2.5294.30 \pm 2.5273.80±3.7873.80 \pm 3.7879.19±8.3679.19 \pm 8.36
Full model✓\checkmark✓\checkmark✓\checkmark95.41±2.5895.41 \pm 2.5874.63±3.5574.63 \pm 3.5580.05±7.9080.05 \pm 7.90

Together, these controlled comparisons provide evidence for the effectiveness of the shared-private construction and the additional components of OSSA. We have added these experiments and corresponding discussions to the revised manuscript.

W2–W7:

We thank the reviewer for the comments on the Method section. We have reorganized and revised the section to provide clearer mathematical definitions, tensor dimensions, and computational details. The main revisions are summarized below.

Eq. 1: We have rewritten the input formulation in Section 3 to explicitly define the state, action, return-to-go, embedding parameters, and Layer Normalization (LN). Eq. 1 now specifies the construction and dimensionality of each input token and the complete input sequence. We also clarify the shifted action input, including its initialization at the first timestep.

Eq. 2: We have clarified the complete Linear–Norm–LIF projection process in Section 3.1. Eq. 2 explicitly defines the generation of Q, K, and V, and the accompanying text specifies that their forward outputs are binary spike tensors. We further explain how the simulation-step dimension is folded into the multi-head representation and provide the resulting tensor dimensions. We also distinguish binary Q/K/V spikes from the real-valued features produced by the subsequent shared-private projections.

Eqs. 3–8: We have expanded the definitions of the shared and private projections in Section 3.1, specifying their dimensions, parameter sharing, and subspace-specific roles. We also clarify how their query-key correlations are constructed and how the shared and private components contribute to each subspace. Eqs. 5 and 6 explicitly describe the subspace attention and the additional causal full-head attention branch, respectively. The revised text defines the causal, local, and valid-token masks, as well as the positional modulation. We also clarify that the attention computation uses direct masked matrix products without softmax normalization or conventional scaling. Eqs. 7 and 8 specify how subspace messages are weighted, fused, and combined with the full-head branch.

Eqs. 9–10: We have explicitly defined the firing rate and spike-time centroid used by the communication gate, including their computation over the simulation steps and the treatment of neurons that never fire. Eq. 10 specifies the construction of the base communication logits and the roles of the learnable mappings.

Eqs. 11–13: We have clarified the complete two-pass information flow in Section 3.2. Eq. 11 defines the first-pass latent representation, action prediction, latent prediction, and prediction discrepancy. The accompanying text explains the stop-gradient operation, predictor dimensions, and initialization at the first timestep. Eq. 12 defines how the predictive context modulates the communication logits, including the subspace-dependent sensitivity and feedback attenuation. Eq. 13 specifies the sigmoid activation, temperature scaling, and mean normalization used to obtain the final communication weights. We also clarify that these weights allow multiple subspaces to contribute simultaneously rather than selecting a single subspace.

Eqs. 14–17: We have revised Section 3.3 to explicitly define the behavior cloning loss and auxiliary prediction loss in Eq. 14, including their normalization and valid-token masking. Eq. 15 defines the gate statistics, while Eq. 16 provides the complete mathematical definition of the gate regularizer. Its three components separately regulate global gate activity, subspace usage balance, and correlations among private projections. We also specify that the regularizer is computed from raw sigmoid gates in the second pass and provide the corresponding regularization coefficients. Finally, Eq. 17 defines the complete training objective and its prediction-loss coefficient.

Overall presentation: We have revised the descriptions throughout the Method section to clarify the dimensions, operations, and relationships between intermediate representations to better illustrate the shared-private architecture, subspace communication, and two-pass predictive feedback process. The revisions are highlighted in blue for your review. We look forward to receiving more of your valuable feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.