{"id":"bb64b4e4-b701-4ae8-bb98-c48f0fb2380c","arxiv_id":"2607.29048","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"SpikePEFT matches or beats full fine-tuning on spiking point-cloud benchmarks while updating only ~5% of parameters via neuron-dynamics and silent-state adapters.","lead":"SpikePEFT adapts frozen spiking point-cloud models to new tasks by tuning neuron firing dynamics and recovering information from 'silent' neurons while keeping most weights frozen. It reports 92.4% on ModelNet40 and 85.6% on ScanObjectNN's hardest split using about 5% of trainable parameters.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SSDA's core premise that near-threshold silent states carry task-relevant information is untested; the cited theory lives in a missing supplement and a random-selection ablation could falsify it.","rationale":"The reader identified exactly the same load-bearing assumption: SSDA's premise that near-threshold silent states carry recoverable task-relevant information, with the missing supplement as a key red flag. I agree this is the most load-bearing concern because SSDA is the primary source of performance gains (Table 4) and is presented as the paper's conceptual innovation. If this premise is false, the method is still an adapter but the scientific contribution ('recovering task-relevant sub-threshold information') collapses, and the method's claimed advantage over generic PEFT (e.g., random sparse adapters) is not established. The concrete test I propose — randomizing the selection mask while keeping all other components identical — directly isolates whether the near-threshold band is the reason for the improvement, rather than the added parameters. This test does not require the missing supplement; it can be run from the existing implementation. The verdict remains CONDITIONAL because the concern does not disprove the empirical results; it calls for additional evidence to validate the mechanism. The reader's original verdict already accounts for the missing supplement and lack of code; my check sharpens the question but does not change the status.","tokens_in":14386,"tokens_out":9054,"duration_ms":94613,"concrete_test":"Run an ablation with SPM on ModelNet40 and ScanObjectNN where SSDA's selection mask Q_t^l is replaced by a random per-token binary mask that selects the same number of silent channels per token as the learned SSDA (i.e., matching the mean |Ω_t,n^l| per block and timestep). Use the same down/up-projection A, B and bottleneck threshold for both variants, and keep IDT fixed. If the random-selection variant achieves accuracy within ±0.5 pp of the full SpikePEFT result, the near-threshold band is not uniquely informative, and the premise is unsupported. If the random variant degrades by >1 pp on the ScanObjectNN RS split, the near-threshold selection matters, supporting the premise.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central novelty, SSDA, is identified in Table 4 as the largest contributor to the reported gains (row 6 with only SSDA reaches 91.7/89.7/89.2/83.9 vs 89.1/83.6/82.8/77.8 for IDT alone). Its design assumes that membrane potentials in the narrow sub-threshold band (V_th^l - δ^l, V_th^l) carry task-relevant information that a learned sparse correction can extract. This is a load-bearing premise because if it fails, SSDA reduces to a generic per-token low-rank adapter whose success is due to added capacity rather than the 'silent-state disambiguation' mechanism claimed in the title and abstract. The paper explicitly defers justification: 'The theoretical motivation and additional analysis of near-threshold state selection are detailed in (cf. Supp. A)' (Method, SSDA subsection), but no supplement is included. Moreover, the premise is not self-evident: the membrane potential distributions in Figure 5 are histograms of raw values, and the shaded interval appears arbitrary; there is no control showing that the band carries more information than other sub-threshold intervals, or that the learned δ selects the informative region rather than overfitting to training data. Without a test that isolates the near-threshold property from the simple act of adding trainable parameters, the central mechanism's conceptual foundation is unverified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"SpikePEFT proposes a parameter-efficient fine-tuning framework for frozen spiking point cloud backbones. It combines Intrinsic Dynamics Tuning (IDT), which applies bounded residuals to LIF membrane decay and firing thresholds, and Silent-State Disambiguation Adaptation (SSDA), which selects near-threshold silent membrane states and applies a sparse event-driven low-rank correction before spike firing. The method is evaluated on ModelNet40, ScanObjectNN, ShapeNetPart, and S3DIS using SPM and E-3DSNN-L/SVL backbones, reporting 92.4% on ModelNet40 and 85.6% on ScanObjectNN PB_T50_RS with about 5% trainable parameters, and claiming to outperform full fine-tuning by up to 1.4 percentage points while preserving SNN energy efficiency. The internal equations (Eqs. 4-17) are consistent, and Table 4 identifies SSDA as the main source of improvement.","tokens_in":14801,"tokens_out":4819,"duration_ms":49070,"significance":"If verified, this is a useful first contribution: it extends PEFT to spiking 3D models, preserves spike sparsity, and reduces per-task storage. The sparse event-driven adapter design is interesting, and the benchmark coverage (classification, part segmentation, semantic segmentation) is broad. The claims are not yet fully supported because the core SSDA mechanism lacks the deferred theoretical justification and a control ablation, and the headline gains are reported without variance. With those additions, this could be a solid paper.","major_comments":[{"comment":"The central premise that near-threshold silent states in (V_th^l - δ^l, V_th^l) carry task-relevant information is explicitly deferred to 'Supp. A', but no supplement is provided. Table 4 shows SSDA is the largest source of gain (SSDA-only 91.7/89.7/89.2/83.9 vs IDT-only 89.1/83.6/82.8/77.8). Without a control that replaces the near-threshold selector by a random or fixed-width sub-threshold selector with equal parameter count, the gains cannot be attributed to 'silent-state disambiguation' rather than added low-rank capacity. Please provide the missing derivations and this ablation.","section":"Method, SSDA subsection and Eq. (9)"},{"comment":"The 'preserving energy efficiency' claim is computed from theoretical 32-bit FP constants (Horowitz 2014: 4.6 pJ/MAC, 0.9 pJ/AC), not measured hardware energy. The 6.22 mJ and '67.1% lower than PointMamba' figures are operation-count estimates under those constants. Please report measured energy (or wall-clock latency) on relevant neuromorphic/GPU hardware, or clearly label the numbers as idealized estimates and soften the claim.","section":"Abstract/Conclusion and Table 6"},{"comment":"Headline gains over full fine-tuning are 0.1-1.4 pp with no error bars, seeds, or significance tests; 0.1-0.4 pp differences are within typical run-to-run variation. Please report mean±std over at least 3 seeds for the main comparisons (especially SPM on ModelNet40 and E-3DSNN-L on ModelNet40, where gains are ≤0.1), or specify fixed seeds and show the comparison is stable.","section":"Table 1"},{"comment":"Implementation details and additional results are referenced as 'Supp. B' and 'Supp. C', and SSDA theory as 'Supp. A', but no supplementary file is included. Essential hyperparameters (e.g., ε_λ, ε_v, δ_max, r, ϑ_a) and the 5.3M segmentation-head parameter allocation in Table 2/3 are not given in the main text. The supplement must be submitted or these details moved into the main text.","section":"Experiments section"}],"minor_comments":[{"comment":"'achieves competitive than full fine-tuning' should be 'achieves competitive performance compared with full fine-tuning'.","section":"Conclusion"},{"comment":"The notation S_t^{l,0} is introduced for the preliminary spike, but Eq. (16) computes S_t^l without explicitly restating the relationship. Clarify the notation for readers.","section":"Method, Eqs. (6)-(16)"},{"comment":"The shaded region is not quantitatively linked to the learned δ; state which δ (or range of δ) is plotted, or add a legend.","section":"Figure 5"},{"comment":"Many references are dated 2026. If this is intentional for the submission timeline, verify all are publicly available; otherwise cite published or arXiv-stable versions.","section":"References"},{"comment":"The row label 'PEFT' is ambiguous; use 'SpikePEFT (Ours)' and clarify in the caption that the ANN baseline is PointMamba.","section":"Table 5"}],"recommendation":"major_revision","confidential_remarks":"The central idea is promising, but the manuscript currently asks the reader to accept the key mechanism on faith (missing Supp. A) and the empirical gains are small relative to likely run-to-run noise. I would condition acceptance on the control ablation for SSDA and either measured energy or a clear idealization label. The many future-dated references may also warrant an editorial check for availability and relevance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the take: SpikePEFT is a solid, clearly written first attempt at parameter-efficient fine-tuning for spiking point cloud models, and the gains over full fine-tuning look real even if they're small. The IDT and SSDA mechanisms are not just renamings of ANN adapters, and the ablations support the design. But the central assumption behind SSDA — that near-threshold silent states carry task-relevant information — is asserted and deferred to a missing supplement. That, plus the lack of error bars and measured energy, means the results are plausible but not fully verified.\n\nWhat's actually new: applying PEFT to SNN point-cloud backbones. IDT's bounded residual tuning of LIF decay and threshold is a natural but non-trivial idea; SSDA's sparse correction of near-threshold silent states is more distinctive. The math in Eqs. 4-17 is internally consistent, and Table 4 shows SSDA is the main contributor, which is honest and matches the narrative. The parameter reduction is real: about 5% of trainable parameters. The accuracy gains over full fine-tuning are 0.1–1.4 pp across datasets — not huge, but consistent, and the energy overhead appears modest.\n\nSoft spots: the load-bearing premise of SSDA is untested. The paper says the theoretical motivation is in Supp. A, but no supplement is included. Figure 5's histograms don't demonstrate that the near-threshold interval is special compared to other sub-threshold bands. A control with random channel selection instead of the δ-based interval would show whether the gain comes from the near-threshold property or just from added low-rank capacity. That's an addressable experiment, but without it the conceptual foundation is shaky. Energy numbers are computed from theoretical pJ/op constants, not measured on hardware, and there are no error bars or seeds. These are common issues in this area, but they weaken the energy-efficiency claim.\n\nThe paper is honest work. The comparisons are fair, the writing is clear, and the citation coverage is appropriate. It's not a breakthrough, but it's a legitimate first step in a direction others will likely build on.\n\nI'd send it to peer review — a good referee would ask for the supplement, a random-selection control for SSDA, and error bars. I'd bring it to a reading group only if you care about neuromorphic 3D perception or PEFT for SNNs. I'd cite it if I were working in this subfield.\n\nRecommendation: conditional accept, with the supplement and control as required revisions.","headline":"A sensible first PEFT framework for spiking point cloud models, but the core SSDA mechanism rests on an untested premise and a missing supplement.","tokens_in":15251,"tokens_out":3454,"would_cite":true,"duration_ms":42403,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"SpikePEFT is the first parameter-efficient fine-tuning framework for spiking point cloud models: it adapts a frozen spiking backbone by tuning membrane decay and firing thresholds and selectively correcting near-threshold silent states, rea","keywords":["spiking neural networks","point cloud analysis","parameter-efficient fine-tuning","membrane potential","silent-state disambiguation","intrinsic dynamics tuning","energy-efficient computation","event-driven"],"falsifier":"Take a trained SpikePEFT model and add noise to the membrane potentials of near-threshold silent states before SSDA is applied; if the accuracy does not drop, those states are not actually carrying task-relevant information. A more direct test is to compute the mutual information between the values in the selected band and the task labels—near-zero would falsify the claim that SSDA is recovering genuinely informative sub-threshold evidence.","tokens_in":14302,"feed_emoji":"⚡","tokens_out":3295,"duration_ms":37756,"temperature":0.7,"pith_summary":"This paper argues that a pre-trained spiking neural network for point clouds can be adapted to new tasks without full fine-tuning, by changing only the intrinsic dynamics of its neurons and by recovering information that binary spikes normally discard. The proposed SpikePEFT framework achieves accuracy slightly better than full fine-tuning on three real-world ScanObjectNN splits, using only about 5% of the parameters and adding a small amount of energy consumption. If correct, this means that the knowledge in a spiking 3D backbone is largely carried by its synaptic weights, and that task-specific adaptation can be done by adjusting membrane decay, firing thresholds, and the sub-threshold states that sit just below the firing boundary. The paper also shows the approach works across classification, part segmentation, and semantic segmentation, and on two different pre-trained spiking backbones.","feed_headline":"Spiking 3D models beat full retuning with 5% of parameters","feed_subtitle":"Tuning decay, thresholds, and near-threshold silent states adapts frozen spiking backbones at a fraction of the cost.","key_machinery":"The core object is the near-threshold membrane state of a spiking neuron. SSDA computes a binary mask Q_t^l that selects membrane potentials between the adapted firing threshold and a learnable lower interval boundary (V_th^l - delta^l), then uses a sparse bottleneck: a down-projection matrix that accumulates only columns corresponding to active events, a binary spike threshold in the bottleneck, and a zero-initialized up-projection that adds corrections only at the selected positions. IDT is the second mechanism, providing bounded residuals to membrane decay (in logit space) and firing threshold (multiplicative), with a tanh bound to keep the dynamics valid. Together they allow adaptation w","core_discovery":"SpikePEFT introduces two mechanisms for parameter-efficient adaptation of spiking point cloud models: Intrinsic Dynamics Tuning (IDT) and Silent-State Disambiguation Adaptation (SSDA). IDT keeps all synaptic transformations frozen and instead applies bounded residual adjustments to the membrane decay and firing threshold of spiking neurons, using a logit-space sigmoid for decay and a multiplicative exponential for the threshold. SSDA addresses the information loss caused by binary spike propagation: it identifies neurons whose membrane potential lies in a narrow band just below the firing threshold—silent yet potentially informative—and applies a lightweight, event-driven bottleneck correcti","pith_inferences":["The same pre-spike membrane modulation principle could apply to other discrete-activation networks, such as binary or quantized neural networks, where thresholding similarly discards borderline information.","The learnable interval delta in SSDA could be interpreted as an adaptive threshold band; a testable extension is to probe whether the learned band width correlates with task difficulty or per-channel sensitivity.","The energy comparison rests on theoretical per-operation costs; actual neuromorphic hardware measurements could change the magnitude of the advantage, though the qualitative shift from MAC to AC is likely to remain.","The success of SSDA hints that pre-trained spiking models may be underfitting because binary spike generation removes useful sub-threshold evidence; this could motivate new spike-encoding schemes or training objectives that explicitly preserve near-threshold information."],"forward_implications":["Frozen spiking backbones can be tuned for downstream 3D tasks with only ~5% of the trainable parameters, reducing storage and memory overhead compared with full fine-tuning.","SpikePEFT outperforms full fine-tuning on the most challenging ScanObjectNN split (PB_T50_RS) by more than one percentage point on both tested backbones, suggesting less destructive adaptation of pre-trained knowledge.","Recovering near-threshold silent states reduces the need for long temporal windows, improving accuracy with fewer time steps and lower latency.","The framework extends beyond classification to part and semantic segmentation, with larger gains in the more difficult point-level tasks.","Because the added operations are mostly accumulation operations (AC) rather than multiply–accumulate (MAC), the energy advantage of spiking models over ANN-based models is retained."],"fun_headline_variants":["SpikePEFT: 5% params updated, 92.4% on ModelNet40","Tune decay and threshold, not weights: 5% params","Silent-state tuning: 5% params for spiking 3D","Frozen synapses, tuned spike thresholds: 5% cost","Preserve SNN energy, update 5%: SpikePEFT"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The central premise is that membrane potentials in a narrow band just below the firing threshold contain task-relevant information that a learned sparse correction can recover; the paper defers the theory to a supplementary appendix that is not included, so the claim rests on empirical support.","fun_headline_variants_meta":{"raw":{"variants":["SpikePEFT: 5% params updated, 92.4% on ModelNet40","Tune decay and threshold, not weights: 5% params","Silent-state tuning: 5% params for spiking 3D","Frozen synapses, tuned spike thresholds: 5% cost","Preserve SNN energy, update 5%: SpikePEFT"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001104,"raw_usage":{"total_tokens":4436,"prompt_tokens":738,"completion_tokens":3698,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":482,"completion_tokens_details":{"reasoning_tokens":3598}},"tokens_in":482,"tokens_out":3698,"duration_ms":26238,"temperature":1.0,"reasoning_tokens":3598,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T14:36:27.566032+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a trained SpikePEFT model and add noise to the membrane potentials of near-threshold silent states before SSDA is applied; if the accuracy does not drop, those states are not actually carrying task-relevant information. A more direct test is to compute the mutual information between the values in the selected band and the task labels—near-zero would falsify the claim that SSDA is recovering genuinely informative sub-threshold evidence.","supporting_citations":[],"review_version":1}