{"id":"b9dcceca-bcf2-45f4-8c99-4e7f4abba242","arxiv_id":"2505.09659","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"LAS converts pre-trained ANN LLMs into fully spike-driven SNNs at 16 time steps using Outlier-Aware Threshold and Hierarchically Gated neurons, reporting near-lossless accuracy on OPT, BERT, GPT-2, and two vision-language models.","lead":"The authors convert pre-trained language models up to 66B parameters into spiking neural networks that compute with binary spikes instead of floating-point numbers, using two new neuron types to handle outlier activations and nonlinear functions. A generalist might read this because spike-driven inference promises large energy savings for LLMs if the conversion truly costs no accuracy, which this paper claims across language and vision-language benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lossless claim rests on unverified HG approximation fidelity and calibration; no error bound, layerwise check, or distribution-shift test is provided, and several of the paper's own tables show non-zero degradation.","rationale":"The reader's weakest assumption is the right one. LAS is a calibration-dependent conversion: OAT thresholds and HG sub-ranges/per-neuron counts are fit to pre-trained activation statistics, and no argument shows the approximation remains negligible after error accumulation. The strongest claim ('without any loss of performance') is therefore vulnerable. I give credit for the released code, for the essentially correct SAA accumulation identity, and for reporting the ViT limitation in Appendix D.2; these make the issue addressable rather than fatal. The existing CONDITIONAL verdict stands: the paper should either add a layerwise or end-to-end fidelity analysis with a distribution-shift check, or soften the lossless claim. Thus the reader's verdict is unchanged.","tokens_in":14947,"tokens_out":10586,"duration_ms":115578,"concrete_test":"Run the released LAS code on OPT-2.7B (or OPT-66B if feasible) and record ANN versus 16-step LAS hidden states at each converted module on (i) the calibration corpus used to fit lambda_i/N and (ii) a held-out distribution-shifted corpus. Compute mean and max relative L2 error per layer, and repeat end-task evaluation over 3 seeds. If mean relative error exceeds roughly 1% on either set, or if the end-task gap exceeds the ANN's seed-to-seed spread, then the calibration-dependent 'lossless' claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The 'lossless' headline requires the HG neuron's piecewise approximation of GELU, exp, and 1/sqrt (Sec. 4.1) to be accurate on every activation that occurs after conversion, not just on the calibration distribution. This is the least secure condition: thresholds lambda_i are 'dynamically adjusted according to the statistical distribution of activation values in pre-trained LLMs' and N is 'optimally tuned per neuron according to each model's error tolerance' (Appendix D.2), but no error bound, layerwise fidelity check, or distribution-shift test is reported. With 66B parameters, small per-neuron errors can accumulate and propagate through softmax and LayerNorm; the two curve plots (Figs. 3-4) are illustrative only. The paper's own tables also show non-zero gaps (OPT-66B: Winogrande 69.98->68.27, RTE 60.55->59.93; Qwen2-VL: HallusionBench 130.66->125.81; GPT-2 WikiText-103 PPL 16.53->16.79), so 'without any loss' is not literally supported. Appendix D.2 further states that the LLaVA vision encoder is retained in analog, so the vision-language model experiments are not end-to-end fully spike-driven.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"LAS proposes an ANN-to-SNN conversion scheme for LLMs. It introduces an Outlier-Aware Threshold (OAT) neuron, built from dual multi-threshold sub-neurons, and a Hierarchically Gated (HG) neuron that approximates nonlinear functions through gated FS sub-neurons. It then reformulates self-attention, feed-forward networks, LayerNorm, and Softmax as spike-equivalent modules, and evaluates the converted models on OPT-2.7B to OPT-66B, BERT-base, GPT-2, LLaVA1.5-7B, and Qwen2-VL-7B at 16 time steps. The paper reports near-ANN accuracy on many tasks, some improvements (e.g., OPT-66B WSC), and several non-negligible degradations. The central advertised claims are that the conversion is lossless and that the resulting models are fully spike-driven.","tokens_in":15304,"tokens_out":6305,"duration_ms":63610,"significance":"If fully supported, this would be a significant advance: converting pre-trained LLMs of up to 66B parameters to event-driven, spike-based computation at 16 time steps without task-specific training. The paper has concrete strengths: the incremental SAA multiplication decomposition in Appendix A is a useful algebraic contribution; the experimental coverage across language understanding, generation, and vision-language benchmarks is broad; ablations isolate the roles of OAT and spike-equivalent attention; and source code is released. However, the headline 'lossless' claim is contradicted by the paper's own tables, and the fidelity of the HG approximation, the distribution-shift robustness of its calibration, and the fully-spiking status of the vision-language experiments are not established. The methodology is promising, but the claims need to be scaled back and supported by additional evidence.","major_comments":[{"comment":"The claim of lossless conversion is not supported by the reported numbers. Table 2 shows GPT-2 Enwik8 BPB 0.97 vs. 0.96 and WikiText-103 PPL 16.79 vs. 16.53; Table 4 shows LLaVA1.5-7B HallusionBench 67.93 vs. 81.36 and POPE 80.79 vs. 82.81; Table 1 shows OPT-66B Winogrande 68.27 vs. 69.98 and RTE 59.93 vs. 60.55. These are non-zero degradations, so the statements 'lossless conversion' and 'without any loss of performance' are not literally supported. Please revise the claims to a clearly defined 'near-lossless' statement, or provide statistical significance or error-bar analysis that justifies treating these gaps as acceptable.","section":"Abstract and §5.2, Tables 1–4"},{"comment":"The HG neuron is the component responsible for approximating GELU, exp, and 1/sqrt, but no error bound is given. The thresholds λ_i are said to be 'dynamically adjusted according to the statistical distribution of activation values' and the number of FS sub-neurons N is said to be 'optimally tuned per neuron,' yet no algorithm, optimality criterion, or sensitivity analysis is reported. Since the conversion is calibrated on activation statistics from the pre-trained ANN but is then deployed on test inputs, the robustness of the approximation to distribution shift is load-bearing. Please provide layerwise approximation-error measurements on held-out activations, and analyze how errors in λ_i and N propagate to final task accuracy.","section":"§4.1 and Appendix D.2"},{"comment":"The claim that the converted models are fully spike-driven is overstated for the vision-language experiments. Appendix D.2 states that 'we retained its analog weights' for the ViT component of LLaVA1.5-7B. Therefore the LLaVA and Qwen2-VL results are not end-to-end spike-based, so the abstract's 'full spiking conversion' and Section 4.2's 'Spike-Equivalent LLM architecture' claim do not cover the full evaluated systems. Please either clarify this limitation in the abstract and scope statements, or convert the vision encoder as well and rerun the experiments.","section":"Appendix D.2 and §4.2"},{"comment":"The derivation of the spike offset for softmax is not evidently correct as written. Eq. (16) defines max operations over time-prefix sums, but softmax requires subtracting the maximum over the n logits at each position. As written, \\ hat z_i(t) = z_i(t) + max(prefix_{t-1}) - max(prefix_t) does not equal z_i - z_max; the notation also makes it unclear whether the max is over the time index or over the softmax input dimension. Please provide a correct derivation with consistent notation, or explain how the time-prefix maximum implements the per-position logit maximum.","section":"§4.3, Eq. (16), and Appendix C"}],"minor_comments":[{"comment":"Figure 5's caption mentions 'STSB task' while the text says the ablation is on RealWorldQA with Qwen2-VL-7B; please make the labels consistent.","section":"Figure 5 and §5.5"},{"comment":"The key matrix is denoted K in the text but appears as Q_k in Eq. (11), and Eq. (18) uses both i and j in the double sum inconsistently; please correct these notation errors.","section":"Eq. (11) and Appendix A, Eq. (18)"},{"comment":"The definitions of S_q(t) and S_k(t) appear swapped: S_q(t) is written as a sum over key spikes and S_k(t) as a sum over query spikes. As printed, the expansion does not match the final product; the definitions should be exchanged.","section":"Appendix A, Eq. (19)"},{"comment":"The paragraph on NLU tasks begins 'FAS achieves near-lossless conversion' but should refer to LAS.","section":"§5.2, NLU paragraph"},{"comment":"The affiliation contains a typo, 'Collage of Computer Science' should be 'College'; the abstract's 'improves the accuracy of 2%' is imprecise because the WSC increase is 2.93 points and percentage points should be distinguished from percent.","section":"Affiliation and abstract"},{"comment":"In the negative threshold branch, the denominator appears as 'n' in '- (H+k)/n θ(t)', which should probably be 'H'; please correct.","section":"Eq. (7)"}],"recommendation":"major_revision","confidential_remarks":"The paper has a useful core idea and broad experiments, but the advertised 'lossless' claim is directly contradicted by several rows of its own tables, and the fully-spiking scope is not accurate for the vision-language models. I would be willing to accept a revised version that reframes the claims, adds the missing HG fidelity and calibration analysis, and corrects the softmax derivation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is the strongest ANN-to-SNN conversion result for generative LLMs I've seen, and the 'lossless' headline is not supported by the paper's own tables. The two new neurons—OAT for outliers, HG for nonlinearities—are sensible, and converting OPT-66B at 16 steps is a real scale first. The spike-equivalent self-attention derivation in Appendix A is coherent and incremental, not hand-wavy. The softmax offset trick in Appendix C also checks out as an algebraic identity. So there is a solid core here.\n\nWhat it does well: the ablation study is informative. Removing OAT costs about three points on RealWorldQA; disabling spike-equivalent attention costs about 48 points, which tells you the attention module is load-bearing. The code is public. The comparison against prior SNN conversion work is fair, and LAS largely dominates on the reported benchmarks.\n\nThe soft spots, in proportion:\n\n1. 'Lossless' is literally contradicted by the paper's data. GPT-2 Enwik8 goes 0.96 to 0.97; WikiText-103 perplexity goes 16.53 to 16.79; LLaVA HallusionBench drops from 81.36 to 67.93 and POPE from 82.81 to 80.79; Qwen HallusionBench drops from 130.66 to 125.81; several OPT rows fall slightly. Some of these are within noise, but HallusionBench is a 13-point drop. The 'without any loss of performance' claim should be softened to 'near-lossless with small, task-dependent degradation.'\n\n2. The HG neuron's fidelity is the load-bearing assumption, and it is not tested. Thresholds lambda_i and sub-neuron count N are tuned to activation statistics and per-model error tolerance; there is no error bound, no layer-wise check after conversion, and no distribution-shift analysis. With 66B parameters, small per-neuron errors could accumulate. The two curve plots (GELU, exp) are illustrative, not evidence of end-to-end fidelity.\n\n3. The LLaVA experiments are not fully spike-driven: Appendix D.2 says the ViT encoder is retained in analog. That is a legitimate engineering choice, but it means the 'fully spike-driven VLM' claim is only true for the language part, and that caveat should appear in the abstract or main text rather than being buried in the appendix.\n\n4. No error bars or multiple seeds anywhere; zero-shot evaluations on small benchmarks can vary. Minor, but relevant to the claimed 2% WSC improvement.\n\nWho it's for: neuromorphic computing researchers and anyone working on energy-efficient LLM inference. The method is a credible recipe for converting large pretrained transformers to spike-based form, and the scale alone makes it worth reading.\n\nMy recommendation: send it to peer review. It needs revision to align claims with evidence—particularly the word 'lossless'—and ideally a robustness or distribution-shift analysis of the calibration. But the core method and experiments are substantial enough that a serious referee should engage with it, not desk-reject.","headline":"A real scale-first advance in spike-based LLM conversion, but the 'lossless' claim outruns the paper's own tables; the method and engineering deserve a serious referee.","tokens_in":15807,"tokens_out":1875,"would_cite":true,"duration_ms":17193,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"LAS converts pre-trained LLMs into fully spike-driven form with no loss of task accuracy at 16 time steps.","keywords":["ANN-to-SNN conversion","spiking neural networks","large language models","lossless conversion","event-driven computation","outlier-aware threshold neuron","hierarchically gated neuron","spike-driven transformer"],"falsifier":"Take a converted model, feed it inputs whose hidden activations fall measurably outside the (lambda_{i-1}, lambda_i] ranges used to set the HG sub-neuron thresholds, and check whether task accuracy drops well below the original ANN baseline; a spike-specific, statistically significant drop on such out-of-distribution inputs would falsify the lossless claim.","tokens_in":14763,"feed_emoji":"⚡","tokens_out":4642,"duration_ms":44976,"temperature":0.7,"pith_summary":"This paper claims that large pre-trained language models can be converted into fully spike-driven spiking neural networks with no loss in task accuracy, using only 16 time steps. The method, LAS, replaces every floating-point computation of a pre-trained Transformer—attention, feed-forward, LayerNorm, softmax, and activation functions—with spike-based equivalents. If the claim holds, pre-trained LLMs such as OPT-66B, BERT, and GPT-2 could be deployed on event-driven neuromorphic hardware with large energy savings and without any task-specific retraining. The paper backs the claim with zero-shot, language-generation, and vision-language evaluations showing accuracy matching, and in a few cases exceeding, the original ANN models.","feed_headline":"LLMs run fully on spikes with no accuracy loss","feed_subtitle":"A conversion method turns OPT-66B, BERT, and GPT-2 into event-driven SNNs that match or beat the original models.","key_machinery":"The machinery is a pair of bespoke spiking neurons plus spike-form rewrites of the Transformer's nonlinear modules. The OAT neuron uses two multi-threshold sub-neurons, one for normal and one for outlier activations, so that a single spike train can encode both the fine structure of typical values and the extreme tail. The HG neuron stacks several Few-Spike sub-neurons, each responsible for a sub-range (lambda_{i-1}, lambda_i], gated by binary masks, and the sub-neuron parameters are fitted on synthetic uniform samples from that sub-range; this hierarchical decomposition is what lets a few spikes approximate GELU, exp, and inverse square root. The spike-equivalent attention decomposes the product of accumulated query and key spike trains into an incremental form using only binary operations, and the spike Softmax reconstructs the max-subtraction, exponentiation, and reciprocal normalization through corrected spike outputs at each time step.","core_discovery":"The central claim is that all floating-point computations of a pre-trained Transformer can be rewritten as spike operations with no loss in output accuracy, provided two new neuron types carry the conversion. The Outlier-Aware Threshold (OAT) neuron splits activations into normal and outlier streams via dual multi-threshold sub-neurons, preventing extreme values from compressing everything into one spike bin. The Hierarchically Gated (HG) neuron approximates nonlinear functions such as GELU, exp, and reciprocal by partitioning the input range into sub-intervals, each handled by its own few-spike sub-neuron, with thresholds set from the statistical distribution of activations. On top of these neurons, the paper constructs spike-equivalent self-attention, feed-forward networks, LayerNorm, and Softmax, so the entire model runs event-driven. The claimed result is that this preserves, and sometimes slightly improves, task accuracy on OPT-2.7B to OPT-66B, BERT, GPT-2, LLaVA1.5-7B, and Qwen2-VL-7B at 16 time steps.","pith_inferences":["The 'lossless' label is best read as benchmark-equivalent performance, not mathematically exact reproduction; the HG neuron's fitted sub-range approximators are only guaranteed to stay accurate while deployment activations fall within the calibrated lambda_i ranges.","For vision-language models the method converts the language backbone while the paper reports retaining the analog vision encoder weights; a fully spiking multimodal pipeline is a direct next step.","Because the HG neurons are fit from synthetic uniform samples on each sub-range, the same calibration machinery could be re-run at deployment on a small set of real activations to adapt the approximation to distribution shift.","The energy model counts only spike-based synaptic operations and drops the analog residual memory and vision-encoder costs; a hardware-aware energy benchmark would clarify the end-to-end savings on chip."],"forward_implications":["Pre-trained LLMs up to 66B parameters can run in fully event-driven form at 16 time steps with benchmark accuracy intact.","Deployment no longer needs surrogate-gradient training or task-specific adaptation; conversion is a plug-in step on already-trained weights.","The time-step budget drops from the 256+ steps required by prior transformer conversion to 16, which directly shrinks latency and energy per inference.","The same component-wise recipe—OAT before linear layers, HG for nonlinearities—extends to other Transformer variants with gated FFNs and different activation functions."],"supporting_citations":[{"why":"Supplies the pre-trained OPT model family that LAS converts and evaluates at scales from 2.7B to 66B.","marker":"[46]"},{"why":"Supplies the pre-trained BERT-base model used for the NLU experiments on GLUE tasks.","marker":"[10]"},{"why":"Supplies the pre-trained GPT-2 model used for the language-generation experiments.","marker":"[29]"},{"why":"Defines the Few-Spikes neuron whose temporal-coding dynamics the OAT and HG sub-neurons extend.","marker":"[36]"},{"why":"Prior ANN-to-SNN conversion work by the authors that serves as a baseline on OPT-7B and provides the comparison for LAS's gains.","marker":"[7]"},{"why":"Baseline conversion method (SPR) with residual membrane potential, compared on BERT and GPT-2.","marker":"[14]"},{"why":"Baseline conversion method (QCFS) with optimal threshold, compared on BERT and GPT-2.","marker":"[5]"},{"why":"Baseline conversion method (COS) with offset spikes, compared on BERT and GPT-2.","marker":"[15]"},{"why":"Prior transformer conversion method (STA) that requires at least 256 time steps, setting the time-step baseline LAS improves to 16.","marker":"[20]"},{"why":"Supplies the synaptic-operations-to-FLOPs energy model and the E_MAC/E_AC constants used to estimate energy savings.","marker":"[30]"}],"fun_headline_variants":["Spiking LLMs, zero loss: LAS converts Transformers to spikes","Lossless spike conversion for LLMs: even 66B models work","Turning LLMs into SNNs without losing a bit of accuracy","Full spike LLMs: no performance drop, sometimes a gain","LAS: spike-driven LLMs that match or beat original accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the assumption that the few-spike sub-neurons, with thresholds fixed from the statistical distribution of activations in the pre-trained model, can approximate GELU, exponential, and inverse-square-root accurately enough across every input the converted model will actually see.","fun_headline_variants_meta":{"raw":{"variants":["Spiking LLMs, zero loss: LAS converts Transformers to spikes","Lossless spike conversion for LLMs: even 66B models work","Turning LLMs into SNNs without losing a bit of accuracy","Full spike LLMs: no performance drop, sometimes a gain","LAS: spike-driven LLMs that match or beat original accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000406,"raw_usage":{"total_tokens":2124,"prompt_tokens":973,"completion_tokens":1151,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":589,"completion_tokens_details":{"reasoning_tokens":1061}},"tokens_in":589,"tokens_out":1151,"duration_ms":8268,"temperature":1.0,"reasoning_tokens":1061,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:38:05.207985+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a converted model, feed it inputs whose hidden activations fall measurably outside the (lambda_{i-1}, lambda_i] ranges used to set the HG sub-neuron thresholds, and check whether task accuracy drops well below the original ANN baseline; a spike-specific, statistically significant drop on such out-of-distribution inputs would falsify the lossless claim.","supporting_citations":[{"cited_title":"Optimized spiking neurons can classify images with high accuracy through temporal coding with two spikes.Nature Machine Intelligence, 3(3):230–238, 2021","cited_arxiv_id":null,"evidence_quote":"Defines the Few-Spikes neuron whose temporal-coding dynamics the OAT and HG sub-neurons extend."},{"cited_title":"Reducing ann-snn conversion error through residual membrane potential","cited_arxiv_id":null,"evidence_quote":"Baseline conversion method (SPR) with residual membrane potential, compared on BERT and GPT-2."},{"cited_title":"Spatio-temporal approximation: A training-free snn conversion for transformers","cited_arxiv_id":null,"evidence_quote":"Prior transformer conversion method (STA) that requires at least 256 time steps, setting the time-step baseline LAS improves to 16."}],"review_version":1}